A study published in Nature Health earlier this year should have been unambiguously good news. Researchers evaluated an AI clinical decision support tool, GPT-4o integrated into the electronic medical records at Penda Health, a primary care chain across 16 clinics in Kenya (full disclosure: Endless is a proud supporter of Penda’s amazing work, although not this study in particular). The AI’s recommendations aligned with clinical guidelines 99% of the time. It surfaced new diagnostic insights in 94% of encounters. It cost half a cent per patient. A companion real-world study found that clinicians using it made 16% fewer diagnostic errors and 13% fewer treatment errors.

By every capability metric, the AI worked.

But buried in the data was a pattern that should concern anyone designing clinical AI systems. When the AI got it wrong, which happened in 7.8% of encounters, clinicians went along with the harmful recommendation, fully or partially, 58% of the time. The human safety net we seem to be designing our entire clinical AI model around wasn’t functioning, at least in this otherwise high-functioning context.

The standard reaction to this well-publicized study I have seen on my LinkedIn feed: we need more friction in clinical decision support systems, more human oversight over AI, better training so physicians can catch AI errors. I’ve spent 15+ years working across health systems in Asia and Africa, and I think that reaction, while understandable, misses several more fundamental (and one radical) questions that have implications for the AI-for-health research agenda.

The uncomfortable truth of the ‘know-do’ gap

Before we debate AI safety, we need to confront what actual clinical care looks like in most of the world.

Jishnu Das and colleagues have spent two decades studying health markets across India, China, and Kenya using standardized patients: trained actors presenting consistent symptoms to real providers. Their findings are devastating. In rural India, 72% of providers knew the correct treatment for simple childhood diarrhea; 17% actually gave it. Fewer than half of cases were correctly managed even by generous definitions. In rural China, 64% of the medications dispensed were unnecessary or outright harmful.

A separate study in Senegal found that overconfident health workers are 26% less likely to manage patients correctly, not for lack of knowledge, but seemingly because confidence substitutes for effort.

So coming back to Penda’s 7.8% harmful AI output rate: compared to what? Multiple experts I work closely with, people with decades of clinical and research experience across India and East Africa, say even older generation AI models like GPT-4 produce objectively better guidance than the majority of real-world practice. Not better than the best doctors on edge cases. But certainly better than what patients actually get from the vast majority of providers on the vast majority of cases.

We’re not comparing AI to a well-rested specialist following guidelines meticulously. We’re comparing AI to Tuesday afternoon in a rural public clinic, where the doctor is reluctantly wrapping up his last few patients while eager to get back to his private practice for the evening to make some supplemental income.

I will be the first one to tell you that guideline alignment isn’t the same as patient outcomes. A doctor who scores lower on a reasoning rubric might still produce better results because she reads the patient’s anxiety or catches something outside the textbook. True. The outcome studies don’t exist yet. But the know-do gap can’t be explained away as a proxy problem; it’s patients receiving incorrect treatment in the majority of encounters across three continents. The Lancet Global Health Commission estimated that poor-quality care kills about 5 million people a year in low- and middle-income countries, more than the 3.6 million who die from not reaching care at all. That ought to be enough to justify testing models of care that look rather different from what we have today.

What the studies show about AI augmentation

Three lines of evidence have been accumulating, and they all point the same direction.

First, AI outperforms doctors on clinical reasoning — even good doctors. A study at Beth Israel Deaconess (JAMA Internal Medicine, 2024) found that GPT-4 scored a median 10/10 on a standard clinical reasoning rubric, compared to 9/10 for attending physicians and 8/10 for residents. These were Harvard-affiliated physicians, not the global average clinician. (The same study found the AI also produced more instances of frankly wrong reasoning than the residents did. Hold that thought; it matters for how we design the human role.)

Second, adding a physician to AI doesn’t seem to make it better. Goh et al (Nature Medicine, 2025) randomized 92 physicians. Those with AI access improved over physicians using conventional resources, but merely matched AI-alone performance. The physician added time, not accuracy. A follow-up analysis (npj Digital Medicine, 2026) of how those physicians actually used the AI found four types of users — copy-pasters, selective copy-pasters, summarizers, and searchers — and none of them outperformed the others. Most physicians used AI like a search engine, not a reasoning partner.

It’s important to note that in most of these studies, physicians weren’t trained to use AI as a diagnostic partner. They anchored on initial assessments and cherry-picked confirmatory evidence. With explicit training, and attitudinal shifts that treat AI as an oft-superior partner, results could differ. But “give doctors AI and let them figure it out” is the current mental model, and it clearly doesn’t work.

Third, the human safety net fails precisely when it’s needed most. Back to Penda: the human-in-the-loop caught and rejected fewer than half of the AI’s harmful recommendations. The oversight model assumes active, critical engagement with every AI output. That engagement isn’t happening. We can keep designing for an idealized fiction, or for messy reality.

Who is the right provider?

Here’s the first controversial hypothesis I will propose, based on 15+ years of working with doctors and other types of health providers across the global south.

In most countries where I have worked, medical schools are among the most elite institutions in society. Admission is intensely competitive, often drawing from a narrow slice of the population. The training cultivates a sense of exceptionalism, and reasonably so. These are people who save lives. The profession carries an aura of nobility that few others can match, amplified by the genuine stakes of the work.

But that same elite formation creates a specific friction when AI enters the picture. When your professional identity is founded on being the one who knows, accepting that a machine might know better becomes nigh-impossible.

Maybe physicians resist AI out of deeper understanding, not identity. They’ve seen tools fail, and that caution has real value. But as we saw above, the resistance shows up even when the AI’s reasoning is demonstrably better. That goes beyond healthy skepticism.

Now contrast with what I’ve seen in West Bengal. Our grantee, Heal India working with the Liver Foundation, trained nurses on an AI diagnostic workflow: recording patient interviews, sending audio to Gemini, getting back syndrome-based differentials in 60 seconds. The nurses adopted the tool with enthusiasm. They expanded its use beyond the original scope, building their own training materials for snakebite treatment without being asked. The tool became a platform for local knowledge creation. A small proof-of-concept pilot with real patients produced consistently strong results.

The nurses didn’t experience AI as a threat to their capabilities. They experienced it as an expansion. This is not just anecdotal; in the literature, nurses show consistently positive, if appropriately cautious, attitudes toward AI.

This brings me to my prediction: nurse + AI will likely outperform physician alone, or even physician + AI. Thankfully, that claim is verifiable: whether providers without the professional identity formation of elite medical training collaborate with AI more effectively and consistently. In fact, we are already testing this. The same Heal India-Liver Foundation study in West Bengal is comparing AI-assisted nurses to physician-only care across a range of quality indicators (stay tuned for some exciting results coming out soon!).

Of course nurses aren’t the only providers without elite professional training. There are many other cadres of providers like paramedics, medical officers, community health workers, pharmacists, and informal “village doctors” — the cadres global health has always treated as the bottom of the clinical hierarchy — who may similarly work more effectively with AI. But if I had to intuit, providers may also need a certain minimum threshold of medical fluency to provide AI the right diagnostic information as inputs, and so there may be a sweet middle spot that nurses and medical officers occupy which may be hard for less trained providers to fill.

Should we let the machine go first?

As if my previous postulation was not provocative enough, let me go even a step further. If AI outperforms AI+Clinicians already (Goh et al above), why aren’t we letting AI become the auto-pilot instead of the copilot?

It is certainly not the first time we have delegated sophisticated skills to a machine. About 90% of flight time on a modern commercial jet is flown on automation, and pilots override when needed. Nobody thinks this diminishes pilots. It focuses them on the moments where human judgment matters most, like when Captain Sully had to figure out how to land US Airways 1549 on the Hudson River in 2009 and save 155 lives.

What would AI-first clinical care look like? The AI could generates the assessment, differentials, and treatment plan, fed by a sophisticated AI-driven intake that is filled in by a nurse or other non-clinician provider. The human clinician reviews, validates, exercises judgment on edge cases. The provider’s job becomes quality assurance, not primary reasoning.

When I started writing this piece back in April, I called this an eminently testable question. The first test has since arrived, and I’m proud to note it was co-authored by our grantees at Heal India alongside teams from Stanford and Microsoft Research. A randomized trial in npj Digital Medicine had 70 US physicians work through diagnostic cases in two configurations: the AI gives its assessment first and the physician reviews it, or the physician commits to an assessment first and uses the AI to verify. AI-first won: 85% diagnostic accuracy versus 82% for AI-as-second opinion, with a clearer edge on the recommendations that actually change patient management, and faster to boot.

But the finding that should reframe this whole debate is what happened to the AI when it went second. In second-opinion mode, the AI’s supposedly independent assessment simply matched whatever the clinician had already concluded in 48% of cases. In first-opinion mode? Only 3%. As is well known already, large language models are sycophants: show them your answer, and they anchor from it, just like humans do. Which seat the machine sits in determines whether you get an independent second mind or a mindless validation of your own flawed first impression.

Put that next to the Penda finding at the beginning of this post and the irony is complete. We put the human second to catch the machine’s errors, and the human waves them through. We put the machine second to catch the human’s errors, and the machine flatters them. The configuration we’ve made the default, where the human reasons first and the AI advises, may be the one that reliably degrades both parties.

I don’t know where we will eventually land on this question. But it is clear to me that the current model is flawed. And I would like to see studies that rigorously tests this question to find the optimal workflow for human-AI collaboration, instead of retro-fitting AI to the present model of care. The stakes are the 5 million people who die from poor quality care every year, so we cannot afford to get this wrong.

(A crucial caveat: AI-first does not mean friction-free. The Penda data shows exactly why judgment gates have to be explicit so that human reviewers cannot passively wave through AI mistakes. Well-designed workflows and UX are crucial to achieve this optimal friction. I mean AI-first here as a design principle, not a shortcut.)

An eminently testable agenda

Above I cited the first few studies that are the canaries in the coalmine telling us we are doing something wrong. They are not the final verdict for sure. What I am aiming to push for in this post is to question our basic assumptions about the models of care based on everything we know about current care models and their failures, human blind spots and prejudices, and the rapidly evolving field of AI underneath our feet.

Finding the optimal path forward would require us to question and investigate those basic assumptions and blind spots, and think from first principles what super-human medical reasoning from an AI affords us to do today that we could not do yesterday. This may mean leveraging abundant non-clinician resources already present everywhere in the world, retraining doctors to follow AI and clean up its errors rather than doing primary reasoning themselves, and much more.

Getting there requires a well designed research agenda, not a leap of faith. Three lines of work, in ascending order of radicalism:

1. Train doctors to collaborate with AI, then measure again. Every negative physician+AI finding so far comes from untrained users treating a reasoning partner like a search box. If explicit training and the right incentives close the gap, the co-pilot model survives. If they don’t, that tells us something too.

2. Run nurse + AI against physician arms, head to head. Not AI versus doctor: nurse-plus-AI versus doctor-alone and doctor-plus-AI, on patient outcomes, not just guideline scores. The Heal India-Liver Foundation West Bengal comparison is an early start. We need ten more like it, across contexts and cadres.

3. Test AI-as-default against AI-as-optional-input. The npj trial ran this on vignettes with US physicians. Nobody has run it where it matters most: in the clinics where the know-do gap lives, and where the alternative to AI-first is not a Harvard attending but an empty consultation room.

I’d love to hear from practitioners, policymakers, and physicians on this — especially those who disagree. What am I missing? Where does this argument break down? Drop a comment or reply to this email.

Leave a comment

We get into the chicken-and-egg problem that traps both providers and payers in fee-for-service, why governments in many countries act as competitors rather than stewards of the health market, and the counterintuitive finding that integration may actually be easier in the hardest markets. We also confront the provocative critique from health economist Jishnu Das that integration creates monopoly power — and what AI and digital health tools might mean for a future where you get the benefits of integration without the organizational costs.

Every conversation with Nachiket and Saul leaves me rethinking assumptions I didn’t even know I had. I hope this one does the same for you. Let me know in the comments what resonates.

Thanks for reading Rubayat Khan for ∞ Endless Health! This post is public so feel free to share it.

Share Rubayat Khan for ∞ Endless Health

CHAPTERS

00:00 Introduction to Integration of Care

01:47 Understanding Fragmentation in Healthcare

06:10 Provider Perspectives on Integration

09:21 Building Integrated Models: The Health Force Story

16:25 Systemic Challenges in Healthcare Integration

23:23 Successful Integration Models in Low Resource Markets

36:37 Negotiating Power in Healthcare

39:48 Challenges in Primary Care Models

41:23 The Role of Health Workers

45:06 Market Dynamics and Patient Choices

50:49 Unique Attributes of Healthcare Markets

53:45 How Digital Health Solutions Could Enable Integration Without the Downsides

58:08 Philanthropy’s Role in Healthcare Innovation

59:59 Optimism for Healthcare Transformation

CONVERSATION SUMMARY

🐔 THE PROBLEM: THE CHICKEN-AND-EGG TRAP

1. Everyone Agrees Fragmentation Is Bad — Nobody Can Fix It

• Healthcare systems worldwide are organized around episodes and transactions, not the person behind the patient

• My US experience: three months coordinating between doctor, pharmacy benefits manager, and insurance to fill one prescription

• Saul’s South Africa: private sector organized for shareholders, not patients — fee-for-service drives fragmentation, duplication, and inflated costs

2. The Chicken-and-Egg Problem (Saul’s Core Insight)

• For payers to pay for integrated care, providers need to organize themselves in an integrated manner

• For providers to organize in an integrated manner, payers need to pay for integrated outcomes

• Neither side can move first without the other — the system is structurally stuck

• Even when startups try to break in, incumbents actively lock them out or replicate the innovation themselves

3. Governments as Competitors, Not Stewards

• In India, Bangladesh, and many LMICs, governments see themselves as players in the market, not facilitators

• They compete with the private sector while simultaneously holding regulatory authority over it — a fundamental conflict

• Despite billions invested in public health systems, populations still overwhelmingly flood to private providers

• South Africa: under-regulated private sector where “forces of capitalism have driven evolution for shareholders, not patients”

🔑 Takeaway: Integration stalls not because people don’t want it, but because the incentive structure creates a deadlock. Payers, providers, and governments are each waiting for someone else to move first.

🔧 TWO MODELS OF INTEGRATION: HOSPITAL-FIRST vs. PRIMARY-CARE-FIRST

1. HealthForce — Engineering Integration Within Fee-for-Service (Saul)

A pharmacy-based primary care model that “hijacked” the fee-for-service system from within:

• 550 pharmacies with nurse-run clinics + telemedicine GP access across South Africa

• 4 million patients served

• Nurse handles 80% of consultations without a doctor; GP picks up from where nurse + tech left off for the remaining 20%

• Compensated clinicians on behavior metrics (patient experience, collaboration, wait times) — not just time billed

• Shared fees between team members to incentivize teamwork, despite collecting fee-for-service on the other end

• Result: 75% cost reduction in primary care by organizing teams around patient need

2. Dual Integration Pathways — The Indian Private Sector (Nachiket)

Based on Nachiket’s paper in Health Care, two starting points for integration:

• Hospital-first (secondary care down): Large hospitals entering primary care to complete the “circle of care” — 40,000+ hospitals in India could do this

• Primary-care-first (subscription up): GPs building subscription-based panels — one provider in Tamil Nadu has 2 million patients and calls the shots with insurers and hospitals

• Primary care “owns the customer” — sees patients 3-4x/year vs. secondary (2% of time) and tertiary (0.02%)

• The banker’s argument: integrated care produces steadier, more predictable cash flows — better for equity markets than volatile fee-for-service

• Examples across markets: Kaiser Permanente (US), CareMore (US, grew across 11 states), Brazil’s hospital-led vertical integrations

💡 Key Insight: Saul and Nachiket reveal two distinct but complementary strategies — Saul engineered integration within the existing payment system; Nachiket argues for building around it with subscription models. Both bypass the need for government action.

🔄 THE COUNTERINTUITIVE FINDING: HARDER MARKETS = EASIER INTEGRATION

1. The Illusion of Choice Trap

• In mature markets (urban India, South Africa), providers are making enough money doing what they already do — no reason to change

• Patients have the illusion of choice among fragmented providers

• Incumbents have outsized negotiating power and actively resist new entrants

2. Why Rural and Underserved Markets Move Faster

• In difficult geographies, the explicit cost savings are visible to patients: one-tenth the cost of traveling 4 hours for basic blood tests

• Nachiket: “The most traction for primary care models is in the most difficult geographies”

• Patients do the math themselves — the value proposition is obvious when the alternative is sending your son on a scooter for half a day

• Less existing infrastructure means less inertia, less corruption, fewer entrenched interests

3. Scope of Practice as a Market Maturity Problem

• Ethiopia allows health workers to do trachoma surgery — “either you do the surgery or you go blind”

• Alaska’s community health workers perform 150 conditions after one month of training — 50% of GP capacity

• Mature markets constrain health worker scope; harder markets expand it out of necessity

• Key insight from Alaska: “Don’t make doctors in charge. The doctor assists the health worker, not the other way around.”

🔑 Takeaway: Don’t assume integration requires market maturity. The most promising experiments may be in the places that look least ready — because that’s where the financial case for patients is clearest and incumbents are weakest.

⚡ PROVOCATIONS: DOES INTEGRATION CREATE MONOPOLIES?

1. The Monopoly Critique

• Health economist Jishnu Das argues that integration concentrates power, creating monopoly pricing risk

• When a single entity controls primary, secondary, and tertiary care, what stops them from extracting rents?

• Competitive environments mitigate this — but only with adequate regulation

2. Regulated Competition as the Answer

• Israel and the Netherlands made it work: multiple managed care entities competing under active regulation

• It’s no longer called “managed care” — it’s called “regulated competition” to emphasize the regulation alongside the competition

• But this requires governance capacity that many LMICs don’t have

3. Patients as Self-Integrators

• Jishnu’s second provocation: patients in fragmented markets are already integrating their own care

• Going to specialists for diagnosis, getting routine tests at the corner pharmacy, carrying packets of documents between providers

• Market price discrimination works to some degree — better doctors charge more

• But: 30% of adults in Kerala have undiagnosed diabetes/hypertension. 60% C-section rates in Indian private hospitals. “The market is not working.”

4. Kenneth Arrow’s Ghost (1963)

• Healthcare violates the first welfare theorem: patients don’t know what they need

• Unlike buying toothpaste, a cheaper stomach ache treatment doesn’t mean better care — it might mean your failing kidneys were missed

• Saul’s three reasons healthcare isn’t a normal market: (1) information asymmetry too great, (2) third-party payer distorts incentives, (3) “pay more, get less” — you can get remarkably more by paying less

🔑 Takeaway: Integration has real risks — but the alternative (fragmented markets with information asymmetry) is demonstrably worse. The question isn’t integration vs. fragmentation; it’s how to get integration’s benefits while managing its monopoly risks.

🤖 THE AI POSSIBILITY: BOTH/AND, NOT EITHER/OR

1. Digital Twins and Personal Health Companions (Nachiket)

• What if instead of organizational integration, your AI twin provides the continuity?

• Companies like HealthifyMe already have chatbots that know your data, predict risks, nudge you on diabetes medication

• “Insurance for me” — personalized pricing based on actual health behaviors, not population averages

• India already has the DPI rails: payment infrastructure, health IDs (800 million issued during COVID), e-prescription systems

2. The Insurer’s Thought Experiment (Saul)

• “If an insurer was starting today with available technology, rather than with legacy systems, how differently would they construct their businesses?”

• The answer: vastly — abilities would outstrip what they’re stuck upgrading

3. Health Assurance vs. Health Insurance

• Hemant Taneja’s “UnHealthcare” concept: guarantee health outcomes, not just pool catastrophic risk

• Voice-based AI collapses the digital divide — converse in your natural language, no text literacy required

• Personal health wallets aggregating data, predicting risk, intervening at the right time

💡 Key Insight: AI may let us get Jishnu’s patient agency AND Saul/Nachiket’s integration benefits — without requiring the organizational mergers that create monopoly risk. The connective tissue becomes digital, not institutional.

🎯 WHAT CAN PHILANTHROPY DO?

1. Don’t Rehabilitate the Panda Bear (Saul)

• Don’t create funding incentives that don’t survive post-funding

• Treat yourself as a product — find product-market fit before deploying at scale

• “The most expensive way to find product-market fit is to build it first and see if it works”

2. Small Check, Big Risk Appetite (Nachiket)

• Small funders can’t follow the market — but they can shift it

• Build public goods, break the journey into pieces, accept that some bets will fail

• “I don’t want a good fit with the current market. I don’t like the market. I want to switch the market.”

• This requires more thought, more imagination, more collaboration — not more money

3. Practical Advice for Innovators (Saul)

• Don’t be precious about your product — be wedded to the outcome

• “You’re not chopping and changing. You’re learning and adapting.”

• If incumbents lock you out, find adjacent companies with large customer bases — retailers, banks, MNOs — that want to enter healthcare

• Healthcare is 4-8% of GDP in every country. There are companies with data on people that could make money by moving into healthcare.

🔑 Takeaway: Philanthropy’s edge isn’t check size — it’s risk appetite and vision. The role is to catalyze market shifts that neither governments nor incumbents are willing to initiate, by funding public goods, testing unconventional models, and building enabling infrastructure.

📣 CALL TO ACTION

1. Reframe How You Think About Integration

• Stop debating integration vs. fragmentation — the question is how to get integration’s benefits (continuity, aligned incentives) while managing its risks (monopoly, rent extraction)

• The chicken-and-egg problem is real — but HealthForce proved you can engineer integration within fee-for-service

2. Look Where Others Aren’t

• The most promising markets for integration may be the hardest ones — rural, underserved, low-infrastructure

• Patients in these markets see the value immediately because the cost savings are explicit

• Don’t wait for market maturity or government action — build subscription models, organize teams, lower costs

3. Take the Monopoly Critique Seriously

• Integration without regulation is dangerous — but regulation without integration leaves patients to self-navigate broken systems

• Push for regulated competition models adapted to LMIC governance capacity

• Use data portability and patient-owned records as structural safeguards against monopoly lock-in

4. Invest in the Digital Connective Tissue

• AI personal health companions could provide integration benefits without organizational consolidation

• Voice-first AI collapses the digital divide — make this a priority for LMIC health systems

• Build the DPI rails (health IDs, payment infrastructure, e-prescriptions) now — they’re the substrate for everything else

5. Fund Market-Shifting, Not Market-Following

• Small, imaginative bets that build public goods > large bets that follow current incentive structures

• Test product-market fit cheaply before scaling

• Find adjacent-company partners (retailers, pharmacy chains, MNOs) who can provide distribution at scale

Mentioned in this episode:

• Nachiket Mor, “Transforming the Indian Private Sector for Universal Health Coverage” (2025)

• HealthForce (Saul Kornik’s venture)

• Kenneth Arrow, “Uncertainty and the Welfare Economics of Medical Care” (1963)

• Hemant Taneja, UnHealthcare (health assurance concept)

• CareMore Health (US integrated primary care model, founded by two Indian entrepreneurs, grew across 11 states)

• PPO Serve (South Africa, GP-led care coordination)

• HealthifyMe (India, AI-powered health companion with chatbot twin)

• Bito

• Jishnu Das, Georgetown University — work on integration, monopoly power, and information asymmetry

• Israel and Netherlands regulated competition models

• Kaiser Permanente (US, originated as employee care plan at a shipyard)

An analog that recently occurred to me is that of horses. Latest AI models are like wild horses — incredibly powerful, but erratic and unpredictable. You need to break them in, get them to internalize the right instructions-to-actions mapping, and saddle them properly before you can make them useful as a mode of transportation (by that description alone, you can probably guess I know absolutely nothing about horses, so I hope horse enthusiasts will forgive my overture).

Similarly, AI models require the right “harnesses” (a recently popularized term describing the range of tools and capabilities LLMs need to be extended with to give them sufficient context) to become truly useful in a professional setting.

In this post, I want to reflect on my own experience over the past 4 months of building harnesses around my own daily AI workflows, and extrapolate and reflect on what’s actually possible for health systems in low- and middle-income countries with today’s technology.

My own journey: Going from Chatbot to “Second Brain”

As of March 2026, I have six automated agents running on daily and weekly schedules. One triages my inbox every morning, surfacing the most urgent ones and asking for decisions on each one-by-one (archive, reply, defer, task, etc.) before executing on them. Another compiles a daily briefing from my calendar, meeting history, and task list, giving me detailed context on each individual/organization I’m meeting, their backgrounds, our past action points, etc. A third synthesizes research papers and reports from my email and curated lists into my Obsidian knowledge vault, which now contains over 900 interconnected notes spanning every meeting, grant, and partner relationship from the last two years. The system drafts emails in my voice, compiles weekly summaries of progress and helps me prioritize tasks for next week, and even evaluates grant proposals against our strategic priorities. This setup (on Claude Co-work) doesn’t just automate/accelerate what I could do on my own; I routinely find genuinely useful connections and insights I wouldn’t have thought of. In other words, AI has recently become an extension of my brain, not just a chatbot.

This has only become possible because of the rapidly advancing field of “context engineering” and the increasingly sophisticated harnesses that are available. A system prompt (Agents.md/Claude.md) defines my identity and ground rules. Skills (my personal favorite — I have 20+ of them) encode my preferred workflows and thought processes as reusable and highly readable markdown files. MCP connectors plug into Gmail, Calendar, Slack, Fathom, and my note vault in Obsidian. Scheduled agents run tasks on cron. A vectorized memory graph (qdrant) persists key knowledge across sessions.

I no longer manage my 900+ note Obsidian vault manually. Here is Claude helping me organize it in real time!

What is important to remember here is that there are as many different ways to put together your context harness as there are people, because each person’s preference re: tools, workflows, analytical approaches is different to begin with. As long as you keep the principles in mind, you can set this up on any set of tools and platforms.

Turning WHO’s ANC guideline into an AI “Skill”

A Nature Medicine correspondence from February 2026 formally validates this approach for healthcare, defining “context engineering in clinical medicine” as the deliberate design of the conditions under which AI systems operate in clinical care.

As I mentioned, I have found Skills (a structured plain text file format) to be an incredibly powerful, elegant and user-friendly way to encode my thought process and logical workflows to constrain LLMs and make them more repeatable/reliable. So I recently wondered, could we encode clinical decision logic in the same way?

So I took the WHO Antenatal Care (ANC) “Smart Guideline” and asked Claude Code to turn them into a skill so any AI model can follow it step-by-step. In a few short minutes, I had a well structured skill and benchmarks to compare the performance of the latest Claude Opus 4.6 model with vs. without the skill. You can find the skill and all evaluation vignettes and detailed results on GitHub. If you want to read just the skill file itself to see what it looks like, find it here.

The benchmark results are striking. With only one iteration, I was able to get the skill to hit 100% of the performance measures across the 12 vignettes. Of course this is not a production grade Skill; I intentionally chose 15 of the most consequential interventions of a total of 49 elaborated in the WHO kit. Will it hit 100% in a real-world setting? No — AI isn’t deterministic, and edge cases will need human judgment. But even at 90%+ accuracy, the tradeoff is worth serious consideration. It currently takes WHO roughly a year to publish CQL-based deterministic Level 3 (L3) logic from their L2 clinical recommendations. This approach goes directly from L2 structured guidelines (think a spreadsheet) to operational software (L4). The skill file itself is readable by physicians unlike CQL, auditable and customizable by Ministries of Health to adapt to country contexts and limitations, and testable with built-in benchmarks so every local adaptation can be evaluated against clinical performance.

This could dramatically accelerate adoption of Smart Guidelines in LMICs by cutting both the time and cost of the current process. This also shows that LLMs by themselves — as great as they are on the clinical benchmarks — will probably never fully replace the years of effort WHO has put into creating these guidelines.

Any Level of the Health System, One Design Principle

The ANC skill is one example. But the harness design pattern applies across every level of a health system. Each scenario below uses a different combination of harness components, and none requires technology that doesn’t exist today.

The patient navigating insurance. I recently spent three months trying to fill a biologic prescription. The problem wasn’t the medication or the doctor — it was coordination failure between the clinic, the insurer, and the pharmacy. An AI personal care companion connected to a personal health data wallet could have triangulated with the AI agents of those respective systems via Agent-to-Agent (A2A) Protocol or MCP and resolved this in days if not hours.

The community health worker losing patients to follow-up. Amina manages 200 households in Kisumu County, Kenya — 37 pregnant women, 84 children needing immunization, and a stack of paper registers. When Grace misses her third ANC visit, Amina doesn’t notice until it’s too late. A WhatsApp-based agent with longitudinal memory graph and scheduled daily scans could generate a morning brief ranking overdue patients by clinical urgency, analyzing each patient’s history via the ANC guideline skill.

The clinic manager rationing drugs. Nurse Fatima in rural Tanzania fills quarterly drug requisitions from memory and under-orders antimalarials before the rainy season every year. An automated agent with the locally contextualized forecasting logic encoded in a skill connected via MCP to DHIS2 (which already has her facility’s morbidity data) could apply seasonal forecasting, pre-fill her requisition, and auto-generate monthly reports that currently take two days.

The junior resident drowning in discharge summaries. Dr. Kofi at Korle-Bu Teaching Hospital needs eight discharge summaries done at 6 AM after a 30-hour call. Multi-agent orchestration with a FHIR MCP connector could pull each patient’s record, reconcile medications, flag interactions, and generate structured summaries. Stanford’s ChatEHR already demonstrates this architecture.

The district health director missing an outbreak. Dr. Sarah oversees 47 facilities in Kilifi County. She copies last week’s situation report because she can’t synthesize data from all of them fast enough. Scheduled surveillance agents with a knowledge graph accumulating seasonal patterns could detect a cholera cluster weeks earlier. Tanzania’s DHIS2 AI triage system already proved the pipeline works, cutting surveillance triage from 36-48 hours to near-instantaneous.

Assembly required

We don’t need to wait for GPT-6 or MedGemma 5 to transform healthcare in LMICs. ‘Capable enough’ models are already here. The contextual data and/or knowledge exists, in humans if not digitally. The harness infrastructure and tools needed to encode the right context is also becoming mature. MCP hit 97 million monthly SDK downloads and is now an ecosystem asset donated by Anthropic to the Linux Foundation with support from every major AI lab. Skills have also been adopted and becoming natively supported by other AI models beyond Claude.

So we have most of the building blocks. What’s missing is assembly: the deliberate design of context, tools, and workflows around the models. The numbers make the case starkly. LLMs score 84-90% on medical knowledge exams but only 45-69% on practice-based tasks, as reflected in the ANC demo earlier. That gap isn’t a model problem. It’s a context problem. Tools, workflows, and guardrails will bridge the delta between bookish knowing and impacting health with fidelity and consistency.

If you’re building context-engineered health tools, or thinking about encoding clinical workflows as skills, I’d love to hear about it in the comments. And if this resonated, please share it with someone working at the intersection of AI and health systems.

Last month, I sat in on the final demo day of Turn.io’s Health AI Accelerator. Ten organizations — spanning from telehealth support in Pakistan, vision screening in South Africa, chronic disease management in Nigeria, to maternal health in Kenya — had spent months building AI-powered health services on WhatsApp. Different countries, different conditions, different languages. All trying to make AI useful for real patients in real world settings.

After all that work, they converged on the same conclusion: “Prompts, tone, discipline, product constraints, negative prompts >> model cleverness.”

Not a single team said they needed a better model. Every single team said they needed better everything else.

I keep hearing variations of this from every implementer I work with and talk to. And it’s made me realize something that I think the global health AI community needs to reckon with: the model debate is over. The bottleneck has shifted. But many of us haven’t noticed yet.

The myths that won’t die

The AI-for-health conversation is still dominated by questions about model capabilities and externalities. Which model is most clinically accurate? Are they going to be equitable in low-resource contexts and languages? Shouldn’t we be worried about hallucinations and bias? What about energy and water use? These are reasonable questions, and for the most part, they’re no longer the right ones.

Because, firstly, the evidence is in. A comprehensive landscape review published in January 2026 by a Harvard-Stanford collaboration (the ARISE network) found that frontier LLMs now match or exceed physicians on structured diagnostic tasks across multiple studies. The review used the word “superhuman” to describe AI performance in several clinical domains. OpenEvidence hit 100% on the USMLE! Google’s AMIE matched primary care physicians on multi-visit disease management scenarios.

Secondly, hallucinations have dropped precipitously with the latest models (0.7-1.5% for GPT-5 according to one source), and is further mitigated with extended thinking modes and the right “harnesses” or context engineering (like RAG, “Skills”, long-term memory, iterative feedback loops, etc. — in a future post, I'll dig into what these "context harnesses" actually look like in practice, how this might be applied to various global health use cases and contexts, and what my own experience building AI workflows has taught me about where the real leverage is).

Thirdly, the cost and environmental footprint per unit of AI intelligence is falling dramatically with every generation. (33x reduction in energy use in one year | each gemini prompt is one-one-millionth of your daily water use)

tl;dr: The models are good enough, and the externalities are within acceptable limits. What isn’t good enough yet is everything we wrap around them. As the ARISE study itself summarized, prospective studies remain scarce and workflow integration is the binding constraint.

Three stories

Voice transcription based reasoning in the Philippines

A well-known tech entrepreneur I spoke with in 2024 worked with Harvard researchers to deploy an AI system for community health workers in the rural Philippines. The setup was simple and elegant: health workers recorded patient encounters as WhatsApp voice memos. Whisper transcribed the audio — nearly flawlessly even at that time, and despite background noise and colloquial Tagalog. GPT-4 analyzed the transcription and flagged clinical errors. The entire codebase was about 3,500 lines. The cost? Data transfer was more expensive than the AI processing itself!

The system was designed to catch the 25% most egregious clinical errors. It did not try to be comprehensive. It was passive, non-threatening, and focused on the worst mistakes. They learned through trial and error that processing directly on the local language worked better than translating to English first. None of these were model decisions. They were design choices — about where AI enters the workflow, what it’s designed to flag, and how it communicates to health workers who didn’t ask to be monitored.

Pro-forma templates to guide clinical history taking (West Bengal)

Our grantees, Heal India and the Liver Foundation, are currently deploying an AI clinical decision support tool with nurses in rural West Bengal. After extensive user-centered design, they landed on a workflow that addressed an inherent limitation in large language models; that they do not gather patient history comprehensively before providing the diagnostic analysis. To address this, their nurses generate a patient history template using a top AI model, then record the patient interview as they follow the checklist comprehensively, then send the recording with detailed patient history to a carefully prompted Gemini “gem”. Within 60 seconds — even from local Bengali dialects, even in noisy clinical settings — they get back a structured case summary, syndrome-based differentials, and a sophisticated treatment plan.

The model was table stakes. What took months was the workflow design. Figuring out that the entry point should be a structured pro-forma. That recording works better than typing. That 60 seconds is an acceptable latency threshold. That nurses, not ASHA workers, are the right users because they have the minimum medical literacy to act on the output. Most remarkably, the nurses have now started using the AI to create their own training materials — snakebite treatment protocols, for example — without being asked. The tool became a platform for local knowledge creation, not just a diagnostic aid.

Jacaranda’s use of contextual data for fine-tuning (Kenya)

Jacaranda Health built Uliza Mama, a maternal health chatbot that serves Kenyan mothers in Swahili, English, and code-mixed language. They built a 3,000-question benchmark from 300,000 real user interactions. And yes, their custom fine-tuned models outperformed general LLMs.

But the lesson here is not that off-the-shelf models are bad. It is that fine-tuning was effective because of context. Jacaranda’s advantage was 300,000 real conversations with Kenyan mothers — their fears, their language patterns, the clinical pathways that actually exist in Kenyan maternal health. Early versions of the chatbot based on off-the-shelf models produced outputs Jacaranda’s team described as “schizophrenic” and “terrifying.” Months of work on safety mechanisms, cultural sensitivity, and balancing medical accuracy with accessible language followed. No model upgrade would have fixed what they were dealing with.

What “context” actually means

When Turn.io accelerator’s participant teams talked about what mattered more than model selection, they kept coming back to the same list: prompts and negative prompts (what not to say is as important as what to say); tone and cultural register (clinical precision doesn’t equal trustworthy communication); local clinical guidelines (not WHO global recommendations, but the actual drugs available and the actual referral pathways); trust architecture (disclaimers, escalation to humans, clear boundaries about what the AI can’t do); workflow design (when does the AI enter the care encounter — before, during, after? passive or active?); safety and evaluation infrastructure (hallucination detection, clinical red lines, ongoing performance monitoring, not just one-time validation); and business model alignment (for private providers, the AI has to fit how they actually make money).

This is a long list. And not a single item on it is a model capability question.

What funders and builders should do about it

If context is indeed the binding constraint, then the field needs not just open source and downsized models that can perform well on edge devices and offline; it also needs an implementation commons — shared playbooks, real-world contextual benchmarks, failure libraries, and design patterns. From Endless, we have started working on this through our partnership with The Agency Fund (TAF) on federated benchmarking (10 organizations are being funded to publish benchmarks based on real-world operational data), and with CHAI and TAF on stewarding an AI commons with other stakeholders. The Center for Global Development and TAF’s four-stage evaluation framework offers a useful structure for thinking about where in the lifecycle implementers need the most help, and their living playbook is a good design primer. I also wrote extensively about various implementation bottlenecks in this post last year.

But you don’t need centralized platforms to start. AI model companies can expose some of their internal tools for benchmarking and evaluating the entire product harness, not just models, so that implementers have better tools at their disposal. Implementers can begin by publishing what their prompt libraries and RAG pipelines and Skill harnesses look like and what didn’t work. Sharing benchmarks built from real user interaction data, the way Jacaranda shared their 3,000 questions. Documenting workflow decisions and the reasoning behind them. Publishing product cards to complement model cards. Open-sourcing the non-model layers. Rigorously evaluating the product, user experience and health impact. Philanthropic funders like us can incentivize this kind of open knowledge creation and sharing as part of our agreements with grantees.

As of 2026, all organizations are starting their journey with superhuman models with excellent clinical performance and incredible multi-modal and agentic capabilities (I would argue that OpenAI’s embarrassing public gaffe with the ChatGPT Health product is more of a context failure than a model failure). Those organizations that understand their context best will have a decisive advantage to succeed. And the faster they share what they’ve learned, the faster the whole field will move forward.

The shifting sands

We’re past the era where AI capability was the binding constraint for healthcare in low-resource settings. We’re now in the era where implementation fidelity and design wisdom are the binding constraints. The models will keep getting better on their own. The context — the prompts, the workflows, the trust, the cultural fit — only gets better when humans do the hard, unglamorous work of deploying these tools and documenting what they learn.

The question isn’t any longer whether AI is ready for global health. It’s whether global health is ready for AI.

If you found this useful, I’d love to hear from you. Are you seeing the same pattern in your work? Where do models still fail? What context challenges have been hardest to solve? Drop a comment or reply to this email.

WHO estimates a projected shortfall of 11 million health workers by 2030—including physicians, nurses, midwives, community health workers, laboratory technicians—mostly in low- and lower-middle income countries. Furthermore, the total obfuscates gross inequities of their distribution: Sub-Saharan Africa, with 11% of the world’s population and 25% of the global disease burden, has only 3% of the global health workforce. There are similar inequities within countries, with rural areas often having as few as one-eighteenth the physician density as urban areas.

On the surface, it seems that the obvious solution is to build more medical and nursing schools and churn out as many doctors and nurses as possible to bridge this gap. Indeed, that is the formula many countries have espoused and largely followed.

In this piece, I will argue two things:

(A) that this approach is inherently flawed and does not lead to expected outcomes for a multitude of reasons, and often creates adverse unintended effects.

(B) that the core assumption of physician (=“clinical intelligence”) scarcity that health systems of today are designed around is no longer true, and there is an opportunity to fundamentally reimagine how we train, deploy and organize our health workforces to meet the needs of the 21st century, which promises to achieve much better health outcomes and patient experiences at a fraction of the cost.

Why training more clinicians isn’t working

1. Time and cost constraints make traditional expansion unfeasible. Medical education requires 5-8 years plus specialty training, while nurses need 3-6 years. Yet the urgency is immediate—NCDs are already killing 32 million people annually in LMICs (74% of global NCD deaths), with 82% of premature NCD deaths occurring before age 70 in these settings.

   Moreover, training a single physician is expensive, with costs ranging from $21,000 (Uganda) to $264,000 (US). At current production rates, many LMICs would need decades and billions in investments to close their workforce gaps through traditional medical education, and that’s before accounting for attrition.

2. Brain drain undermines all investments: Every year, Ghana loses half of its newly graduated doctors within 4.5 years. Nigeria has seen approximately half of its licensed physicians emigrate, and Philippines has had a similar experience with their nurses. Sub-Saharan African countries have lost more than $2 billion from training doctors who then migrated to just four destination countries: Australia, Canada, the United Kingdom, and the United States (recipient countries have benefited ~$4.55 billion in saved training costs, creating perverse incentives that perpetuate the crisis).

3. Even doctors who stay cluster in urban areas: The fundamental access problem is not just quantity, it is also distribution and retention. Of Kenya’s 5000+ registered physicians, only 600 work in public hospitals. Vacancy rates for rural positions in Bangladesh can range from 40-80%. Even this appalling number may hide the actual reality — in Bangladesh, I witnessed how public hospitals in rural areas would have physician positions filled on paper, but the actual doctors would practice part-time in next-door private clinics or not show up at all. Even well meaning efforts by the government, such as setting up biometric logins for hospitals, were met with sabotage; most of the sensors were found to be broken within weeks.

4. Qualifications ≠ Quality: Despite conventional wisdom, it seems credentials are not always a good proxy for quality, and the presence of doctors doesn’t necessarily mean better care. In a groundbreaking (and controversial) study by Prof. Jishnu Das and colleagues in India, public sector doctors spent only 2.4 minutes with patients, only gave correct diagnoses in 4% of cases, and prescribed unnecessary or harmful treatments 74% of the time. In fact, these fully trained MBBS doctors showed no diagnostic advantage over informal and untrained practitioners. More tellingly, the same doctors in private practice performed 0.76 standard deviations better on quality metrics.

   In contrast, there is a multitude of emerging evidence that task-shifting to lower skilled providers can deliver quality care. For example, nurse driven care has been found to be “as safe or safer” than physician-led care for mortality and hospital admissions, and led to longer consultations and greater patient satisfaction. Other studies have found similar results, and at least one systematic review has confirmed that task shifting for chronic disease management, screening, patient education, and even certain diagnostic procedures achieves equivalent or superior patient outcomes at a fraction of the cost.

5. Doctor lobbies throttle production of semi-skilled providers: While points 1-3 above are widely known, here’s an uncomfortable truth rarely discussed: the medical establishment has systematically restricted the supply of other health providers by invoking quality concerns (irrespective of evidence to the contrary). Using their control over medical accreditation bodies and Ministries of Health formally, and using strong lobbies through their unions and associations, they block production of accredited but less skilled providers, including nurses (in 2023 alone, AMA defeated 80+ bills across US states aimed at expanding scope for nurses, pharmacists, and physician assistants). I faced this firsthand in Bangladesh in my effort to accredit and mainstream private pharmacy practitioners (who happen to provide 70% of primary care in the country); I was often humiliated and thrown out of meetings because I was “teaching thieves how to steal better.” The same systematic review cited above also found that the biggest barrier to task shifting was onerous regulations and restrictions on prescribing medicines by semi-skilled providers. And the main reason this argument has been so persuasive and effective so far in maintaining strict gatekeeping has been legitimate concerns about clinical rigor and patient safety.

6. Clinicians aren’t best suited to address 21st century health needs: With ageing a major crisis even in Africa, and NCDs surpassing infectious diseases as the major cause of death across the world, clinical medicine is no longer sufficient in addressing the health needs of the 21st century. What we need is a population health management approach — more proactive follow-ups, medical adherence support, lifestyle counseling and behavioral therapies, patient education, mental health support, and high quality palliative care. Yet, fifty years after Alma Ata promised people-centered primary care, health systems remain primarily configured for acute, episodic care delivered by clinicians.

The democratization of “clinical intelligence”

Every health system ever designed to date has operated under one fundamental assumption: clinical intelligence (the knowledge and skill required to diagnose and treat diseases) is scarce and expensive. This scarcity shaped everything—from care team composition to scope-of-practice regulations to care protocol design. But artificial intelligence is shattering this core constraint.

Over the past 3 years, AI has rapidly improved to the point where it now beats the average human doctor on almost all clinical reasoning tasks. In direct comparison studies, AI outperformed physicians on rare disease diagnosis with 78.9% accuracy versus 26.3% for physicians. In 5 years, by the time a freshman medical student of today will graduate from medical school, AI will likely surpass the ability of the best human doctors, with up-to-the-second medical knowledge and infinite memory (OpenEvidence recently scored a superhuman 100% on the famously difficult USMLE exam).

More impressive still is the cost trajectory. The cost of AI inference—producing medical-grade clinical reasoning—has fallen 240x in just 18 months, from $90 to less than $0.4 per million tokens (roughly 750,000 words of medical text). The trend line is clear: the cost of clinical intelligence is dropping at least 10x per year and rapidly approaching the marginal cost of electricity. At this rate, it will soon become a utility that can be metered and deployed at will anywhere there is an internet-connected device (even on-device, offline intelligence through small language models is rapidly becoming feasible). Importantly, these AI systems can operate 24/7 without fatigue, emigration, or urban bias, and is already showing promise in reducing clinical errors in real-world settings.

We are therefore witnessing the fundamental economics of healthcare flip on its head. This isn’t about replacing doctors, who aren’t reliably present in most places anyway—it’s about democratizing clinical intelligence anytime, anywhere. When a community health worker in rural Kenya or a rural pharmacist in Bangladesh can access the same diagnostic reasoning as a Harvard-trained physician, we have the opportunity to completely reimagine the composition and function of the health workforce.

From Workforce Scarcity to Workforce Abundance

If clinical intelligence becomes abundant and cheap, we can completely rethink workforce composition from first principles. Instead of asking “how do we train enough doctors?” we can ask “what are the various functions needed to be performed by health systems to achieve the outcomes we want, and who can deliver those tasks most effectively and safely if augmented by AI?”

This simple but fundamental reframing could enable us to think out-of-our-current-box and transform our workforce scarcity to potential abundance.

Firstly, consider already existing infrastructure we’re currently ignoring:

1. Private drug shops and pharmacies: In contexts as diverse as Bangladesh and Nigeria, there are hundreds of thousands of providers delivering 70%+ of primary care. That’s millions of touchpoints with established trust, existing supply chains, physical infrastructure, and—crucially—sustainable business models that don’t require government salaries. These pharmacies are already doing clinical triage and basic diagnosis, just without oversight or tools to do it well. Now imagine equipping them with AI clinical decision support. In India, Wadhwani AI has extensively tested tuberculosis screening based solely on cough sounds, achieving sensitivity comparable to traditional diagnostics. A pharmacy worker with a smartphone could easily screen for TB without lab equipment or specialized training, as soon as a patient walks in with a persistent cough.

2. Informal providers: Every country has traditional healers, midwives/doulas, village doctors, faith healers, and other informal practitioners who communities trust but health systems ignore or actively oppose. Rather than fighting them (a battle we’ve been losing for decades), we could formalize their role with clear scope of practice, equip them with AI tools, and create oversight mechanisms that actually work. When a traditional healer in rural Ghana can use an AI diagnostic tool and knows when to refer to formal facilities, they become part of the solution rather than being marginalized or pushed underground.

3. Community volunteers: There are numerous young people, women and the elderly in every community who volunteer time to support their communities. Supporting health needs in their communities would be an attractive and respectable undertaking for many of them.

4. Family caregivers: Perhaps the most overlooked resource of all. Literally millions of people—predominantly women—provide unpaid care to elderly parents, chronically ill family members, children with disabilities. Equipped with AI health companions, they could do far more than physical caregiving. They could assist with lifestyle and behavior change, ensure medication adherence, monitor vital signs, triage symptoms for escalation, and provide the consistent daily support that chronic disease management requires (Noora Health’s training programs already show that family members can be empowered as caregivers, resulting in 89% feeling more prepared and confident, with one hospital seeing a 24% reduction in readmissions for heart surgery patients). My mother helping my dad take his diabetes medications on schedule isn’t currently considered “healthcare.” But if she had an AI tool that reminded her when he missed a dose, explained why adherence matters, and tracked his blood glucose trends, she would be performing a critical health function.

Secondly, imagine completely new cadres of health providers of all different scopes of practice that do not exist today, like:

• A special needs nanny, who specializes in caring for children with learning disabilities while babysitting.

• A health navigator, who knows how best to guide patients through the confusing processes of choosing providers, securing appointments and navigating insurance.

• An adherence coach, who specializes in motivational interviewing techniques to get patients to comply with their treatment regimen.

• An end-of-life carer, who can provide the at-home emotional support and handholding needed for terminally ill patients.

This isn’t about replacing formal health workers. It’s creating a distributed, decentralized ecosystem where many health actors, each operating within their scope, collectively deliver comprehensive care. When you add up all these potential contributors, the 11 million worker shortage becomes an eminently solvable problem of coordination and enablement rather than an impossible gap in formal training capacity.

Towards a “Gig Economy” of Health Tasks

We could take this even one more step further, beyond traditional employment models entirely. Why shouldn’t family members or community volunteers be compensated for providing health system functions that save the health system downstream costs and improve outcomes?

Key features of this model could be as follows:

• Stackable and verifiable micro-credentials on secure digital wallets: A 32 year old woman in peri-urban Dakar, Sarah, who takes care of her hypertensive mother at home, can complete a 20-minute online course on blood pressure measurement and earn a digital credential in her cryptographically secured “health competency wallet.” She can now perform blood pressure monitoring for any patient in her neighborhood, not just her mother. The local pharmacist similarly completes training modules on diabetes management, TB screening, and mental health first aid—each adding capabilities, earning credentials and expanding scope of legitimate practice.

• Task-based compensation: Rather than traditional employment, participants earn micro-payments for specific functions. Sarah receives a $1 per week stipend for regularly monitoring her mom’s blood pressure (verified by photo upload), which is far less costly for the health system than a CHW dispatch, but is still meaningful supplemental income for Sarah. She additionally gets a bonus if her mother’s hypertension stays controlled for three consecutive months. The pharmacist similarly earns a fee for each TB screening conducted, with bonuses for successful referrals that result in diagnosis and treatment initiation. Platforms like CommCare Connect are already proving that this is possible, reaching 77,000+ clients and distributing $161,000 in verified payments to gig-working CHWs while demonstrating 97% worker acceptance rates for autonomous learn-deliver-verify-earn models.

• Open access platforms: Instead of proprietary systems, imagine open digital infrastructure—a “healthcare operating system”—where any trained person can plug in, see patients who need services they’re credentialed for, complete tasks, and get compensated. Think Uber for health, but designed with appropriate safeguards, quality assurance, and outcome accountability.

• Progressive capability building: As participants complete more training and demonstrate competency through successful cases, they unlock additional scopes of practice. A caregiver who excels at medication adherence support might add chronic wound care to their capabilities. A pharmacist who demonstrates excellent TB screening performance might be trained and allowed to initiate DOTS treatment under remote physician supervision.

This creates multiple benefits simultaneously:

1. Scalability: You’re no longer constrained by medical school capacity or government hiring bottlenecks. Anyone with basic literacy and a smartphone can start with entry-level health functions and build from there.

2. Flexibility: Health workers can work part-time, seasonal, or full-time depending on their circumstances. A high school graduate can gain competency and real-world work experience while waiting to go to college. A young mother can do medication adherence support while caring for her own children. A retired teacher can become a health educator in their community.

3. Retention: The brain drain problem largely disappears. You’re not training expensive specialists who then emigrate. You’re building capability within communities through stackable, portable micro-credentials that create local economic opportunity.

4. Quality assurance at scale: Every interaction generates data. AI monitors quality continuously, flags concerning patterns, requires additional training when performance drops, and identifies high performers for advanced responsibilities. This oversight—which providers must agree to as a precondition for getting compensated—could be designed to be far more rigorous than current systems where doctor performance goes unmeasured behind closed doors.

The Economics of the Paradigm Shift

The financial logic is compelling. For the net cost of training and effectively retaining a single doctor, a health system could afford tens of AI-assisted community-based health providers (CHWs, pharmacies, etc.) or hundreds of gig workers playing micro-functions.

Beyond the obvious benefits to healthcare experiences and outcomes, it can also support economic development. Health sector employment is countercyclical (recession-resistant) and creates 3.4 downstream jobs for every direct health position. The jobs created through this model are also flexible, humane and equitable:

• Youth-friendly: No need for 8 years of expensive education. A motivated 20-year-old can start earning within weeks while building capability over time.

• Women-friendly: Flexible hours, work from home options, and culturally appropriate roles (family caregiving, peer counseling) create opportunities in contexts where women face employment barriers. AI-enabled workforce expansion could formalize, scale, and fairly compensate millions of women already doing health work informally.

• Dignified: These aren’t low-skill, menial, repetitive jobs—they require compassion, judgment, and interpersonal skills. But they’re accessible without gatekeeping credentials.

• Locally rooted: The economic benefit stays in communities rather than concentrating in urban hospitals or leaking abroad through brain drain.

If we invest in building this decentralized, AI-enabled health economy in earnest, we could realistically create 30-40 million jobs globally over the next decade—far exceeding the 11 million “shortage” while simultaneously improving health outcomes and care experiences for the same cost to the taxpayer wallet.

Redefining the physician’s role

This doesn’t mean physicians become obsolete. Rather, their role evolves dramatically. In this reimagined system, physicians become:

• In-person diagnosticians and interventionists: Whenever cases require complex physical examinations or procedures/surgeries that cannot be done by a lay worker or even nurse. Emergency rooms will likely require doctors for the foreseeable future.

• Specialist consultants: Handling complex diagnostic dilemmas and judgement calls that even advanced AI cannot make, such as deciding whether to initiate a treatment for a patient who does not have access to or able to afford the right diagnostics, or whether to recommend a patient for high-risk surgery, or how to adjust protocols to resource constraints. One specialist could support dozens of community health centers through telemedicine, reviewing the flagged complex cases rather than spending time on routine hypertension follow-ups.

• Clinical supervisors: Reviewing AI performance, auditing care quality across the distributed ecosystem, and providing ongoing training to frontline providers. Think of it as quality assurance at scale—one physician might oversee care delivered by 100+ augmented workers.

• System designers: Working with AI developers to improve algorithms, designing care protocols, and ensuring clinical validity of automated decision support.

• Researchers: With millions of AI-documented encounters generating rich data, physicians can focus on discovering what works, testing interventions, and advancing medical knowledge.

In Costa Rica’s EBAIS primary care model, which achieved remarkable health outcomes (including a 13% reduction in all-cause mortality), the physician-to-patient ratio was much lower than in typical systems. But what they had was strong community health teams with clear protocols, comprehensive empanelment, and continuous follow-up. Now imagine that same model supercharged with AI and expanded to include pharmacies, caregivers, and other community actors in a coordinated ecosystem.

The North Star

The most important mindset shift for policymakers is this: stop thinking from infrastructure forward and start thinking from health outcomes backward. Finance ministries and health system leaders must recognize that the traditional training and delivery models are not effective to begin with, and certainly not scaling fast enough to meet the health needs of the 21st century. Aging populations, rising chronic disease burden, and persistent inequities require fundamentally different approaches.

Every cervical cancer death prevented, every diabetic complication avoided, every medication adherence improvement achieved matters regardless of who delivers the intervention. A trained community member using AI decision support to detect hypertension achieves the same health outcome as a physician making the same diagnosis. A family caregiver following AI-guided protocols for post-surgical care produces the same recovery benefits as a visiting nurse. A patient self-managing diabetes with AI coaching can achieve better long-term control than sporadic physician visits.

Practical considerations and call to action

This transformation won’t be simple. There are a lot of pieces to figure out — decentralized measurement and auditing of outcomes and quality, the infrastructure for open access learning and credentialing, and the AI orchestration engine or “operating system” to govern all of this, among a few.

This paradigm shift also cannot happen overnight. Medical professional associations will resist scope-of-practice expansions. Regulators will demand safety evidence. Traditional health system leaders will question quality assurance for non-traditional providers. Patients will require time to build trust on AI-enabled providers.

But it’s far more achievable than scaling medical education by an order of magnitude while overcoming brain drain and urban concentration. Here are some pragmatic next steps for different stakeholders:

To policymakers and health officials: Create regulatory sandboxes for AI-augmented care. Measure outcomes rigorously. Stop protecting incumbent interests when evidence supports better models. Your legacy will be defined by whether you had the courage to break from convention.

To funders and investors: Stop financing the construction of medical schools that will graduate their first doctors in 2035. Start backing AI-augmented community health models that can scale now. Demand sustainability from day one. Don’t just fund point innovations, but rather the operating system and unsexy “plumbing” infrastructure stack that will make everything fit together.

To medical educators: Reimagine curricula for the AI age. Emphasize behavioral approaches and relationship building. Prepare medical students for roles as surgeons, supervisors and consultants, not primary care gatekeepers. Design micro-modules and credentialing systems for non-physician providers based on health system functions and 21st century needs.

To technology builders: Develop AI systems that work offline, on low-cost devices, for diverse populations. Make them interoperable, not proprietary. Train on representative data. Build in accountability and transparency. Partner with community providers, not just academic medical centers.

To health workers and professional associations: This transformation is coming whether you embrace it or resist it. Those who adapt early—learning to work alongside AI, advocating for appropriate regulation, helping design better systems—will shape the future. Those who cling to monopolies and outdated scope-of-practice laws will eventually be bypassed.

Conclusion

The 21st century health workforce won’t look like the 20th century’s, any more than the 20th century’s looked like the 19th century’s. The question is whether we’ll design this transition intentionally and ride the AI wave with equity and outcomes at the center, or whether we’ll be crushed under the tsunami of disruption.

The need is urgent. The foundational technologies already exist. The evidence base is emerging. What’s missing so far is the openness to unlearn and the courage to reimagine.

Let’s not waste this moment.

What do you think? Are there legitimate concerns I haven’t addressed? Examples of this transformation already happening that should be highlighted? Push back on my arguments, share your experiences, or point to blind spots I’m missing. This conversation is too important for any of us to get complacent.

Leave a comment

If I may say so myself, every minute of the hour is filled with profound insights! I hope it sparks ideas, challenges assumptions, and opens up new pathways. Let me know in the comments what resonates.

CONVERSATION SUMMARY

🩺 THE PROBLEM: THE NCD PARADOX

1. Massive Burden, Minimal Resources

• 70% of NCD deaths occur in LMICs, yet only 2% of global health funding addresses them

• By 2030, NCDs will be the leading cause of death in Africa

• 45 million annual deaths globally, 18 million premature (under 70)

2. Systems Built for Episodes, Not Continuity

• Healthcare conceptualized as "sick care"—engagement only when symptomatic

• Rob's stroke patient: catastrophic endpoint of 20-30 years of missed prevention

• Patient with BP 150: took meds for a week, thought he was "cured"

3. The Double Burden of Poverty

• Poor populations face disease plus impossible economic trade-offs

• Prevention requires investments that compete with daily survival

• 90-95% of African patients pay out of pocket—patients are also payers, so affordability of prevention is key!

🔑 Takeaway: We've built systems that wait for catastrophe rather than prevent it, treating discrete episodes instead of managing lifelong conditions.

🔧 SOLUTIONS IN PROGRESS: TWO GROUNDBREAKING APPROACHES

1. BP Sawa – Penda Health

A WhatsApp-enabled hypertension management program treating chronic care as a "sales funnel":

• 70% remote connection rate with patients

• 40-50% return for follow-up blood pressure checks

• Continuous optimization at each touchpoint

• Proactive follow-up drives both loyalty and health outcomes

2. Population Health Management Framework – Ariadne Labs

Based on Costa Rica's success, a simple but powerful approach:

• Empanelment: Know your population/denominator

• Assessment: Measure outcomes across the population

• Management: Intervene to improve

• Iteration: Continuous measurement and improvement

Can be implemented at any level—from individual provider to national system. In fact, it works best where every level is working towards an aligned goal, kind of like OKRs.

💡 LESSONS & INSIGHTS

1. The Sales Funnel Revolution

• Rob: "Old school public health says 'lost to follow up, too bad.' With a sales hat, you ask: what different message could we send?"

• Each touchpoint optimized like a tech company would

• Business incentives align with health outcomes because proactive follow ups are welcomed by patients and creates loyalty and repeat visits

2. Costa Rica's Pencil-and-Paper Miracle

• 13% all-cause mortality reduction

• 19% cardiovascular mortality reduction

• 22% elderly mortality reduction

• Achieved through commitment and care culture, not technology.

3. Relationships Over Transactions

• Providers need to understand what motivates each patient (picking up grandkids? artistic pursuits?)

• Patients felt cared for when followed up via WhatsApp or a voice callk

• 5-year journeys to control A1C more rewarding to providers than quick antibiotic fixes

4. Culture Beats Financial Incentives

• Costa Rica rejected performance pay as "commodification" of care

• Professional pride and peer coaching outperformed monetary rewards

• Bottom 20% of clinics get coaches, not penalties

🧱 CHALLENGES & BOTTLENECKS

1. The Timeline Problem

• NCDs require 10-30 year horizons to show mortality impact

• Funders want results in 3-5 years

• Political cycles don't align with prevention timescales

2. Provider Training Gaps

• Medical schools don't teach motivational interviewing

• Providers trained for diagnosis, not longitudinal relationships

• New skills needed: counseling, behavior change, patient engagement

3. Fragmented Financing

• Vertical disease programs crowd out integrated care

• Kenya spends only $80-90 per capita annually on health

• No sustainable financing model for prevention

4. Demand Generation Void

• Patients don't understand chronic disease management

• No public education campaigns at scale

• Providers can't create both apparatus and demand simultaneously

🔭 VISION: WHAT 21ST CENTURY NCD CARE LOOKS LIKE

• Every provider practicing population health: Individual clinicians asking "who needs something today?"

• AI-powered personalization: Messages adapted to language, tone, emoji preferences—90% resonance vs 30%

• Cognitive offloading: AI handles clinical decisions so providers focus on relationships

• Radical cost reduction: Hypertension management at $1/month

• Proactive by default: Systems that reach out, not wait for catastrophe

• Trust-based care teams: "By foot, by boat, by horse, providers will arrive"

🎯 Paraphrased from Madeline: "Empanelment is the linchpin. We can't be proactive without knowing who we're responsible for. It's about providers knowing their patients and patients knowing their providers."

📣 CALL TO ACTION

1. Reframe the Mental Model

• Stop thinking episodes, start thinking about the person behind the patient

• View chronic care management as behavior change at scale

• Apply "sales funnel" thinking to patient engagement

2. Start Where You Are

• Any provider can do population health management with their panel

• Don't wait for perfect systems—use WhatsApp, spreadsheets, or simply pen and paper

• Measure what matters: controlled BP, not just diagnosed BP

3. Advocate for System Redesign

• Push for empanelment as foundation of all care

• Demand public education campaigns on NCDs

• Challenge vertical programming that fragments care

4. Leverage Technology Wisely

• Use AI for personalization and cognitive offloading, not replacement

• Start simple (IVR, WhatsApp) then layer complexity

• Remember: tech supercharges good systems, doesn't fix broken ones

5. Change the Financing Conversation

• Stop searching for silver bullet payment models

• Explore new financing models like value-based care that aligns incentives with patient experiences and outcomes

• Focus on driving costs down to affordable levels

• Blend grant funding for setup with sustainable business models

Watch the full conversation for deeper insights into BP Sawa's implementation, Costa Rica's remarkable success, and practical strategies for NCD management at scale.

Disclaimer: The views expressed here are my own and not necessarily those of Endless.

The progress of the models in medicine has indeed been astonishing. OpenAI's o1 preview demonstrated “superhuman performance” surpassing top physicians on clinical reasoning tasks months ago; it’s successor, o3, is already faster, cheaper, and more accurate. Hallucination rates, among the biggest concerns about using LLMs in medicine, has declined dramatically to negligible levels in cutting edge models. Google’s recently released MedGemma now offers multimodal clinical reasoning (including dermatology and radiology images) in an open-source model, rivaling state-of-the-art models from mere months ago. Models are progressing so fast that research is becoming obsolete soon after being published, and new benchmarks like HealthBench are popping up monthly to extend the goalpost. While it is not easy to appreciate the improvements in medical reasoning intuitively, here’s a visual analog: AI video generation models can now turn a single text prompt into HD video clips with sound that is indistinguishable from reality—a mind-bending leap from the bizarre and muted clips of Will Smith devouring spaghetti from just two years ago.

AI adoption has similarly accelerated at a pace unprecedented in the history of technology. Extrapolating from OpenAI’s numbers indicates that close to a billion people chat with an LLM every week, with up to 10% of those conversations focused on health. My own parents, aged 73 and 66, told me recently they're talking to ChatGPT regularly about their health issues (and everything else, like gardening), with zero encouragement from me. I've been doing the same for over two years, and several friends got life-changing benefits managing chronic conditions, children’s autism, and more. The barriers to self-care have never been lower: voice input in local languages (my parents use Google keyboard’s amazing voice-to-text feature in Bangla), 24/7 availability, free to use within reasonable constraints, infinite patience and attention to detail. While LLMs are far from infallible, people seem to intuit that the average human provider is likely much worse in accuracy, and certainly can’t compete on speed and availability.

The risks of health delivery organizations — public health non-profits, national public hospitals, or private clinics and diagnostic centers — not keeping up with this revolution is manifold. Health data is being fragmented and walled inside private non-health organizations like OpenAI and Google without guardrails. National and local clinical guidelines are likely being disregarded. There are no ways to monitor and mitigate when LLMs inevitably provide harmful and even life-threatening advice (such as the mental health downward spirals ChatGPT is sometimes causing). Care from a typical pharmacy or community health worker, which is how billions of people across the world receive primary healthcare, looks laughably unsophisticated when compared to the advice one can get from an LLM from the comfort of their home — as a result, actual providers are further and further distanced from people’s healthcare, and utilization and trust in them is bound to erode further.

I therefore find it baffling that health systems are not taking this with the seriousness or urgency it deserves, and in most cases not even doing anything differently than what they did pre-ChatGPT. I have the privilege of working with some leading-edge innovators and adopters of AI through my work at Endless, such as projects with Penda Health in Kenya, E-Health Africa in Nigeria, and Dimagi in Zimbabwe, but beyond these early adopters, most non-profits and social enterprises, and especially traditional hospital and public health systems, are scarcely considering the implications and reckoning with this historic disruption.

After spending the last year talking to dozens of model developers, health organizations, hospital systems and funders of AI pilots, I've identified eleven specific (but often interrelated) barriers that explain why the AI adoption by our systems are so far behind the people they aim to serve.

Bottlenecks slowing AI adoption

To make this digestible, I've organized the barriers into three layers, going from the internal (psychological and knowledge gaps) and zooming out all the way to systemic issues.

Below is a TL;DR summary table of each of these bottlenecks and some pragmatic suggestions on how to overcome them in the near term. Subsequently, I will elaborate on each of them if you are interested to read more.

A. INTERNAL BARRIERS

1. Mistrust and fear of AI, and the infamous human ego

AI is poorly defined, and even more poorly understood. Given the myriad doomsday forecasts out there, it is hard to have a focused constructive conversation with a public health professional or doctor about the true promise of AI in medicine (i.e. the democratization of clinical intelligence), and how to manage the genuine risks that inevitably accompanies that superpower.

Many doctors in the West, overburdened for decades by documentation and process-creep, are generally positive about AI’s potential to simplify life and help them focus again on patient care, partly also because they don’t see an imminent threat to their jobs under the onerous regulatory environments they operate in. But many senior clinicians in poorer countries, who also tend to lead Ministries of Health, live in this complacent bubble of overconfidence and skepticism that their craft is far beyond any computer to replace, and hence are often unable to take the disruption and opportunities seriously. As a friend at Harvard Medical School who was a leading Internal Medicine specialist in India told me, "Even the free version of 4o is better than most doctors I've encountered in developing countries." Yet decision-makers continue operating from outdated assumptions about AI capabilities.

This is also why they still over-index on AI hallucinations being a dealbreaker (which is a much smaller problem today than it was 2 years ago), declare that AI is not ready after a small pilot with the outdated GPT-3.5, and latch on to any indications that AI is not “true intelligence.” I saw a version of this when I ran Jeeon, where senior doctors in Bangladesh declined to even acknowledge the widely researched fact that 70% of healthcare in the country was delivered by the humble neighborhood drug shop.

To be fair to doctors, however, this inflated ego is a broader problem with human experts in general!

2. Limited AI Literacy and artificial self-imposed constraints

Ego and mistrust naturally prevents people from deeply and objectively engaging with the promises of AI. There are serious limitations in people’s understanding of LLMs and their capabilities, and many clinicians and public health professionals use it (when they do) just as they would Google Search, and no better.

As a result, even when organizations undertake efforts to pilot AI, they tend to get stuck within the “chatbot” paradigm. As Rob Korom from Penda Health told me, “it is very easy to build a cool chatbot by layering a 3-sentence prompt over ChatGPT”, and people tend to stop innovating beyond that novelty. But chatbots alone can rarely solve a real problem in healthcare. I will cover this issue and available design approaches in more detail in points #4 and 5 below.

On top of this, organizations often impose artificial and poorly informed constraints on themselves which lead to failure, such as choosing outdated models like GPT-3.5 or open source ones citing cost or data protection constraints. What they often fail to understand is that costs for an equivalent level of performance are coming down at 10x per year or more. Recently, only months after releasing o3-mini, OpenAI slashed API prices by 80%! Open source (e.g. DeepSeek) is also catching up fast and is probably at most 6 months behind the frontier.

By the time pilots are concluded (ideally with the best frontier models, even closed ones) and funding is secured for scale up, costs are bound to come down to an acceptable level, and open source models available that can replace the closed models without little or no performance tradeoff. But piloting a substandard model is certain to lead to underwhelming results, dampening excitement and adoption.

B. ORGANIZATIONAL BARRIERS

3. Devices and connectivity

There's an incorrect assumption that smartphones and 5G connectivity are prerequisites for AI deployment. Yet even basic setups—like those demonstrated by Viamo in Nigeria—can effectively utilize AI through innovative design:

This is therefore more of an internal mindset barrier than an actual blocker to progress. AI can be the most democratizing health technology of all times, and it is incumbent on us to take it to as many people as possible, starting with the lowest hanging fruits. We should start with facilities and providers (hundreds of thousands by this point) who already have devices and connectivity. Billions of people rely on CHWs or pharmacies who would get an instant boost in care quality if AI was used to upgrade these touch-points, even with phone/IVR/Whatsapp based solutions like Viamo’s. We should not let perfection be the enemy of the good.

4. Unclear goals and success criteria (“Hammer in search of a nail” problem)

Because AI is all the hype these days, and so much money is available for AI tests, many organizations jump into pilots without clarity about what they are trying to achieve with AI. As a result, pilots are rarely targeted or well designed to meet the goal, and fall flat after some initial excitement with the “cool” factor.

Organizations should start out their AI journeys with a clear understanding of the “nail” — for most, it will usually come down to one or more of these four goals:

• Reduce costs (e.g. by automating/simplifying processes)

• Increase impact (without massively increasing costs) by solving a bottleneck, e.g. that of scarcity of call center agents, clinicians or radiologists

• Increase revenue (e.g. by adding new services or increasing throughput)

• Improve patient or provider experience (e.g. care orchestration, ambient scribes, etc.)

5. Workflow Integration & Design Failures

Even when organizations have a clear analysis of the problem(s) to solve, and specific definitions of success, they often lack design processes or expertise to integrate the right AI in the right format to address the problem, learn from failures and iterate towards an optimal solution.

While everyone naturally gravitates towards chatbots, these rarely fit into existing health worker workflows seamlessly. If you start from the problems, however, there are actually an infinite number of ways AI can be put together to solve a problem, such as:

• Ambient scribes that transcribe conversations and populate medical records, solving the key challenge of getting frontline providers to reliably input information into EHRs.

• Passive monitoring agents that flag concerning patterns in routine data, helping address critical mistakes or gaps by providers or systems.

• Competency-based coaching systems that assess healthcare worker competencies and provide personalized refreshers, helping address knowledge/skill gaps at a low cost.

Different models or instances can also be stitched together into integrated workflows, e.g. taking inbound messages from clients and triaging/categorizing them using Model 1, then passing to specialized sub-agents to answer different categories of queries (Models 2.x), and supporting humans in drafting responses to the few critical or unclear ones that require human involvement (Model 3).

Organizations therefore need to put their design thinking hats first and foremost, and co-design with users what AI-enabled solutions might alleviate their challenges. Early prototyping could even involve human actors pretending to be AI to gauge the right input-output combinations. Only after the problem and solution are conceptualized manually can one explore the right AI models and prompts to produce the right outputs, and design the evaluation metrics and thresholds that AI has to pass in order to constitute success and deserve scale-up.

6. Resource constraints:

Traditional health systems such as hospitals typically lack the expertise in-house to champion and build AI literacy internally, and struggle to attract such resources to their organizations even if they want to. Implementers like health non-profits and social enterprises, for their part, are often too time and resource constrained (even more so in the current aid crisis) to think creatively about AI. All of this is compounded by the fact that AI resources are scarce and extremely expensive in most markets.

One obvious answer to this is to invest in shared resources that can spread that cost across organizations — organizations like Audere, IDInsight and Turn.io are already trying to do this, as are fellowship and acceleration programs at large model developers. We need many more AI and design experts supporting non-profits and health systems to realize the true value and potential of AI, and handhold them through the journey.

7. Inertia and Change Resistance

Most large organizations are set in their ways, and health systems are no exception. Doctors and health professionals who have spent their entire formative years studying and operating within those doctrines are incredibly resistant to change and disruption, especially from something they don’t understand or trust. Even nimble organizations like Penda Health who have been on the frontiers of scaling up AI have told me that the hardest part was getting the providers to trust and use the AI guidance, not the technology innovation itself.

Even when early adopters and internal champions have successfully run pilots, it is therefore extremely important to think through the change management efforts needed for wider scale up and use. This may involve securing buy-in on success criteria and investment requirements from top leadership ahead of time, or mapping against provider pain points and co-designing with them to ensure the intervention does not languish on the proverbial shelves post-deployment.

C. SYSTEMIC BARRIERS

8. Tooling gaps for model orchestration, evaluation and more

The process for exploring, evaluating, selecting, integrating and maintaining model deployments in real world clinical and health contexts is non-trivial. Currently, in the absence of proper implementer-friendly “Dev-Ops” infrastructure, most implementers are walking this journey solo and unsupported, and hence often recreating rudimentary “wheels” to get the job done.

Here are some basic questions every implementer inevitably faces at some point in their AI adoption journey:

• What are the available models, tools and development strategies for my context?

• Do I build on open or closed models? What are the trade-offs I should consider (accuracy, flexibility, latency, cost, data safety, long-term adaptability, etc.)?

• What approach should I use to adapt and specialize the model for my use case (RAG, fine-tuning, prompting techniques, etc.)?

• How do I evaluate which model is the right one for my specific language, dialect, context, use case, etc.? What is the right size and structure for an internal benchmark dataset?

• How do I orchestrate different models/instances to get the job done, and maintain context between them?

• Are models performing according to my expectations post-deployment, and how consistently? Is model performance degrading over time?

• How happy are users with the responses, and how do I identify and fix errors made by LLMs in production?

• When and how do I upgrade to the newest models, and how much effort would that entail in redesigning the prompts and debugging problems?

For example, Penda's WhatsApp system needs to classify user intent, route to appropriate agents, maintain conversation context, and escalate complex cases to humans. But as Rob explained, "You can't build all of that context and flow into a single prompt. You need tools that define how the bot should behave, categorize user intent, and maintain context between agents. The models aren’t really the biggest gap; the software that helps you put the model pieces together and evaluate them is the biggest gap."

Implementer after implementer I spoke to echoed similar frustrations with the tools that exist today. For example, they need tools to evaluate "Maternal health triage for ASHA workers in Gujrati" or "TB adherence coaching for Nairobi slum dwellers in Sheng"—combinations of language, domain, and use case that no centralized benchmark can anticipate or represent.

From Endless, we are coordinating an effort with some leading benchmarking organizations and ecosystem partners (e.g. Harvard-EPFL LiGHT Laboratory, Audere, CHAI, Qure.ai, IDInsight, Agency Fund, etc.) to fill some of these tooling gaps, but this inevitably requires a system wide effort and significant resources.

9. Coordination failures and “tragedy of the commons”

The above is one example of a broader set of coordination failures across the ecosystem. Despite AI being a ‘blue ocean’ opportunity, current incentive structures, funding crises, and a prevailing “scarcity” mindset are preventing open sharing of successes, failures, and best practices, which slows down progress for the entire ecosystem.

For example, multiple NGOs are independently collecting and labeling hundreds of hours of audio in the same languages, while this could be done a lot more efficiently at an ecosystem level. Organizations that should collaborate are instead competing.

While efforts like PATH’s community of practice around AI for Health is a good step in the right direction, we need many more shared platforms and coordination mechanisms, and even more importantly conducive incentive and funding architectures, for addressing these market failures.

10. Regulatory & Liability Hurdles

Few countries are getting ahead of the AI disruption with enabling regulation. The US is notorious for its litigious legal environment which discourages any meaningful experimentation in patient-impacting AI prior to lengthy FDA approval processes. Even poorer countries sometimes have unfounded and unhelpful requirements such as not allowing health data to leave the country’s borders, despite local data centers incurring much greater security risks than an AWS cloud server. There are of course, exceptions to this, most notably Rwanda within the African continent. But as a general point, there are path dependencies based on decisions made in the 2010s’ digital health era that are hindering experimentation and adoption of AI in many countries.

Even non-technology related policies, such as rules defining what each cadre of health worker can and cannot do must be revised to take proper advantage of AI to task shift responsibilities to lower levels. One possible solution to this is to create “sandbox” environments where governments can relax existing laws and rules to allow AI innovators and implementers to experiment freely, traded against more intense oversight and evaluation requirements. Once these experiments prove successful, policymakers can decide how to update the laws and regulations to allow those innovations to scale.

11. Misaligned and Perverse Incentives

There are of course numerous perverse incentives and interest groups in each ecosystem that will resist AI disruption. Most notable among these is probably doctor’s associations, which tend to be extremely powerful lobbies in many countries. Ministries of Health, which are often run by doctors, are particularly susceptible to such lobbies. During my work with Jeeon in Bangladesh, we repeatedly encountered doctors' associations resistant to any new idea (e.g. delivering care through pharmacies) that was deemed to undermine their authority and status.

As a mentor recently pointed out to me, successful innovations often succeed by “going around” existing interests rather than trying to co-opt them. Mobile money succeeded by working with telecoms, not banks. Medical abortion technology scaled by working with pharmacies and community health workers, not doctors. Similarly, AI in healthcare may find its biggest wins by empowering non-physician providers—pharmacists, paramedics, community health workers—rather than trying to convince doctors to change their workflows. While health ministries might be reluctant, finance ministries and insurance companies have stronger incentives to invest in cost-effective prevention and early treatments, and as a result might be bigger champions of AI if the ROI case can be made to them effectively.

Looking forward

As should be amply clear by now, the bottleneck is no longer "Can AI do X?" The question is: "Can we organize ourselves fast enough to use what AI can already do?"

In other words, it is not about models anymore —it's about mindsets, workflows, financing, and politics. The tools to close that gap are knowable, and likely vastly cheaper than the billions being spent in better and better model development.

What we need now is the collective will to build shared infrastructure, align incentives, and get serious about execution. But first, we must recognize that our old mental model for health systems is headed for a dead-end, and we must reimagine our role in this brave new world.

Disclaimer: The views expressed here are my own and not necessarily those of Endless.

We explore two cutting-edge approaches, both proudly supported by Endless Health—Community Insights to Action (CITA) and Journey Mapping 2.0—and reflect on what it takes to go from feedback to action. This episode is a full-length dive, and I’ll be releasing a shorter cut for Spotify soon.

I hope it sparks ideas, challenges assumptions, and opens up new pathways. Let me know in the comments what resonates.

CONVERSATION SUMMARY

🩺 THE PROBLEM: HEALTH SYSTEMS THAT DON’T LISTEN

1. Designed Without the User in Mind

Most public health systems—especially in LMICs—were built to deliver services to people, not with them.

As Alex put it: “They were never designed for real-time feedback.” The result?

• Top-down planning with little ground-level insight

• Tokenistic civil society participation

• Delays, drop-offs, disengagement

2. Invisibility of Lived Experiences

Rubayat and Emily both highlight how data systems obsess over outputs (coverage, supply chains) while neglecting outcomes like patient trust or satisfaction.

• Pregnant women delay antenatal care—not out of neglect, but due to prior bad experiences

• Caregivers want immunizations—but inaccessible hours, rude service, or inconsistent outreach push them away

3. Informal Systems Fill the Gap

In Bangladesh and beyond, people often choose informal providers over government clinics. Why?

Because responsiveness, convenience, and dignity often matter more than credentials.

🔑 Takeaway: A health system that doesn’t listen drives people elsewhere—even when services are technically available.

🔧 SOLUTIONS IN PROGRESS: TWO APPROACHES

1. Community Insights to Action (CITA) – VillageReach

A framework to routinely gather, analyze, and integrate community feedback—especially from the most underserved.

Key features:

• Embedded in existing monitoring systems

• Co-designed with government and community

• Feedback via hotlines, QR codes, surveys

• Strong focus on closing the loop with communities

2. Journey Mapping 2.0 – PSI

An evolution of traditional patient journey mapping using tech-enabled, longitudinal engagement.

• Remote follow-up with pregnant women via IVR, WhatsApp, AI avatars

• Builds a narrative of real decisions and barriers across care episodes

• Aims to make insight generation continuous, cost-effective, and scalable

🔑 Together, they form two complementary approaches towards an integrated solution:

CITA builds deep system-wide integration; JM2.0 explores lean, digital pathways to feedback at scale.

💡 LESSONS & INSIGHTS

1. Tech Must Follow Trust, Not Replace It

• IVR worked better than WhatsApp—not because it was newer, but because it was simpler and more familiar

• People felt “heard” by automated calls because they were introduced via trusted community workers

2. Start Simple, Then Layer Tech

• CITA began with government process design—not AI dashboards

• Emily: “You can’t tech your way out of a broken feedback loop.”

3. Stated Preferences ≠ Real Behavior

• People say they prefer WhatsApp, but hardly use it

• Voice response feels safer than self-recorded audio

• Cultural comfort and digital literacy shape use far more than survey data suggests

4. Granular Data = Actionable Data

• Macro-level feedback creates fatigue (“We never hear back”)

• Micro-level insights (like evening immunization preference in DRC) prompt real change

🧱 CHALLENGES & BOTTLENECKS

1. Financing & System Incentives

• Health budgets remain vertical and inflexible

• Local facilities often lack autonomy to respond to community needs

• Feedback mechanisms are often housed in the wrong place (e.g., communications departments, not PHC leadership)

2. Multilateral & Global Blind Spots

• Even WHO and the World Bank rarely emphasize responsiveness as a core metric

• Physician-led systems tend to be system-centered, not patient-centered by design

• Emily: “We’re not institutionalizing experience data. Even globally, it’s not measured.”

3. Structural Power Dynamics

• Feedback can be seen as threatening

• Requires champions at high levels to mandate change

• Without accountability for using feedback, it becomes performative

🔭 VISION: WHAT RESPONSIVE SYSTEMS COULD LOOK LIKE

• Feedback as a default system input, not a side project

• Integrated, interoperable systems that connect feedback with quality, cost, access, and outcomes

• Patient stories valued as much as metrics

• Flexibility within structure—systems that can adapt, not just deliver

• Creative care models—from schools to pharmacies to local gatherings—enabled by community input

🎯 Quote from Alex: “A system where feedback is routine, disaggregated, acted on, and visible to the community—it’s not just better policy. It’s trust in action.”

📣 CALL TO ACTION

1. Build Communities of Practice

• Share models, frameworks, lessons across organizations and governments

• Learn from failure, not just success

2. Institutionalize Responsiveness

• Integrate feedback into national systems

• Build low-cost, scalable models for governments to adopt without donor dependence

3. Push Multilaterals to Join the Table

• Co-create guidance, global standards, and advocacy platforms

• Elevate responsiveness and experience as central to UHC and SDG goals

4. Link Feedback to Outcomes

• The ultimate proof is not the dashboard—it’s better trust, access, and health results

In summary, health systems won’t change just because we measure more. They will change when we measure what matters—and act on what we hear. If we don’t, we will continue to invest billions and see little returns.

Thank you, Emily and Alex, for reminding us that responsiveness is not a mere feature of health systems—it’s the foundation of people’s trust in them.

Share

Disclaimer: The views expressed here are my own and not necessarily those of Endless.

A couple of weeks ago, I found myself in a deeply unsettling situation. Like millions of others, I had entrusted my genetic data to 23andMe to learn about my ancestry and health predispositions. Now, with the company filing for bankruptcy and seeking acquisition, I was horrified to discover that I couldn't fully download or delete my information due to servers overloaded by fleeing customers. My DNA — perhaps the most intimate data I possess — suddenly felt out of my control.

This might sound like a first-world problem, at least for now. But lack of data sovereignty or control is not new to me or unique to the developed world. During years of building digital health solutions across a dozen developing countries, I have repeatedly seen how the rush to digitize healthcare often prioritizes institutional convenience over individual agency. In most of the systems I helped build, patients had little or no access to their own records. Consent was rarely sought. Information was siloed in “walled gardens” with poor interoperability and minimal cybersecurity—because budgets were barely enough to build the tools, let alone secure them. As a result, millions in Bangladesh, myself included until recently, have to carry stacks of medical papers from one provider to another—yet continuity of care remains elusive.

Why this matters now

While there are many urgent issues in global health, control over personal health data is becoming pivotal due to three converging and urgent factors:

1. AI is here. As I explored in previous posts (singularities; self-care), AI is set to deeply personalize healthcare. Tech giants like Google, Apple, and OpenAI —and countless startups—are already building AI health coaches that will rely on, and accumulate, unprecedented volumes of personal data. Who owns and controls that data will shape healthcare for decades and expose people to immense risks.

2. The economics of data are shifting. There’s a growing financial incentive to centralize health data, driven by AI’s hunger for it. The more information Google or OpenAI has about your health, the more personalized your experience, and the less likely you are to switch to another AI companion. Patient-centered alternatives, like personal data ownership, offer better outcomes—but lack the same profit motive; if anything, it runs counter to all the financial incentives! Such perverse incentives have historical precedence in the adoption of closed EHR systems like Epic, which prioritized internal efficiency over patient agency and continuity across systems.

3. The Global South is on the cusp. Only 35% of lower-middle income and 15% of low income countries have national EHRs today, but many are newly digitizing. Their decisions will shape health systems for generations, so it is very important to get it right.

First Gen Mistakes we must learn from

In developed nations, health digitization was driven by goals like efficiency and data-driven care. While 96% of U.S. hospitals now use certified Electronic Health Records, this impressive statistic masks critical flaws:

• Centralization creates vulnerability: Epic Systems alone holds records of 78% of U.S. patients - about 325 million individuals. It also controls 3% of all non-US health records (285 million people).

• Patients lack control: Only 54% of U.S. adults have been offered access to their medical records, and of those, only 57% actually accessed them.

• Systems don’t talk to each other: 70% of hospitals report issues with data exchange. Four out of ten still rely on printouts.

These failures affect me personally. My medical history from living in Boston (2010-2013) is inaccessible in my current MyChart, which only starts in 2023 (when I was 39 years old!). Records from Bangladesh? Completely missing! I had to repeat a full set of MRIs at significant cost and discomfort (I get claustrophobic inside MRI machines) after moving to the US—because my fairly recent reports from Bangladesh were not accessible/acceptable.

The Risks of centralization

And the risks of all this centralization goes far beyond just poor continuity of care. Let’s look at the mounting evidence:

⚠️ Massive Breaches

Even sophisticated tech giants like Facebook and Google have reported breaches of their client databases. Here are some breaches reported by various health systems, but given the lack of mandatory reporting requirements in most countries, this is likely a gross underestimate:

• Change Healthcare ransomware attack (2024): exposed data of 190 million patients - more than half the U.S. population

• 23&Me (2023): exposed genetic data of 6.9 million individuals.

• SingHealth (Singapore, 2018): Compromised 1.5 million patients' records, including the Prime Minister's!

• AIIMS (India, 2022): up to 40 million patient records potentially exposed

🕵🏽‍♂️ Government Misuse and Abuse

Even governments cannot be fully trusted to be effective stewards of their citizen’s health data, both due to widespread incompetence and poor governance, and (occasionally) sinister motivations.

• Brazil: 243 million health records leaked online and undetected for six months, due to credentials uploaded mistakenly to Github by an employee. Very few Ministries of Health have basic security measures in place, let alone strong cybersecurity teams to fend off malicious attacks.

• China's COVID health codes were weaponized against protesters

• The UK's NHS partnership with Palantir raised privacy alarms

• In Bangladesh, the deposed Prime Minister’s son was sued for selling the citizen’s national ID database for personal profit!

If all of this does not ring an alarm bell for you, for a moment imagine being a persecuted minority — queer in Uganda, Uyghur in China, or Venezuelan migrant in present-day USA. Entrusting your data to governments that could persecute you is not just irrational — it is an existential threat!

💸 Corporate Exploitation

Similar to 23 & Me’s current situation, there are numerous possibilities for data in corporate hands being misused beyond its original intention and authorization.

• Google's Fitbit acquisition raised concerns about health data being used for advertising

• The GEDMatch genealogy database was acquired by a forensic genetics company for use in law enforcement, raising privacy concerns

• Insurance companies gaining access to genetic data such as 23&Me’s could use it to deny coverage for pre-existing conditions

Overall, the key takeaway is that centralized health data amplifies power, and power can be misused. But what's even more concerning to me are the "lygometric" risks no one is talking about — those hard-to-anticipate, emerging threats that become plausible as technology advances. In the current political climate, it is not inconceivable for conservative US states to identify who got an abortion from their period-tracking app data (just this week IRS agreed to share tax records with ICE to aid in deportations). Using rapidly advancing and democratized tools like CRISPR and AlphaFold, genetic data could be used by even non-state rogue actors to develop targeted bioweapons (FYI — the 2023 attack on 23andMe’s servers specifically targeted the genetic profiles of Ashkenazi Jews!). AI systems analyzing health records could predict and manipulate voting behavior, similar to the Facebook-Cambridge Analytica scandal of 2016.

Health records, unlike credit cards or passwords, can't be changed once exposed. The impact of these breaches can even cascade across generations - affecting not just individuals but their children and grandchildren through their genetic information.

What can we learn from this?

Above, we have established the myriad grave risks of centralizing health data, following the Western model. Without going into too much detail into the history, it is sufficient to conclude that there were missteps made on the path to digitization in developed nations, which has caused path dependencies that are hard to course correct or reverse. For example, one of the core drivers of centralized health records were the strong incentives for systems to have “internal interoperability” (e.g. various parts of a hospital system talking to each other), but there was little corresponding motivation for “external interoperability” (different systems talking to each other) which would actually benefit patients.

Newly digitizing countries in the global South have the benefit of learning from these mistakes. They also have newer technology paradigms to start their journey with; for example, standards like the Fast Health Interoperability Resources (FHIR) did not exist before 2012, a gap that partly contributed to the data fragmentation in the West. In contrast, Jean Philbert, the Chief Digital Advisor of Africa CDC, declared with characteristic gusto during the Open Digital Health Summit conference in Nairobi in December 2024 — “We are going to put Africa on FHIR!” This is undoubtedly an exciting prospect.

We have seen this “leapfrogging” effect before with mobile phone adoption and digital wallets, and this is another huge, and time-critical, opportunity to bypass failed legacy systems altogether.

Core principles, examples, and design of a patient-centered health data architecture

Starting from first principles, we must first acknowledge the basic patient rights when it comes to health data:

• Right to access and understand their data.

• Right to consent before data is collected, shared or used.

• Right to delete or correct data (including right to be “forgotten”).

• Right to be informed about risks, including secondary findings in genetic testing.

• Right to benefit from data use (e.g. in shared research gains, personalized AI, etc.).

• Right to protection from misuse or exploitation.

Any system we advocate for or build must at a minimum fulfill these principles. Thankfully, there are a few early experiments in different parts of the world that we can learn from.

Early experiments and bright spots

Estonia offers perhaps the clearest example of what's possible through decentralization:

• Every citizen owns their health data and can give granular permissions to access it.

• There is an auditable log of every time their records have been accessed and by whom. This promotes accountability to patients.

• Records stay with providers, not in a monolithic database, but can be pulled into an unified view on demand through a secure data exchange network called X-road

• Blockchain technology ensures transparency and prevents unauthorized changes

• 99% of health data is already digitized, and 99% of prescriptions are electronic, while maintaining patient control

India's ambitious Ayushman Bharat Digital Mission (ABDM) is attempting something similar at massive scale, with over 670 million health accounts created. There are however questions about infrastructure readiness in rural areas and whether the system will exacerbate digital divides. While too early to call it an unqualified success, it demonstrates that even large, complex health systems can prioritize patient consent and control.

Beyond India and Estonia, many countries now have Smart Health Cards which enable people to have portable but verifiable digital clinical records, such as vaccinations or test results, developed in response to the COVID-19 pandemic.

Innovative startups are also emerging in this space. HealthTAG in Thailand is using blockchain to enable patients to own and access their medical records across providers. They've developed an NFC card system where patients can access their consolidated records by scanning the card and entering a PIN. Providers store data locally, while blockchain indexes and enables sharing based on patient consent. HealthTAG is also exploring data monetization models where patients can sell anonymized data for research—essentially becoming "CEOs of their health data."

The World Health Organization-endorsed International Patient Summary (IPS) Standard provides an early blueprint of an globally interoperable summary health record that individuals can carry across borders. EPIC and other major EHR systems announced support for the IPS standard in 2024. Google has launched Health Connect, an open protocol for health data sharing across the 3B+ device Android ecosystem, although a cross-platform standard is still to be developed. The Open Wallet Foundation is convening an event in July to discuss how digital wallet protocols could provide more agency and control to individuals over their various forms of data (including health and financial).

It is important to learn from these early experiments and design robust and decentralized frameworks that can be adapted and adopted in different contexts. Abstracting away, we can identify the following basic requirements any system should fulfill to be future-proof:

• Technically decentralized and auditable (data stays with source systems, i.e. no central repository that can be hacked or exploited, any access logged).

• Based on open standards like FHIR, HL7, IPS so that it is completely interoperable across platforms and data stores

• Patient controlled and owned (support for granular, revocable data sharing, record deletion, exporting to other platforms, etc.)

• Inclusive (designed for various levels of digital literacy and access, such as through a voice AI interface in ones’ natural language)

• Politically robust (usable even in fragile or authoritarian states; i.e. prevents misuse and abuse by design)

• AI-ready (supporting personal health companions through protocols like MCP, but with strong governance of secondary data use)

Below is a simplified illustration of how such a model would be superior to current models:

Practical considerations

Implementing this vision faces significant challenges. First and foremost, no one model will serve all contexts — South Sudan and South Africa will have very different capabilities (resources, regulatory, governance, etc.) by definition, and so the overall approach may have to be adapted to each unique context.

However, irrespective of local variations and constraints, it is clear that left purely to market forces, the intrinsic incentives will dominate the choices, which in all likelihood will not lead to patient-centered systems. Even governments, except forward thinking and “benevolent” ones like Estonia, are unlikely to fully embrace a patient-centered paradigm. For example, critics point out that India’s Health Data Management Policy contains loopholes for use of data for surveillance and monetization purposes.

There is therefore a strong case for market shaping and policy approaches to ensure patient needs are prioritized. Multilateral organizations like WHO and Africa CDC, and funders of health systems such as World Bank and the Gates Foundation, have a major role to play in taking a stance on patient data rights, and defining conducive normative standards and reference architectures. Concerted efforts from these agencies and patient advocacy groups could force governments to set strong data protection laws, invest in basic digital health infrastructure, and require health sector players to comply with interoperability and data sovereignty requirements, including and particularly for the new breed of AI-for-health companies.

Another complication of decentralized ownership and management is that people often don’t understand the concept of data privacy, ownership and control in many parts of the world. The most poignant example of this I saw was while building a mHealth intervention for BRAC in the urban slums of Bangladesh. Contrary to our fears that pregnant women in a conservative Muslim country would be skeptical about sharing their data over the internet, they showed up excited and all dressed up because “the doctor is going to see my picture.”

We therefore need a lot of consumer advocacy and education — including the risks of having their data fall on the wrong hands. This is another important role for patient advocacy and support groups, who can support digital health literacy initiatives. I can imagine conversational AI tools greatly assisting in such patient education as well.

The third major complication of data decentralization is regarding access to consolidated data for research. Centralized databases (e.g. EPIC’s Cosmos dataset) are conducive to running research at a large scale. Under a decentralized model, there would have to be incentives built into systems to motivate patients to share non-sensitive information with researchers (e.g. through directly compensating patients for data sharing or participating in trials).

In summary…

We have two starkly different alternative futures ahead:

1. A future where tech giants and governments amass ever-larger health datasets, promising efficiency through centralization - but at the cost of individual autonomy, privacy and grave risks of breach, misuse and abuse.

2. A world where individuals own and control their health information, choosing when and how to share it - enabling personalized care while maintaining privacy and agency.

The Global South stands at a unique moment in history, where the choices made by countries today will determine which of those realities we inherit into the future. We must also protect at all costs against AI companies consolidating and “owning” patients’ health information simply by virtue of the private conversations people have with their chatbots and agents. [A small personal advice here: if you are using an AI chatbot for health conversations, make sure the organization’s business model does not involve directly monetizing your data in any shape or form. Personally, I feel much more comfortable with a monthly subscription fee!]

While the challenges of building patient-centered health data systems are significant, the cost of inaction — or of blindly following any variation of the North's centralized model — is far greater. We will have to iterate our way through the inevitable challenges and mistakes, but we must first lock our eyes on the goal.

I hope this post has given you something to think about. If you agree that patient sovereignty over health data is vital, share this with others who care about the future of healthcare and digital rights. And I'd love to hear your thoughts in the comments: Do you know of any new experiments and experiences with promoting patient data sovereignty? How could having true control over health data change people’s relationship with the healthcare system?

Share

Disclaimer: The views expressed here are my own and not necessarily those of Endless.

Firstly, I recognize the immense privilege I have of saying this from a position of relative safety and stability. Secondly, I completely understand that this may be triggering for some of you who are suffering right now. I will encourage you to stop reading beyond this if so.

But given I trust this community so much, I hope you will take this with the intent this is meant; i.e. that of radical self-reflection and a strive for true and lasting impact. If I thought me (a puny Substacker with ~500 subscribers) writing about this would make the world any worse for our sector, I would certainly not have written it.

***

Like all of you, I've spent the last ~2 months processing the seismic shock that hit global aid and development starting on January 20th. The sudden freeze of USAID, and the swift and decisive dismantling of this 50-year old institution sent shockwaves through the entire sector. My own friends, family members and colleagues lost jobs overnight. One organization I helped co-found in Bangladesh had to lay off or furlough a majority of its staff to survive the sudden drying up of projects, as did numerous other organizations. Most devastating of all, I couldn’t stop thinking of the millions of vulnerable populations who depended on aid-funded health services and were now left in limbo.

In Nairobi's Mathare slum, a TB screening program had to let go of their X-ray technicians when salaries vanished. In rural clinics across Africa, shelves of HIV medications are running low with no resupply in sight, and children are already dying as a result. Women affected by domestic violence are suddenly left with no legal resources, their only lifeline, to turn to. Clinical trials testing critical new treatments were suspended mid-stream, leaving patients without care. The human cost is staggering and will continue to mount in the months and years to come, unless other funders (philanthropists, governments, etc.) step in to bridge the gaping hole in last mile services and products.

As someone who has worked in global health for over 15 years, my initial reaction was profound shock and grief. How could decades of progress be upended so suddenly? My empathy and solidarity for the affected people were clouding my every thought. But as weeks turned to months, I found myself wondering why I’m suddenly bemoaning the “death” of USAID when I have long criticized the entire model of aid-funded development for decades? In order to remain true to my beliefs and values, I must be able to simultaneously hold my empathy for my colleagues in the sector AND own up to my belief that the way aid was done was a huge part of the problem, as evidenced by the way health systems are faltering like a house of cards as a result of one country’s political whims. And that led me to the next (and in my belief, all-important) question: What if this crisis, painful as it is, finally forces us to fix what was broken in global health all along?

The Broken Aid System We Created

To understand where we go from here, we need to look clear-eyed at the system that just collapsed. The post-WWII aid architecture which culminated in the creation of USAID under JFK in 1961, while well-intentioned, created some fundamental problems:

1. In the quest to reduce mortality, it reduced people to statistics. In order to show up on most public health charts, you would have to be dead. But for billions around the world, death can be a welcome end to a lifetime of suffering, as I have heard in various forms from numerous poor patients during my work at mPower and Jeeon. Unfortunately, we rarely if ever understood or measured suffering, and focused all our attention on averting deaths. A corollary to this is that we approached “poor people’s health” in a very different way to how we would approach our own healthcare, where continuity of care, quality of care or the experience delivered mattered much less than superficial and rudimentary measures like “access”.

2. It favored vertical disease programs over integrated primary care. This was needed to secure quick wins initially (for example to avert deaths), but became a liability when funding flows through separate channels for HIV, TB, Malaria, MNCH, etc. crowded out investments in primary care and health systems more broadly. Take the example of Rwanda - between 2003-2008, it received $187 million exclusively for HIV/AIDS programs (affecting ~3% of the population), over 5x its entire annual health budget of $37 million. Many health problems went unaddressed due to such narrow focus — e.g. acute respiratory infections, which, despite being 26% of the communicable disease burden globally (and which the same HIV patient might be suffering from), received only 2.5% of all aid. Finally, aid-funded programs often poached providers from the public sector, weakening health systems further — for example, a Global Fund funded program paid triple the average pay for local medical staff in Ethiopia.

   Aside from skewing priorities, such vertical funding silos also caused tremendous fragmentation. I saw this firsthand in Bangladesh - different agencies (multiple public agencies and non-profits) would send their own cadres of community health workers to the same households for overlapping tasks around a pregnancy. The inefficiency was stunning. Moreover, it caused tremendous confusion and intimidation for a poor patient to navigate this complex maze — as a result, many of them would default to bypassing the health system altogether, and relying on informal providers like drug shops.

3. It promoted donor dependency over self-reliance. If you ask anyone in the aid sector, they will idealistically tell you that their goal is to work themselves out of a job. Yet, in 50 years of aid-funded development, we have made little progress towards self-sufficiency and sustainability. In 2013, while designing a $80M USAID FTF funded agriculture program with a renowned global NGO, I advocated for obvious sources of earned revenue so that the project could sustain beyond the 5 year contract term, only to be summarily told off that “we don’t have the apparatus to manage revenue”! Just last week, I was speaking to an impressive community-based non-profit in Kenya called Afya Research Africa who told me that an aid-funded program providing free malaria medication destroyed their previously sustainable business model around treating malaria, and eventually left the community with no services after the donor-funded program packed up and left 5 years later. In these and many other ways, top-down aid warped national priorities and created vulnerability to exactly the kind of funding shock we're seeing now.

4. It prioritized compliance over outcomes. As I wrote about in a prior piece on incentives, the aid industry's obsession with process metrics and reporting requirements often came at the expense of actual health impact. In the same 2013 proposal for the $80M USAID grant, the DC-based consultant declined to even visit the implementation sites with me — the priority clearly was to check the proposal boxes in the three days she had before going back, despite working with a very superficial understanding of ground realities.

The Silver Lining: Yet Another Chance to Build Back Better

So, is the aid freeze/reduction a crisis or an opportunity? It’s both. In the short term, it’s unquestionably a crisis – lives are at risk, and urgent efforts are needed to fill the gaps. But it’s also a stark illustration of why the status quo was flawed. Many of us in global health have long recognized these issues and advocated for change. The WHO's 2016 Framework for Integrated People-Centered Health Servces called for a fundamental shift in how care is funded and delivered, following up on the promise made in Alma Ata almost 50 years ago. Yet inertia and vested interests made real transformation difficult.

Now, perhaps more than ever, this crisis creates an urgent imperative to do what we should have done all along.

I see three BIG opportunities, which we should have prioritized after COVID, but didn’t:

1. Prioritizing Holistic, People-centered, and Outcome-oriented Primary Care

The evidence is overwhelming - strong primary care systems deliver the best health outcomes per dollar spent. Look at Costa Rica: by investing in community-based care teams (EBAIS) that provide integrated services, they achieved stunning results:

• 90%+ reduction in deaths from communicable diseases, and 13% reduction in all-cause mortality!

• Life expectancy of 80-81 years (rivaling much richer countries)

• All while spending less on health (~7-8% of GDP) than the world average (10%)

Most importantly, their model reduced health inequalities. Between 1980-2000, premature mortality fell 48% among the poorest populations compared to 39% among the richest. By 2009, there was essentially no geographic disparity in infant mortality.

With aid funding shrinking, countries will need to focus resources where they get maximum value for money. The math clearly favors integrated primary care over fragmented disease-oriented programs. Governments and local actors will also have more autonomy to design locally relevant, people-centered systems rather than be perversely incentivized by lucrative aid dollars to prop up fragmented vertical programs. One promising early example of this is Nigeria, where the government stepped up to allocate $1B in funding in response to the USAID freeze, and geared it more towards health system reforms than vertical programs, earmarking significant portions for primary health care, workforce training, and maternal/child health.

There is yet another reason to invest in robust community-based primary care. As we start to face new and grave health challenges like anti-microbial resistance (AMR), chronic diseases (NCDs) and Climate Change, our natural knee-jerk response has been to allocate new verticals of funding to address these needs. Yet, I would argue that a robust PHC approach would treat these new challenges as opportunities to build systemic capacity, such as strengthening workforce capacity to do empanelment and proactive routine follow-ups, motivational interviewing (for lifestyle and behavior change), address heat strokes and vector-borne diseases, and institute/enforce rational antibiotic use. Costa Rica’s model is illustrative here as well — when dengue fever emerged as an issue, the EBAIS teams incorporated dengue control into their routine work. Now, as diabetes and hypertension rise, the teams are managing those by doing home visits for at-risk patients and emphasizing health education. The system adapts because it’s built to be comprehensive and locally relevant.

2. Leveraging Private Sector Infrastructure

One of my biggest and long-standing frustrations has been watching donors support the building of parallel and duplicative health infrastructure while ignoring existing and trusted community-based providers like drug shops completely. In Bangladesh, 70% of primary care happens through local pharmacies and informal providers. Instead of bypassing this vast cadre of providers, why not upgrade their capabilities and integrate them into the health system?

The funding crisis creates urgency to be more pragmatic. Private providers like pharmacies already have sustainable business models. With proper training/accreditation, oversight and incentive alignment, they could deliver many basic health services (there are numerous success stories from across the world spanning family planning, TB, immunizations, malaria, NCDs, etc.) more efficiently than building new facilities or cadres of providers.

There are also numerous social enterprises and non-profits —big and small— around the world that are trying to deliver high-quality, integrated care. Developing health financing models that reward these providers for population health outcomes they generate can be an effective model for building robust, decentralized and outcomes-oriented health systems that rely on local innovation and leverage private capital rather than external aid.

3. Using Technology to Achieve More with Less

As I wrote in my post on AI and healthcare, we're seeing breakthrough capabilities that could help bridge gaps in health workforce, digital literacy and care delivery. AI can now match or exceed the vast majority of human doctors in many diagnostic tasks. Multimodal language models can provide accurate, empathetic health guidance in local languages and even over voice.

These tools won't replace human providers, but they can dramatically extend their reach and capabilities. A community nurse in a rural clinic, supported by rapid diagnostic tools, AI clinical decision support and telemedicine, could handle a much wider range of cases. Community health workers armed with AI assistants could provide higher quality care to more people. Interoperable data standards and “digital public good” platforms can help make care delivery more transparent, accountable, efficient, and continuous. People could have 24/7 access to evidence-based triage and self-care advice from the comfort of their homes.

Above all, we need to stop treating digital technologies as their own silo of interventions, and bake them into the core fabric of future health systems. We need to be able to show that they improve outcomes, reduce cost, and enhance system efficiencies and capabilities.

The Path Forward

To reiterate — the USAID shutdown is incredibly disruptive in the short-term, and completely unnecessary. The human suffering it's causing alone is unconscionable. But it is equally true that the prevailing aid architecture was largely responsible for the brittleness, ‘top-down-ness’ and fragmentation of health systems globally.

Given the rise of right-wing, inward-looking politics across the world, I think it is fair to assume that significant reductions in aid are going to last at least a decade, if not more. I believe it is imperative we leverage this crisis as an opportunity to rebuild a better, more resilient system that is long overdue.

What might that look like? I envision health systems that:

• Center on strong community-based primary care

• Integrate services instead of fragmenting them by disease

• Leverage existing private infrastructure rather than duplicating it

• Use technology to amplify human capabilities

• Measure and reward actual health outcomes

• Build local ownership and sustainability

Most importantly, we need health systems that serve people's comprehensive needs, not donor priorities. The Alma-Ata Declaration was ahead of its time in 1978. In 2025, we have no excuse – we must turn that vision into reality. The freeze of old funding models might just be the thaw that lets something new grow: a health system that truly serves and survives.

What do you think? How are you seeing the impacts of the USAID freeze play out? What opportunities do you see to build back better? Let me know in the poll and comments section.

After months of reflection and recovery, I am finally returning to this blog with renewed determination to envision a better future for healthcare. I hope to do you justice and bring you at least one post a month for the next few months, before getting back to my bi-weekly schedule. I am deeply grateful for your patience, and the many kind souls who reached out through this trying time.)

In my last three posts, I reflected on current gaps in health systems, how AI is both stretching the limits of our current systems and opening up novel opportunities, and the central role of self-care that is currently ignored in our health systems. A logical next step would be to try to envision a 21st-century health system that leverages all we know from our historical mistakes, and the emerging technological opportunities, to deliver a truly people-centered, integrated, and outcomes-oriented care experience for all. I’ll peg the timeline at 2040 to be more specific, but also (hopefully) to leave sufficient time for some of the behavioral and regulatory inertia to be overcome.

The Design challenge

I am going to approach this as a constrained optimization problem starting from first principles (my econ professors may finally be proud!). I will first set up my criteria, assumptions and optimization function, and then try to outline the design that this leads me to. Please note that I am intentionally NOT including as constraints the inertia of existing systems — comprising of regulatory barriers, lobbies, special interests, etc. — which benefit from and try to protect and perpetuate the status quo. I will touch on this again a bit at the end.

Optimization Objective:

Maximize physical and mental health outcomes while ensuring “positive experiences” during care.

Design constraints:

1. Must be applicable with contextual adaptations to most low-and-middle-income countries (excluding those that devolve due to man-made or natural calamities).

2. Must address the healthcare needs of the masses within these countries, and not just within the wealthiest segment of the population that can afford concierge services or fly to another country for care. Please note that this definition may or may not include the most marginalized, such as refugee populations.

3. Must address the three prevailing challenges of fragmented, poor quality, and perverse incentives (post with details here).

4. Must be feasible to accomplish using existing resources or less in most developing nations.

Assumptions:

• Current rate of development in technology, especially digitization and AI, will continue linearly (need not be exponential) at least for next 15 years until 2040.

• Current rates of socio-economic progress in LMICs continue.

• The overall amount of development assistance going to health programs at least remains steady.

Healthcare 2040 - foundations and design

Person-centered healthcare is impossible without a holistic view of the person and their health and wellbeing. In 2040, two fundamental objective sources of truth about each human being will enable this holistic view:

1. Their genome: Extrapolating cost reduction trends (which happen to outpace Moore’s law) in human genome sequencing, a full human genome sequence will cost ~$10 in 2040. As a one-time cost yielding significant lifetime dividends, all but the poorest countries will find it worthwhile to sequence everyone’s genome at birth.

2. Their longitudinal health history: Projecting forward current and emerging trends, each individual can have a biometrics-based, comprehensive health record that stays with the individual from birth to death, and aggregates information about every care encounter, self-care conversation, biomarkers (from wearable devices, biochips, etc.) and hospitalization record throughout their lifetime.

This firm grounding in each individual’s unique and continuously evolving health profile will enable healthcare in 2040 to be fully personalized, and form a life-long care continuum that starts at birth and ends only in death.

Such extensive data aggregation about personal genomics and health, while extremely promising, certainly also poses ethical and logistical challenges, particularly in ensuring data privacy in LMICs. I strongly believe this individual level health profile or record must be biometrically encrypted and legally owned and controlled by each individual (not their government or some private entity like EPIC), to enable maximum aggregation while minimizing risk of breach/exploitation. Here is another article where I dive deep into this:

Overall architecture

By 2040, we must flip 180 degrees the existing top-down and provider-and-system-centered design of healthcare (e.g. emphasis on hospital-based care and vertical disease oriented programming) and amplify the role of individuals, families, and communities as the central engine of health, going even beyond the original vision of Alma Ata. Most of the resources in this new model will be used to strengthen the care continuum at the community level, enabling local health systems to handle as much as 90% of lifetime health needs of the community. Referral linkages to facilities will only handle the remaining 10% of cases that cannot be addressed locally.

Instead of the often blurred boundaries between primary, secondary and tertiary care (which frankly always confused me, and leaves out the all-important role of self-care), I propose a simpler and more comprehensive categorization, in terms of home-based, community-based and facility-based care. Below I will flesh out each in more detail.

1. Home-based care

This would include:

1. Educating and supporting people on all the self-care activities people do (or could do) at home

2. Supporting (and even compensating) families and caregivers to take care of their dependents (children, elderly, disabled, post-operative or terminal patients, etc.) in the best way possible

3. Routine doorstep check-ins by community health workers (CHWs) for things like ANC, PNC, family planning, HIV ART, NCD follow ups, disease surveillance, etc.

4. New technologies like patient-to-doctor teleconsultations, rapid self-testing (such as the COVID home test), cheap bio-chips (e.g. blood glucose monitoring patches), e-pharmacies and medicine home delivery, and most importantly AI personal health companions, will all greatly enhance the range, continuity and fidelity of care that can be provided at home.

As I discussed in my earlier AI post, let me re-articulate the central role of AI in this shift. By 2040, AI health companions with better medical training than the best doctors on the planet today will become virtually free in terms of cost-of-inference (price-performance ratio of LLMs is decreasing by 10x each year). With any internet connected device, a family could have the assistance of such an omnipresent family health companion (closest analog may be the Google Assistant/Alexa of today) that is accessible at any time to answer queries, track vitals and flag deviations, triage a new symptom, manage a chronic disease, and more, individually for each member of the household. And all of this can happen over voice in the natural language of the user.

It’s 3 AM, and your child is coughing hard? You can ask your AI in Bangla what to do. Based on just the sound of the cough, the AI will be able to triage if it is something you need to worry about, and automatically alert your community nurse if so. You’re not sleeping too well lately? It will advise you to take a warm shower, manage your stress, and guide you through a breathing exercise before sleep time. You need an antibiotic for a bacterial fever? It will be able to do effective triage (vs. a viral infection), write a prescription if necessary and order home delivery for you (as long as the supply chain exists). It could even initiate and facilitate telemedicine encounters with the community nurse or the specialist doctor in the city, summarizing the case history for the provider and explaining their recommendations to you afterwards.

The affordability of the AI will enable most governments to distribute this assistant for free to any household with a device. For those without devices, it might still be available over a phone call, or via a kiosk at the local pharmacy (Babyl’s AI tool was available at kiosks at Rwanda’s community clinics and powered the national health helpline before its mother company went bankrupt last year).

Not only that, every conversation with AI companions will incrementally populate the health history for each household member, and create the most granular, rich and continuous health record we have ever seen. As a result, unlike today, home-based care will become a completely visible, transparent process, making rich personal health and behavioral data available to inform community- and facility-based care interactions if and when escalation is needed.

2. Community-based care

Not everyone can manage self-care effectively. Some are too young, elderly, burdened with chronic conditions, or may not have anyone to care for them. Even otherwise healthy people sometimes fall sick from infections, trauma/accidents and other extraneous causes.

People will therefore need health providers and facilities in the community to handle such cases, without needing to travel long distances. Many community-based providers already support patients in their health journeys, such as pharmacies and drug shops, public sector/NGO community nurses and community health workers, and more. However, many of them are currently undertrained, poorly equipped, not digitized, and poorly paid. This must change by 2040, and we must empower local providers with the knowledge and tools to take care of their communities optimally.

With strategic investments, by 2040, in most parts of the world we can have local mini-clinics such as a publicly funded/subsidized Community Health Center (CHC) — staffed by a qualified nurse or paramedic from the same community who can provide the vast majority of outpatient care with support of AI clinical decision support tools, basic lab tests and ultrasounds, and high-bandwidth video tele-consultations with remote doctors and specialists. These technologies will greatly amplify the range and accuracy of medical services these non-physician providers are able to provide, and hence eliminate the need to have physicians be physically present at the community altogether. Indeed, container-clinic models like AccessAfya or North Star Alliance or pharmacy++ models like Ilara Health’s AfyaNzima franchises or mPharma’s Mutti pharmacies already provide a glimpse of what these CHCs might look like.

The vastly superior experience of seeking care at such a CHC (assuming we can fix the “incentive problem”, which I will tackle in future posts), as opposed to traveling to distant facilities like hospitals for outpatient care, are manifold:

• Convenient and more accessible at any time of day

• Savings in travel time and cost

• More respectful, responsive, empathetic, and “equal” encounters from someone you already know from the community.

• Greater accountability of a local provider as opposed to an impersonal doctor who might never meet this patient again

• Option to request house calls for the sickest patients

Not only will each individual interaction with such a health provider be more rewarding, but better interoperability and linkages between providers will ensure that the total is greater than the sum of the parts. At present, you might go to your local pharmacy for routine blood pressure checks for hypertension, but the public clinic nearby has no clue that you have suffered from uncontrolled hypertension for a decade (that is, until you show up at the hospital with a stroke). By 2040, I anticipate that we will have resolved data interoperability challenges (including resistance from private sector, lack of technical capacity in-country, etc.) and established national standards be able to facilitate seamless flow of data and information between all different types and sectors of providers within an ecosystem. An early test of this that we are supporting from Endless is D-tree’s Afyatek model in Tanzania, which seamlessly connects pharmacies, community health providers, and facilities through a unified data structure and platform. Another example is the FHIR-based standardization of health records within both the government and BRAC community health systems in Bangladesh, which for the first time is allowing an unified view of over 120 million patients’ health journeys and helping avoid duplication of efforts.

Of course, community awareness events and outreach campaigns such as immunization drives, diabetes camps, etc. will also fall technically under community based care, and may be managed and executed by this same “care team” comprised of the CHC nurse/paramedic and loosely affiliated CHWs and pharmacies.

3. Facility-based care

Only for the remaining few complicated or emergency cases would people need to travel long distances to the hospital, clinic or diagnostic lab with its specialized doctors, radiology equipment, surgical facilities, and inpatient setups. Currently, hospitals are often overwhelmed with outpatient cases causing provider burnout; in this future model they will not serve outpatient cases without referrals.

Bringing it together

This is “task shifting” at its best, where the goal is to do as much of it possible closest to the home (even at home!) and with the lowest skilled personnel available (even the patient themselves!). Below is an illustration of what this ecosystem might look like:

Note that any specialized care pathways (HIV, TB, etc.), rather than having their own vertically integrated and siloed infrastructure, gets embedded into this infrastructure using intelligent protocols. For example, a patient might get flagged as high risk for HIV by their AI companion, walk to the pharmacy for a rapid test, and referred to the CHC for confirmatory tests and ART enrollment if they test positive. Once they come back home, they are reminded to take their ART pills regularly, and if they do not refill their prescription on time as expected, an flag is created on the CHW’s platform who makes sure to check on the patient during their next household visit.

Let’s look at another example of a condition current systems struggle to manage — NCDs, in particular hypertension and diabetes. In this new model, anyone over 35 with any risk factors will be routinely screened by the CHW for high BP or blood glucose. If flagged, they will be referred to the CHC for a confirmatory test, and an AI or virtual doctor will help determine the personalized care plan for the patient, setting scheduled reminders to both the patient and their various providers to ensure the right steps are being taken on time. And at home, the AI care companion will always be ready on a moment’s notice to answer any questions, guide people through diet and lifestyle modifications, and more.

Because it is a closed loop system, you can design the system to be proactive and intelligent in this way, and orchestrate providers’ workflows and schedules around the patient’s needs, preferences and behaviors. I am calling this “ambient systemic intelligence” (which complements and builds on the “personal intelligence” in everyone’s pockets). Of course we have implemented rule-based scheduling for many years, but they tend to be rigid and one-size-fits all. AI can make these much smarter and more clinically informed, enabling completely novel capabilities such as personalized care planning and orchestration.

And indeed, the vast majority of experts I have spoken to seems to agree that such a system would generate much greater bang for the buck than the fragmented, convoluted mess we have in place today.

Is this realistic?

Before I conclude, let me turn to the all important question on everyone’s mind who is reading this — is this vision really feasible given all we know about how global health works today?

I would posit that all of this is plausible, within the constraints laid out at the very beginning. However, there is a reason I left out the inertia and rigidities of “how we do things around here” out of my list of constraints at the very beginning of this design exercise. What stands on the way of implementing this vision is indeed our own conventional wisdom, biases and prejudices, and systemic inertia. Conversely, by unlearning past mistakes and adopting new approaches (e.g. reorient global health funding flows from vertical disease programs towards systems strengthening, acknowledge that insurance is not the best way to fund primary care, etc.), changing our culture of care (e.g. passive to proactive and transactional to relational), and letting go of our unconscious biases and prejudices (such as that towards private drug shops and pharmacies, assuming we as experts know better than patients what they need), this vision is indeed very much possible to bring to reality.

Even if we are unable to make all these profound paradigm shifts globally, I strongly believe that some countries, especially those in the Global South with innovative and prescient leadership and a blissful lack of legacy infrastructure and systemic rigidities, will invest and successfully build out similar compelling models of care. I was tremendously inspired by Costa Rica’s localized and highly proactive care teams, proving to the world for the first time with evidence that holistic primary care can indeed produce unprecedented results (in this case 13% reductions in all-cause mortality). Sadly political shifts have since dampened much of the success there, but I am certain we will see others step up (my best bet would be on countries like Indonesia, Malawi and Rwanda). However, sustaining political will in these pioneering countries, and measuring results to generate evidence systematically to inform the global conversation, will be key.

In the meantime, the clock has started ticking, and we don’t have time to waste. As innovators, funders, and practitioners, we collectively hold the power to examine our own blind spots, disrupt this dysfunction, and chart a new course for global health.

It all begins with a simple declaration: “We can do better!”

***

I intentionally left out some important questions in this post for brevity’s sake — such as about financing and incentives, or the public-private split of this model — which I will try to tackle in future posts. But in the comments, please let me know if I have missed any other obvious aspects of public health and care provision that is unaddressed by this model. Your feedback and insights will be invaluable as we refine this vision together.

***

What is self-care and how important is it?

If you are anything like me, you probably spend a LOT more time in health-related self-care than you spend under professional/clinical healthcare. I don’t mean self-care in the capitalistic way the Western world has reinvented it, conjuring images of expensive detox spas and hot yoga retreats. I mean the banal, everyday things we each do to take care of our own health, starting from brushing our teeth, exercising, having nutritious food, meditating/praying, or maintaining good hygiene and sleep schedules, all the way to managing our pain with an OTC medicine, wearing a condom to prevent STDs, self-testing at home for pregnancy or COVID, or monitoring our diabetes and blood pressure to manage chronic diseases. Here’s a rather comprehensive facts and definitions page from WHO.

A rough back of the envelop calculation yields ~15 hours each week I spend in such activities. Compared with the ~5 doctor and lab visits I have to do each year, in terms of sheer time spent the ratio is at least 150:1! I can also confidently attest that my physical and mental health status is at least 90%+ a result of my daily choices and actions, and <10% attributable to competent clinical care (doctor visits, prescription medicines, therapy, etc.) for the few things I cannot do on my own.

Of course this is also dependent on genetics, age and other risk factors. As a relatively healthy 40-year old living a stable life, my need for clinical care is much lower than anyone with an assortment of inherited illnesses, a displaced refugee, or my healthy 73-year old dad. If you hold these exogenous factors constant (since we cannot influence them easily), I would further posit that the earlier you start on disciplined self-care, the later and lesser in aggregate you will need clinical care. Or more formally:

The most eye-popping example of this I found relates to the emerging evidence around Type-II diabetes reversal. In a randomized trial in Qatar, the treatment group receiving an intensive lifestyle adjustment intervention showed complete remission of diabetes in 61% of cases, and normal blood glucose levels in 33%. Compared to the control group which just received evidence-based clinical care, the lifestyle modification group was 12X (OR=12, 95% CI 5.17-28.03) more likely to be free from any symptoms of diabetes.

This means that not only is self-care powerful in preventing disease, it can even reverse chronic diseases that have long been assumed by the clinical establishment to be lifelong and irreversible. Contrast this with the estimate that diabetics on average lose 13+ years of life from the point of diagnosis, and imagine the potential gains in terms of reducing human suffering and cost burden on the health system from just one lifestyle intervention!

A conceptual model

If we follow this train of thinking to its logical endpoint, the ideal design of a health system should be one in which self-care and clinical interventions are part of an integrated care continuum that starts at birth and ends only in death:

In this conceptual model, each time you leave a doctor’s chamber or CHW visit, it informs and hones your self-care routines and practices, which keeps you in good health until the next scheduled encounter or acute incident prompts a touchpoint with the formal health system. Simplistically, your goal is to maximize the self-care loop and minimize the clinical care loop as much as possible.

Indeed, this is kind of what it looks like from the typical patient’s perspective already.

Why are health systems designed so differently then?

It is, at least on the surface, rather mind-boggling how rarely health systems across the world are designed factoring in this essential and primary role of self-care. It does, however, make more sense when you visualize the care continuum as it currently seems from the perspective of a clinician or the health system:

In Bangladesh, I would often hear public sector doctors refer to poor people’s “self-medication” at the pharmacy or their belief in traditional medicines in a derogatory sense, implicitly implying that they are acting against their best interests when it comes to health. I myself have been admonished with “Don’t trust Dr. Google”, and “Don’t come to me if you believe in snake oil” when I went to renowned allopathic doctors after doing my research and trying out alternative therapies.

There are at least a few underlying factors that drive this lack of understanding and appreciation of self-care within our health systems:

1. Clinical bias: To a hammer, everything looks like a nail. Trained doctors, who tend to drive public health policy as well, are conditioned to think in terms of diagnosing and treating problems, and certainly not from the patients’ perspective, who would rather avoid going to the doctor at all in the first place.

2. Lack of visibility: We don’t understand what we cannot see. The moment a patient exits the chamber, they become invisible to the “system”, and what they do at home and in their communities are a complete black box.

3. Lack of actionability and personalization: Even if you knew what people are doing, there are few tools, and even fewer resources, to targetedly intervene in each unique case. So why bother? Might as well focus on blunt tools like behavior change campaigns and hope some of it sticks!

4. Financing for “sick-care” over true healthcare: Our legacy systems monetize sickness and finance “interventions”, and are not designed to reward keeping people in good health. Hence, preventive mechanisms like self-care are not financed or prioritized.

5. Prejudices and lack of real empathy for the poor: Last but not least, I have often felt that there is also an implicit bias and prejudice that prevents many clinicians and health systems folks (who tend to represent the upper echelons of society) from truly empathizing with the poor and thinking of their health in the same way we think of our own health needs and aspirations. I fall victim to this bias myself and often have to remind myself of my own needs and experience as a patient and care-seeker.

Given how little we understand and appreciate how people approach and take care of their own health, it should come as no surprise that we struggle to deliver outcomes on anything that requires more than a one-off intervention, and a semblance of a partnership with the individual, their families and their community — antibiotic resistance, tuberculosis, malaria, dengue fever, NCDs like diabetes and hypertension, and mental health are all glaring examples.

If delivering health outcomes is our core objective in designing our health systems, [assuming we can’t greatly influence people’s socio-econo-environmental constraints (a.k.a. social determinants of care)] then we must act as partners in helping them manage their own health better (the 99%) before bringing in the clinical “hammer” for the remaining 1%.

The promise of digitization and AI

Perhaps needless to re-emphasize, but people all over the world, including the poorest, do their best to take care of their own health and well-being given the much larger set of constraints they operate within, which are largely determined by their socio-economic and physical environments, and often also includes poor information.

Digitization has already helped alleviate the information and visibility bottlenecks significantly, even though we have barely taken notice on the public health side of things. In fact, 7% of all google searches (1 billion queries per day) are health related. In India, Prime Minister Modi on an election year was only 2.5x more popular than “diabetes” as a search topic, was at par with “exercise”, and was less than half as popular as “skin” related queries.

Google clearly already has the largest database of self-care queries, and by proxy, interest, anywhere in the world. Combine that with Google Fit (or similar platforms), which is an activity tracker pre-loaded on most of the 3 Billion Android phones in the world, and we also have the world’s largest database of self-care behaviors and activities (daily steps, sleep times, BMI, etc.), even among less well-off populations.

Although there are a plethora of apps for this in the private sector, from a health systems perspective, we have not yet leveraged this newfound visibility to design better self-care oriented interventions or inclusive care continuums. But the size and scope of this opportunity will be amplified another 100x with the advent of clinically competent AI personal health companions.

For example, many of those 1 billion search queries a day are going to become actual clinical consultations, informed by each person’s health history and records, and often done over voice in the user’s natural language. If we design these well, we will gain highly nuanced insights from these conversations about the actions people are taking, and the impact that is having, on their health and well-being. But more importantly, each of those conversations will also be an opportunity for us to inform them with the latest clinically validated information adjusted to their prevailing level of health literacy and understanding (ask ChatGPT to explain Quantum Mechanics for a 6, 12 and 18 year old, to see what I mean), and to influence and nudge their behavior in a highly personalized and contextually appropriate manner. In this new reality, the role of the health system will be to plug into these conversations for insights into people’s health, and to provide targeted resources and interventions to remove bottlenecks. You’re too poor to afford quality food? Let’s connect you to the government rations program. You are struggling to manage your diabetes through nutrition and exercise? Let’s assign a CHW to visit you routinely and supply you your Metformin.

It is important to highlight how significant the personalization aspect of this is. Traditional behavior change campaigns often fail to have significant impact because each individual’s context, preferences and challenges are completely unique. AI can solve the personalization problem once and for all. For example, if I am a poor pregnant woman in rural India and show signs of iron deficiency, it might advise me to grow some spinach on my backyard and increase my lentil consumption:

Just this week, OpenAI and Arianna Huffington’s Thrive Global embarked on a high-profile joint venture to build exactly such an AI “Personal Health Coach” that will remember everything about an individual and provide hyper-personalized and contextually aware nudges to improve one’s preventive behaviors such as exercise, nutrition and sleep. This is only the beginning.

I discussed this impending amplification of the agency and autonomy of individuals, and the role of the health system in this new reality, in more detail under Singularity #3 of my AI post.

A call for re-imagination

Of course building trust in AI health coaches and figuring out their safety guardrails will take some time, and the equity issues are important to think through. We also already spoke about the other systemic barriers we have to overcome, which are non-trivial. Below is a list of the main barriers and potential enablers to help overcome them.

But the first and inevitable step is one of intent and imagination. If we truly believe in building people-centered health systems and want to keep people in good health rather than only treat sickness, we have to prioritize and position self-care as a core, if not the main, engine of healthcare. As a by-product, we might be surprised to stumble upon the most cost-effective way to achieve health outcomes, and a healthy, happy population.

Reorienting ourselves from a top-down, clinically-oriented and system-centered approach towards one that maximizes people’s agency and autonomy to take care of their own health, is LONG overdue. AI might just provide us the best opportunity we have ever had to make that vision a reality, but it is in the end up to us to design and build that future.

Over the past 18 months, I have as a result been obsessed with AI developments, especially the increasingly capable and sophisticated GenAI models targeted at medical use cases. I spent late nights and weekends reading newsletters and arxiv papers to keep up with the rapidly advancing technology and the flood of emerging research. Here are a few curated highlights:

• AI Passing medical licensing exams: ChatGPT and other AIs passed the US Medical Licensing Exam with flying colors, and soon after, they excelled in exams in Japan, Germany, and Israel. The rate of improvement was also notable - while GPT-3.5 was scoring below passing marks, GPT-4, released only a few months later, often scored in the 90th+ percentile of medical students.

• AI better than human doctors in diagnostic conversations and even empathy: Paper (later published in Nature [paywalled]) showing that a conversational AI outperformed human doctors in a blinded trial on 26 of 28 key dimensions of care (which, mindblowingly, includes empathy)

• Generalist foundation models with some creative prompting techniques outperforming specialized medically “fine-tuned” models (“Medprompt” Paper by Microsoft), possibly indicating [my opinion here] that cross-disciplinary expertise somehow provides an added benefit to medical reasoning.

• GPT-4o gaining the ability to comprehend top 50 languages and speak in natural sounding voice across those languages. Also, Swahili model fine-tuned in Kenya, and African startup fine-tuning voice recognition for 200 African accents of English.

• Nvidia and Hippocratic AI joined forces to train AI Nurses that can converse on video and only cost $9 an hour.

• Multimodal models able to analyze medical images: GPT-4 (not the latest Omni model) performed at a post-grad year-2 student level on radiology exam questions, so fine-tuned multimodal models could realistically achieve at-par performance with human radiologists in the near future.

• AI can even outperform human experts on diagnosing the long-tail of uncommon diseases, even without expressly human-labeled datasets (humans under-index rare disease probabilities, whereas AIs are more objective. AIs also “pick up subtle patterns humans miss”)

• GPT-4 approaches expert level performance on ophthalmology exam questions, and significantly outperforms trainees and junior doctors (paper)

• Early indications that hallucinations are greatly reduced in larger context windows, with RAG (i.e. searching a knowledge base) and other techniques. This has been the single biggest concern about using GenAI in medical use cases.

• AI has superhuman persuasive abilities, which may be relevant in nudging people towards healthy behavior change to manage chronic diseases.

As I kept up with these developments, I thought a lot about the upcoming disruption to healthcare as a result of generative AI. Most conversations I heard were about how AI would fit within existing health systems at the margins (e.g. transcribing medical conversations, improving clinical decision support, etc.). But I felt that the vast majority of health experts were missing the big picture and not analyzing the trends and their potential impact from first principles. That’s what I will attempt to do today.

After I read a recent article by Ethan Mollick (one of the most influential applied AI researchers now, founder of the Wharton GenAI Lab and author of the One Useful Thing newsletter which I highly recommend), I realized that my thoughts on this could be framed best by the concept of a “Singularity” — i.e. points at which changes become so profound and radical that is is hard to predict the future beyond them. Although the timeline is still a guess right now, I’m quite convinced these singularities will relatively soon transform how we think of, design and experience healthcare.

At least the first singularity I speak about is being widely discussed already, and I will just summarize it, but I think the ones that follow are logical consequences of the first. And #3 on the list, which I have not heard anyone talk about, might be the most consequential of all.

Singularity #1: Explosion of clinical therapies and novel drugs

We went from pre-schooler level AI (GPT-2) to advanced high-schooler level AI (GPT-4) within four years. With current exponential growth and significant investments in compute, the next leap to PhD-level capabilities or beyond should occur in roughly two years.

Combine this with the revolution in protein mapping (thanks to AlphaFold compressing a billion years of research into months), CRISPR-based gene therapies, and in-silico drug development —with each of these intersecting and compounding on each other—and we’re heading for an unprecedented acceleration in clinical research and drug development. We’re already seeing the early indications of precision therapies like the ‘miracle drug’ that helps obese patients lose weight, or the drug currently in Phase 1 trials that enables adults to regrow their teeth. In the not-too-distant future, expect therapies that cure cancer, prevent/delay onset of inherited/chronic illnesses like diabetes and Alzheimer’s, drugs that slow aging and provide metabolic benefits similar to exercise, and vaccines that prevent deaths from currently fatal diseases like Ebola.

(Please note that I’m intentionally not focusing on intra-country and global equity issues in this post for brevity’s sake. Of course there will be friction with adoption and regulations. Prices of these latest therapies might be more or less expensive than current drugs, depending largely on patent regulations. Existing inequities w.r.t. R&D funding and baseline income may be amplified. Etc. etc.)

Not only will research in labs accelerate, but research will enter the consultation chamber and the lines between practice and research will start to blur. As more and more clinical encounters get digitized and transcribed (again, thanks to LLMs and other natural language processing AIs, soon we will have ambient listening as a way to capture pertinent data to update patients’ medical history from the raw conversations — instead of nurses/doctors having to type into EMRs, which has historically been the greatest barrier to adoption and regular usage), we will have real-time and accurate data from millions of patient encounters that will help validate the outcomes from therapies in real-world settings, and update the algorithms and protocols that determine the optimal therapy needed for a particular patient given their unique health history, ethnicity, age, sex and other information. Digitized clinical encounters will also enable running clinical experiments automatically in a similar way as A/B tests are done in software development to optimize outcomes, although ethical standards and IRB protocols will be crucial to think through for such novel types of research. And when we finally allow AI to drive the protocols guiding the care encounter, any updated protocols could be instantaneously deployed all over the world for better clinical encounters and outcomes (the current time it takes to go from research to practice for a new protocol is ~17 years!).

While it’s currently hard to imagine such an exciting future in low-resource settings like Sub-saharan Africa and South Asia, I think, counter-intuitively, some of these countries might actually leapfrog ahead of many western nations and their fragmented and onerously regulated health systems. Already, initiatives like GIDH by the WHO are showcasing success stories from countries like Bangladesh and India that have made tremendous progress in digitizing their health systems, and providing a roadmap for other low-resource countries to adopt interoperable data standards and digital platforms for every type of care encounter.

Singularity #2: Whether and how we train medical professionals and allocate them in health systems

This rapidly accelerating research will lead us, or at least contribute to, the next singularity. The last studies I know of estimated the rate of doubling of medical knowledge at just 73 days in 2020 (and it has surely shortened further since then). Even very conservatively, what this means is that a medical student entering a 7 year program leading to a medical practice license is experiencing at least 5-6 doublings (a whopping ~50x) of medical knowledge by the time they graduate, which is impossible to reflect in updated curricula. Beyond licensure, no amount of continuing education can help a practicing physician keep up with the flood of new information and protocols, partly contributing to the egregious delay in translating research evidence to practice.

Another dimension of this problem is the hyper-specialization it causes these days in medicine, leading to increasing fragmentation, conflicting advice and physician egos, more drug interactions, and even mistakes in diagnosing conditions accurately. A friend who recently went through a terrifying episode of post-natal complications and surgeries at Johns Hopkins hospital had her gynecologists fail to identify a drug induced fever, which was predicted far ahead by her internal medicine specialist husband. As the subbranches of medicine proliferate and the volume of material in each expand exponentially, this problem will likely get from bad to worse.

Last but not least, the quality of doctors is highly variable, and Johns Hopkins doctors are likely in the top 1% of all doctors worldwide, while the average doctor in a place like Bangladesh likely falls in the bottom two quartiles.

Now contrast this with this the fact that AI is rapidly reaching the level of top doctors, and will soon surpass them, in history taking, clinical conversations and diagnostic capabilities. With increasing context windows, it will be able to take a holistic view of all of a patients’ medical history, and follow the latest protocols and medical evidence accurately. It will even be able to seamlessly traverse multiple specialties relevant to the case at hand, and “consult” with other AI specialists without any ego conflicts or time lags.

Such a super-human clinical agent will be available 24/7, just as much in Botswana (which had <1 oncologist per million population in 2010) as in Boston, with as little as an internet connected device, and at a fraction of the cost of training and paying a human doctor (estimated at $21-48K in sub-Saharan countries). And “they” will never flee the country for better opportunities, and happily serve in their designated rural outposts.

People are rightly worried about AI clinicians replacing human doctors, citing bias and hallucinations as the top two concerns. We must absolutely have the right evaluation frameworks and localization mechanisms to ensure AI performs well in each context. But we must also remember that we are competing against human doctors in real world settings — where poor training and skills, out-of-date and/or forgotten knowledge, short attention spans, implicit bias and stigma, and perverse incentives cause even the best doctors to make mistakes and suggest inaccurate treatments. That is not a very high bar to clear.

I think the biggest challenge will be that even these superhuman AI agents will have to make high-stakes decisions with imperfect information (e.g. a patient who cannot afford a CT-Scan or even travel for an X-ray), and treat with the limited therapies that are often available in resource constrained settings. AI guardrails and benchmarks will have to be nuanced enough to navigate such ambiguous moral and ethical dilemmas.

Even after we figure that out, regulatory and liability considerations might still prevent AI from taking the place of human doctors in the Western world for the foreseeable future, but in countries like Botswana, the economics of it will soon make it unjustifiable to train any more human physicians, at least in its current form. Medical schools might pivot to train many more human “caregivers” (e.g. nurses, paramedics, CHWs) instead, who are trained on the soft skills of taking care of patients while being clinically guided by an AI.

Singularity #3: The transition from “public” to personalized health systems

Even if health systems don’t respond rationally to these developments due to lobbies and regulatory capture, such as doctor’s unions (which I have personally faced in Bangladesh with Jeeon), people will likely make personal choices that could seriously undermine health system function and performance unless they can be fundamentally reimagined.

Let’s break it down a bit. Public health as a discipline serves broadly four key functions:

1. Managing people’s health needs for common health problems/conditions they face, including infectious and non-communicable diseases, pregnancies, etc. in the absence of clinicians to handle every case individually

2. Educational, preventive and promotive communications and interventions to prevent risky behaviors and promote healthy ones

3. Macro scale activities such as immunizations, surveillance of diseases, predicting epidemics, tackling antibiotic resistance, etc.

4. Addressing non-clinical (social, behavioral, environmental, emergencies, etc.) drivers of health

When AI clinical agents reach human parity, one of the key binding constraints on the first function —clinical expertise— will suddenly be lifted. With tailored medical advice and treatment plans based on the latest clinical knowledge, on-demand 24/7, each person will have access to their own personal AI doctor, massively disrupting Function #1. Around the same time, AI will also be able to converse empathetically in local languages and dialects, which means people will likely prefer the convenience of consulting such an AI at home over making inconvenient, time-consuming, and expensive trips to health facilities. Not only that, the AI will be able to persuade and nudge people towards better health in a personalized way as well, informed by the latest in behavioral sciences (significantly improving on Function #2 and eliminating the need for one-size-fits-all communications campaigns).

(If you’re still not convinced, GPT-4o is already freely available to anyone on the planet. Try downloading the ChatGPT official app on your phone and have a voice conversation with it regarding your most recent health problem, starting by asking it to “Act like a qualified health professional”. Then project the rate of progress forward a few years. You’ll know what I mean.)

Resisting this inevitable shift in user behavior will be counterproductive for health systems, because we will lose out on opportunities to localize and de-bias these systems, ensure proper guardrails, target high-risk patients with health system interventions like a CHW visit, and detect trends and patterns from the day-to-day conversations people have with their AIs. Indeed, the micro data from millions of conversations with AIs can also easily be aggregated up to develop the most high-resolution and real-time disease surveillance system we have ever seen (Function #3). Most importantly, perhaps, such deep integration and fundamental redesign can massively improve access and care experiences for patients, and save vast amounts of health resources while improving outcomes. Conversely, if health systems fail to adapt, they will likely be completely bypassed by people voting with their feet (or rather, fingers).

***

The above is my own (hopefully logical) extrapolation from current trends. However, it is still conjecture, and I have not addressed many pertinent questions. For example:

• What will be the new binding constraints after clinical knowledge is democratized? Will it be access to cutting-edge diagnostics, therapies and drugs due to poor infrastructure (cold chains, etc.) and affordability? Will it be the inability of regulatory mechanisms to keep up with the pace of change? What else?

• When we don’t have the blessing of ignorance and can pinpoint exactly which hundred million people around the world need an urgent heart surgery or cutting-edge gene therapy (thanks to their AI doctor), but only have the resources and infrastructure to deliver it to a tiny segment of well-off patients, how will we deal with the resulting moral dilemmas?

• As AI takes on more diagnostic and treatment roles, how do we address the ethical implications of AI making life-and-death decisions? Who will we hold legally accountable for AI-driven medical errors? How will we choose between conservatism (only allowing AI to practice if it exceeds human accuracy by a large margin, as in the case of self-driving cars), or pragmatism (allowing AI to take the place of vacant doctor positions if it clears the relatively low bar for a medical license)?

I therefore think this topic deserves a wider conversation within this community.

What do you think? Please take 5 minutes to leave a comment or question on this post or join this Chat room. Feel free to also reply to this email newsletter directly to share your reflections with me.

• Burnout and New Beginnings: Post-COVID, an opportunity to build a global health portfolio for Endless came my way. I embarked on it with extensive research and triangulation to zoom out and broaden my horizons. My first-principles analysis eventually landed me on some key insights about persisting gaps in Global health, that I am setting as my North Star orientation.

• Key persisting gaps in Global Health from a people-centered lens:

  • People-Centered Health Systems: We need people-centered, responsive health care that meets people where they are, i.e. through robust and holistic care at people’s homes and communities. Need a true paradigm shift, not just HCD as a checkbox item. 4 underlying gaps I’m focusing on going forward:

    • Integration of Care: Horizontal fragmentation of care across sectors and providers causes disjointed care experiences and poor health outcomes.

    • Quality of Care: Despite widespread access today (at least when you count all types of providers), poor quality care persists due to knowledge gaps, fragmented, uncoordinated services and misaligned incentives among providers.

    • Provider alignment with patient experiences and outcomes: Misaligned and perverse incentives and lack of feedback mechanisms from a patient-centered perspective contribute to poor alignment with long-term, comprehensive and outcome-focused care that people need and want.

    • Self-care: In the context of public health, self-care is a fairly new and emerging topic, although people spend a vast majority of their health-focused time on self-care compared to provider interactions. This is a huge opportunity for innovation, especially in the context of AI.

I do hope you find the time to read, reflect and comment on the full post.

By mid 2022, I was an emotional wreck. Although the economy and society were starting to recover from COVID, I was utterly burnt out from two grueling years of pandemic response work. Jeeon’s financial runway was dwindling down to its final months, and I had to face the reality that there was no path back to a commercial trajectory. On top of this, there was an immense void in my personal life as well, since my family had relocated to Baltimore in the middle of the pandemic for my wife’s graduate studies, and staying away from my two kids for months at a time (and alternately, trying to manage the company remotely across a 12-hour time difference) was unbelievably painful.

To support my wife through her final months of studies and to give myself a break, I took a four-month sabbatical in the US for reflection and healing. (I wrote a series of posts in a blog titled “The Recovering Entrepreneur” at this time — if you ever find yourself in an existential/identity crisis, you may find a kindred spirit there).

During this period, I first considered moving to the US full-time. This meant dissociating with my primary identity as a social entrepreneur in Bangladesh, which was daunting. Fortunately, an opportunity to help Endless build a global health portfolio unexpectedly came my way.

We will write about Endless’s vision and mission in an official blog post on our website (and link to it when it’s up), but here, I want to document the incredible learning journey I have gone through in the past ~20 months since taking that leap of faith, and how I have come to orient my compass with respect to my convictions and activism around global health today.

Broadening my horizons

My deep immersion in rural Bangladesh through Jeeon taught me a lot about the gaps in public health programs driving people to private providers, the relationships between informal providers and their communities, the challenges of introducing technology, and the power of incentives in shaping behaviors. If you haven’t read the earlier post, it goes into a lot more gory detail here:

However, after nearly 10 years immersed in one ecosystem, I became tunnel-visioned and lost sight of the global health landscape. I needed to zoom out, recalibrate, and understand how global health (and digital health in particular) had evolved and how different contexts compared to rural Bangladesh.

I did that first through a thorough literature review, voraciously reading as many (hundreds) papers, publications and reports I could get my hands on which seemed relevant to our broader mission of ensuring high-quality comprehensive care to billions (I can’t thank my wonderful colleague Bridgette enough for supporting me through that painful process). I then started synthesizing the recurring themes and triangulating them against my own experiences from mPower and Jeeon. I also started to reach out to my community of expert mentors (many of whom are in this distro - you know who you are! 🙏🏾) to validate and find gaps in my learnings, and distill it down into a set of key challenges that are fundamental yet unsolved. Finally, I conducted some field visits (both in Bangladesh and neighboring West Bengal) to ground-truth and stress-test these hypotheses with various frontline health organizations.

Initially, I felt quite lost to distill the immense volume of information. Ultimately though, through this iterative process over ~6 months, I arrived at this deceptively simple summary of the most critical problems to solve in global health, taking a first-principles lens:

Each of these areas deserves its own post to cover the research and emerging solutions. Here, let’s frame and articulate these gaps and their intersections in just a bit more detail:

People-centered health systems (or lack thereof)

At an overarching, philosophical level, I realized that we have not really walked our talk about “people-centeredness” within our health systems. Human Centered Design (HCD), promising as a methodology as it was, became yet another jargon within an ever-expanding checklist to tick-off when pitching funders for money. We hardly internalized it as a philosophy, and stopped short of embracing as a core value that it is important to yield power and voice to our clients (or as they are tellingly termed in global health, “beneficiaries”).

However, if you genuinely approach healthcare from a people-centered lens, by definition you would have to design health services and systems that meet people where they are. This is why we designed Jeeon the way we did, leveraging people’s existing touchpoints and trust with pharmacies, as I wrote in my last post. Contrast that with the 18,000 community clinics built in Bangladesh spending billions, which are so clueless about people’s needs that they remain open (in the ~50% of cases they open at all!) from 10 AM - 2 PM, when most people are either tending the fields or cooking their meals. The busiest hours of pharmacies are usually before 11 AM and after 5 PM.

Meeting people where they are from a health system design perspective means investing in robust, responsive, comprehensive and effective healthcare systems that reach into their homes and communities. This is hardly news — we have known since Alma Ata that well-functioning primary care systems should be the centerpiece of any country’s health system, can address 90% of a community’s health needs, and save 60 million additional lives per year. Yet, in our single-minded —and dare I say, outdated— focus as a global health ecosystem to eradicate fatal infectious diseases, and the inertia of the resource allocation processes we have developed around it, we have categorically failed to address this fundamental need (and indeed, human right) of communities.

During my research, I was honestly pleasantly surprised to find the WHO (along with numerous other voices) advocating for the same people-centered philosophy I have always believed in, and equally horrified that there was so little departure from the status quo in the last eight years since the adoption of the Framework.

In terms of bang-for-buck, I therefore decided that continuing to lend my voice to this chorus — and focus my energy and efforts towards mainstreaming a people-centered paradigm in healthcare — still made the most sense.

***

When you double-click into that HUGE box and look beyond the philosophical and into the operational, however, there’s at least three big, interrelated and largely unsolved challenges that needs addressing, each of which is emphasized within the WHO’s vision, and which I had also personally run into at Jeeon many times over during the past 10 years:

1. Integration of care

Due to fragmented public health programs and the multitude of providers in mixed health systems, people receive disjointed and uncoordinated care. I concluded that we have done slightly better with vertical integration of care (within certain disease silos) than horizontal integration across various vertical programs and sectors, but that huge gaps persist in both dimensions:

1. Vertical integration: We have done relatively well in this area, especially within the disease-focused silos of most global health programming. In order to combat the HIV epidemic, for example, systems needed to be designed to screen, confirm, treat, and follow-up on the same patients effectively and over time (as represented in the 95-95-95 goal for 2030). However, as soon as you go beyond the handful of priority diseases, this integration breaks down completely. The same person found to be diabetic during an unrelated visit to a public hospital might therefore be rediscovered years later by a CHW belonging to the same public healthcare system. Few countries have well-designed triage+refer mechanisms or data sharing (especially for open-ended general health conditions) between the various providers even within the public sector, let alone across sectors. Last but not least, there is often no mechanism for reverse referrals, i.e. the ongoing follow-up and monitoring of care at the community level after one completes a visit or procedure at a facility, such as for post-operative care or chronic diseases.

   While building Jeeon, we were faced numerous times with the dilemma of a patient who needed to be treated at a facility, but without a 1:1 agreement and data sharing arrangement set up with each referral facility (which was resource and time intensive), there was no guarantee that the patient would be treated (and treated well) and referred back to us for ongoing follow-ups.

2. Horizontal integration: At each level (primary, tertiary, etc.), there are often a variety of providers to choose from, sometimes even from different sectors (public, private, faith-based, non-profit, informal, etc.). At the community level, for example, there are typically public/social sector community health workers (CHW)/nurses/midwives, private pharmacies, faith-based healers, and even friends-and-family as informal providers of care. People typically make choices on a case-by-case basis based on availability, urgency, cost, and a mental model of who-is-good-for-what. For example, within the window of a single pregnancy, a woman would typically opt for routine ANC/PNC visits by the health worker, handle acute symptoms like bloating or back-ache with pills “suggested” by the local drug shop, share depressive or mood disorders with peers or the local priest, and seek nutrition and lifestyle counseling from the mother-in-law.

   Without data sharing and integration of care across these various sectors and types of providers, continuity of care breaks down completely, causes confusion, and leads to duplication and inefficiencies. I have seen this first-hand in Bangladesh, where different providers (two departments within the Ministry of Health, as well as BRAC and other NGOs) might send CHWs to the same households for largely overlapping tasks around a pregnancy. At Jeeon, we were catering to many of these same pregnancies (e.g. for pregnancy-induced back-pain or gestational diabetes), but there was no data sharing or coordination between our systems, so these discreet but related encounters remained completely invisible to each other.

   Horizontal fragmentation is no doubt perpetuated and exacerbated by the fragmentation of global health funding, which might introduce thick walls and “not-my-problem” dynamics between providers. As a result, each program might screen for a particular condition it is designated for, but fail to account for or refer to adjacent ones.

Not only is this fragmentation a frustrating and confusing experience for patients, it also directly causes poor quality care and health outcomes. And it is intricately linked with the incentive structures that animate each individual provider and/or system.

2. Quality of care

We have long obsessed about access as our primary goal. This could be true in some contexts where low density of rural populations might mean long distances to any sort of care (although studies in India have shown otherwise). However, as I experienced viscerally in rural Bangladesh, people often don’t suffer because of lack of access, they suffer despite it. Once you count all the various types of providers ignored by the health system, people more often than not have access to some health provider within 15 minutes from their house. Many informal providers even do house calls, and will readily show up at 3 am in the morning when your child is sick.

If it isn’t access that is causing the problem, what is? According to the flagship 2018 report by the Lancet Commission on Quality, up to 15% of all deaths (8 million people) in the developing world are attributable to poor quality of care.

A few data points may be illustrative as to how this typically plays out:

• The average doctor consultation in Bangladesh lasts 48 seconds, which is the lowest among 67 countries surveyed. The closest second was India, at ~2 minutes! (study)

• In Madhya pradesh (India), as many as 63% of interactions in PUBLIC sector clinics happened with providers who had no medical qualifications! Moreover, there were only minor differences between trained and untrained providers in terms of following clinical guidelines. (paper)

• Indian public sector doctors spend 65% more time with patients, diagnose them better, and are more likely to offer correct treatment in their private practice than THEIR OWN day jobs at public clinics. (study)

• 70% of primary care interactions in Bangladesh happen with the informal private sector, which is completely unregulated, untrained/unaccredited, and operate in a vacuum with no support or oversight. 95% of health providers in Bangladesh are in the informal sector. (Bangladesh Health Watch)

So an extremely simplified version of the story is this: public sector care in these countries are often poor quality, which drives people to seek care in the unregulated private sector, which offers equally poor clinical quality but more convenience and generally a better experience, which people find worthwhile enough to pay for out-of-pocket.

Indeed there is a gap in the number of qualified medical professionals in the absolute sense, which causes some of this quality gap, but that is clearly not the entire story. Firstly, qualified physicians tend to crowd urban areas and rarely visit their rural postings (“absenteeism”) causing huge density variances. But it is not even clear that qualifications translate to clinical skills, and that those skills translate to effort. In other words, many providers, even qualified ones, don’t know well how to treat patients (or to treat patients well). Lastly, even when providers know what to do, they often don’t do it. This will lead us to the problem of incentives in the next section.

What complicates the quality problem further is that measurement of quality is not directionally consistent. In education, reading scores are good when going up, and bad when down. In healthcare, contrastingly, no medicines might mean poor quality (inadequate care) in one case, while excessive medicines may indicate poor quality in another. One imperfect solution to this may be to rely more on subjective patient experience and self-reported outcomes. However, although the Lancet Commission on quality strongly advocates for including such measures in health system performance assessments, to the best of my knowledge there are still no well-validated and consistent indicators measuring patient experience that can be applied universally.

3. Misalignment of incentives driving poor quality and fragmentation

It should be fairly evident by now what I mean by incentive gaps. Just to clarify, I don’t mean just the tangible financial incentives that are misaligned, but also the softer elements of the “choice architecture” and non-financial motivations (reputational considerations, peer pressure, etc.) that providers are embedded in. Both the integration and quality gaps are ultimately fueled and perpetuated by a wide variety of incentive misalignments.

At the micro level, public providers, who are paid irrespective of performance (or even presence for that matter, lack incentives for quality care. Private providers, while better incentivized to deliver services, have other forms of perverse incentives, such as the informal provider above who would clearly make a huge margin from the 95 injectable drugs he prescribes and then sells. Complicating this further, patients themselves often don’t have a good mental model for clinical quality, equating more drugs with better care, and demanding the latest antibiotics as quick fixes. In my hometown of Jashore, Bangladesh, the most popular private provider by far is a pharmacist who injects a potent “secret recipe” concoction of antibiotics, steroids and painkillers to every patient, irrespective of their condition.

At the meso level, there is rarely any incentive or motivation for providers across organizations to collaborate and coordinate with each other in care delivery, leading to the fragmentation challenges. For-profit clinics will often try to vertically integrate to hold patients longer than specialize and link up with complementary service providers. Public health officials will often see private clinics as competition rather than potential collaborators in ensuring health of the population, in the absence of performance metrics that cover the breadth of the health system (not just services rendered within public facilities). Last but not least, there are no incentives to share knowledge and learnings for accelerating innovation, leading to many reinvented wheels and repeated mistakes.

At the macro level, funding silos and poor measures of health system performance are the main culprit. Global health financing, as discussed above, perpetuates silo’ed thinking around diseases and not integrated life-course oriented care for populations. Payments are also typically tied to outputs (people served, widgets delivered, etc.) rather than effective measures of quality and value creation. But funders don’t deserve all the blame. Governments themselves are largely driven by their own perverse incentives — 5-year election cycles that lead them to prioritize shiny objects like building hospitals rather than meaningful but invisible system strengthening investments that pay off long-term.

I wrote about incentive misalignments further in the context of Antimicrobial resistance in the following post:

In addressing the above challenges, below are some key strategic focus areas that could yield high leverage and that I think we must explore deeply and systematically as an ecosystem (certainly not exhaustive):

1. Technology and business model innovations: What can a modern, holistic and integrated primary care system look like that fully leverages 21st century technology? How might we be able to tackle NCDs proactively at a low cost? Can we anticipate the impending “singularities” that AI will engender, and exploit it for furthering these goals?

2. Measurement and financing innovations for incentive alignment: How could we create enabling environments for value-based care delivery? For co-opetition within and across sectors? For more patient-centered thinking and design of systems? For tackling AMR?

3. Integration architecture: How should we design our health data and systems to be safe, private, patient-driven and interoperable? What kind of infrastructure and institutions might enable public systems and large funders to easily identify, test out and scale new and promising innovations in #1 and #2?

4. Global health financing and resource allocation processes: Can we start to shift how large health funders think about resource allocation? Can we move the conversation from incremental reallocations (“diagonalization”) to dedicated and significant allocation for a comprehensive redesign of integrated and holistic primary health systems?

I’m certainly under no illusions that we can solve all these issues globally in my lifetime, or that I can make a meaningful dent on any of them, let alone all. Tackling these issues is a collective responsibility for all global health actors.

However, it is useful to have a North Star orientation at all times to figure out how best to contribute at any given time going forward, hopefully for the rest of my career in global health, whether it is in ecosystem roles like at Endless, or entrepreneurial ventures like mPower and Jeeon (or the next one!).

In (re-)setting this orientation for what has long been my professional mission, I am reminded of the following quote from an incredible book called “Mastery” that I am currently reading:

For a master, the rewards gained along the way are fine, but they are not the main reason for the journey. […] if the traveler is fortunate — that is, if the path is complex and profound enough — the destination is two miles further away for every mile he or she travels.

For a mission as long-term, complex, evolving and unpredictable as this, I will do well to remind myself to make this more about the journey than the destination.

***

As I wrote in the last post, our journey started with an anthropological immersion in rural Bangladesh that opened my eyes for the first time to the day-to-day sufferings of people that are often unaccounted for in global health, but also to the tremendous opportunity of over 200,000 small drug shops and pharmacies littered across the country who account for more than two-thirds of primary care interactions and out-of-pocket health expenditures in the country. That’s up to 4M primary care encounters, EACH DAY! Not only do these drug shops (often called chemists or druggists in other contexts) dispense all forms of medicines (thanks to a highly competitive and aggressive pharma industry in the country), they are also often the descendents of “local healers” and “herbalists” who have served these communities for millenia, and hence are the closest thing to a primary care physician for a vast majority of the population, despite their lack of credentials or formal clinical training.

We posited that by leveraging this existing ecosystem that has deeply rooted trust and a vast number of footfalls cutting across every possible health condition, we could perhaps get closer to our mission than was possible through the public and NGO infrastructure alone.

Nearly 10 years and numerous wrong turns and dead ends later, we still believe strongly that this infrastructure can be a huge asset to any health system — depending on whether we understand, recognize and leverage it effectively. Indeed, many organizations, most notably mPharma in Africa and SwypeRx (formerly mClinica) in East Asia, are proving this hypothesis correct. Jeeon still continues to exist as a non-profit for that very reason, and this is also why Endless Health believes in upgrading, connecting and mainstreaming informal providers of care everywhere (more on that in the next post).

A chronological history

Phase 1: Telemedicine (2014-2016)

Excited with these revelations, we jumped headfirst into our first product — a telemedicine platform — within months of that original immersion. The idea was to connect these disconnected pharmacies with urban doctors to bridge the clinical knowledge gap, helping the former provide more accurate care for a wider range of conditions, while enabling the latter to increase their reach to remote areas from the comfort of their bedrooms.

The Doctor-in-a-tab service in a promo video

Not only did we launch a pretty sophisticated telemedicine service for its time, we attempted to do it (too idealistically and naively so, if I might add!) in some of the remotest and poorest areas in the North of the country which were submerged under water for half the year. Our rationale, simply, was that the need for a telemedicine solution was greatest in those areas that had the least access to formal health infrastructure.

We also had an elaborate vision of how this would form the nucleus of a broader strategy to link pharmacies to the rest of the healthcare ecosystem, unlocking other revenue streams for us in the process:

But as we found out the hard way, elaborate plans seldom pan out the way you expect! Despite some early traction and positive feedback from both pharmacies and patients (we had 84% positive patient reported outcomes and a consistent 70%+ net promoter score), we failed to eventually scale the model. Main among the reasons was that delivering positive outcomes requires the coming together of many discrete elements - medicine supply chains, clinical knowledge, diagnostic equipment, referral linkages, patient’s own health awareness, etc. - which were all more or less broken in the rural and remote environments we chose for our intervention. We found ourselves scrambling to solve too many problems at once, which drove up our costs significantly. On the other hand, rural Bangladeshis had very slim health wallets, and just recovering our costs would require charging prices that would keep the service out of reach for the vast majority (I wrote a ~50 page report on our learnings from telemedicine for a funder, which I encourage anyone to read to dig deeper into these and many other learnings).

This was a major turning point in our journey, because we had already spent most of our seed investment in perfecting and fine-tuning this model, and were forced to go back to the drawing board and start from scratch. In hindsight, the fact that we were structured as a for-profit also introduced undue pressures to figure out a business model quickly, which caused us to scramble for the next year or so without a clear vision of how to move forward. Many team members were also highly emotionally invested in and attached to the telemedicine model, so we lost some key people and morale dropped significantly.

Phase 2: Prototypes (2017-2018)

Over the next two years, we did what we probably should have started with in the very beginning, i.e. prototype a variety of different products and services to systematically explore the “design space” and find product-market fit. We did this in part through a “skunkworks” like team of our most enthusiastic, creative and energetic people with the mandate of identifying nuggets of opportunity based on our learnings to date, and trying out rapid prototypes of those ideas. Most of these ideas seemed promising initially, but to our utter dismay, turned out to be dead-ends for various reasons:

• A clinical decision support system for “village doctors” to use during patient encounters: They did not want to look at a screen during patient interactions, which they considered detrimental to the eye-contact, warmth and natural flow of the encounter.

• A debt and cashflow tracker for drug shops (“Haal Khata”): We noticed that many pharmacies maintained paper credit-books, and we thought this might be an interesting problem we could solve. Turns out credit recovery was never their aim; they write off most of this debt (often amounting to ~2 months worth of income) at the end of the year anyway, through a festive celebration during the Bengali new year, as a way of building goodwill in the community. The cashflow and accounts management feature failed because the transaction sizes were too small for the users to key it in for each patient.

• A e-learning app for “village doctors” to build clinical skills and knowledge: During our telemedicine model, we had curiously observed that pharmacies would secretly xerox doctor’s prescriptions and study them later to learn treatment approaches. Indeed, when we launched this prototype, we quickly reached about 5000 users within a month. Many users used the platform for hours a day, thanks to its built-in Duolingo-like leaderboards and practical case-solving games simulating real patients. However, the engagement dropped off a cliff after the second month once the novelty wore off, likely due to lack of an end goal (e.g. certification or accreditation), preventing us from monetizing the platform through sponsored content or product recommendations as we had initially planned.

• Non-pharma product sales and associated services: We started off with eye-care, successfully piloting with VisionSpring a model to screen for common eye problems and selling reading glasses (for near-vision disorders) in the process. We quickly reached 35,000 screenings and 15,000+ pairs of glasses sold with only ~150 pharmacies, but the supply chain was too complicated (due to the high number of SKUs of prefab glasses) and expensive to set up (buying vehicles, etc.) in the absence of reliable third-party logistics partners, and the pipeline of additional products that would fit this market was also too dry.

• Referrals to surgeries and other procedures: We successfully demonstrated the potential of using pharmacies to identify and refer cases that required tertiary care, such as cataract and cleft deformities. This also had some revenue potential, but did not generate enough volume and flow to justify investments in building out the nationwide referral network.

Phase 3: E-commerce (2019-March 2020)

Ultimately (and to our great relief!), we found a prototype that started showing traction, which was an ordering solution (branded “JeeonConnect”) for small and hard-to-reach pharmacies to procure quality drugs directly from manufacturers. We decided to double down on this e-commerce approach, had a couple of successful trials connecting ~3000 pharmacies with a renowned pharma company, and by early 2020, secured a contract with the company to aggressively scale up the platform to 20,000 pharmacies by the end of the year. We were energized after a long time because this model not only had a direct revenue model (a commission on every order, compared to the indirect business models of all the other prototypes), tapped into the core business of pharmacies and hence had a relatively large addressable market, relied on the pharma company’s own distribution and hence was not capital intensive, and would get us to a meaningful enough scale to hopefully attract other pharma company partners, raise our Series A and layer on other patient-focused services (the vision for how this would all come together was represented in the 2019 promo video below). Not to say all was a bed of roses with this model (there certainly were thorns!), but it was certainly the closest we had ever been to a scalable business model that could get us to break even.

Phase 4: COVID Response (March 2020 - 2022)

In an ironic twist of fate, just as we were ramping up the roll-out of JeeonConnect, the first cases of COVID were identified in the country, and within a few days, the entire country went on lockdown.

Although we anticipated an increase in sales on our digital platform due to the restricted movement of pharma sales reps, this hardly materialized. Apparently, sales reps were unable to reach doctors for their usual marketing promotions, so they went all-out trying to push sales with pharmacies directly, regularly visiting even the smallest and hardest-to-reach shops that were JeeonConnect’s staple client base.

At the same time, however, we were also hearing earnest pleas for help from our pharmacy clients, who felt vulnerable to this unknown virus and deeply anxious without any access to authentic knowledge or protective equipment, despite a rapid uptick of symptomatic patients who they couldn’t turn away.

It was an extremely difficult choice, but one that we made unanimously as a team (and probably the one decision I’m most proud of in all my 10 years of running Jeeon) — to put our commercial operations on hold and invest wholeheartedly in COVID response work, irrespective of the existential threat it represented.

We launched an online certification course on community-based COVID response (which was accredited by the Pharmacy Council of Bangladesh) within 5 days, which ended up certifying over 25,000 community pharmacies nationwide. From our own coffers, we supported over 4000 pharmacies with PPEs. We worked shoulder-to-shoulder with the government and other NGOs and multilaterals in the Risk Communications and Community Engagement (RCCE) pillar of the country’s COVID strategy. Our engineers helped design the standardized national API that centralized millions of COVID-19 case reports from all corners of the country over the next two years. We conducted symptomatic surveillance through pharmacies in partnership with a Yale research group, disseminated behavior change content and up-to-date verified information to communities through our pharmacy network in partnership with BRAC, and helped the Savar health officials (a large industrial suburb of Dhaka) develop and document a highly effective COVID response strategy engaging communities and pharmacies directly. We eventually wrote up this latter experience into an Action Toolkit, which was published by Monash University, and since disseminated to all regional public health officials in the country as a best practice in community engagement during public health emergencies.

In the process of all these activities, we also built a massive database of over 120,000 pharmacies (complete with phone numbers), and over 35,000 pharmacies joined our closed Facebook community for access to authentic COVID related information and educational content.

Although we raised some philanthropic resources for all this work, by 2022 we had exhausted all our capital and without a path back to our commercial trajectory, were left with no choice but to convert to a non-profit, aided in no small part by the supportive generosity from our investors.

Phase 5: Jeeon as a Non-profit (2022 - present)

Today, Jeeon is a non-profit with a small team of part-timers and volunteers, who are keeping alive our large online community of pharmacies through regular virtual learning sessions on timely health topics, knowledge-based discussion forums and exchanges, and collective advocacy efforts for a more conducive regulatory environment. Some conversations with potential partners to leverage this network to roll-out various products and services are also under way.

Key Learnings and Takeaways

It is hard to summarize the thousands of big and small learnings from such a long and intense period of experimentation and discovery, but I have tried to pick the top three below.

Technology solutions must follow a deep understanding of the problem, not precede it

Like many technology-enthusiasts, we often got carried away by the “silver bullet” fallacy. We got excited about the solution before investing deeply in peeling the various layers of the problem to get to the root issues.

For example, with telemedicine, we were trying to solve the pain point people repeatedly emphasized in our focus groups — lack of access to doctors. As we rolled out the service, we started to feel that the underlying need was more about the quality and completeness of the treatment provided, which included the accuracy of the treatment, but also required a stock of quality medicines, access to diagnostics, and more. I did not realize until much later that the actual deepest need was for a quick, inexpensive fix to whatever problems poor people came with. Indeed, as we found through our research, people typically give a pharmacy 72 hours to “fix” the problem before switching to a competitor (and usually there are 10-50 such competitors in any rural bazaar). This was why drug shops as they operate now — delivering the two-day dose of the most potent antibiotic or steroid in their stock, sometimes even on credit — was such a “perfect” solution.

In hindsight, I’m not sure telemedicine would ever solve this particular need, unless you could get people to care about the long-term costs of these quick fixes, set up monitoring mechanisms on providers and impose punitive measures for mistreatment, AND offer an alternative that was just as convenient and inexpensive (cannot even imagine what that might be). In other words, unless you can change poor people’s definition of “quality” or address their desperate need for a “quick fix” in some way, it will be really hard to design a private sector service (technology-based or otherwise) that truly delivers health outcomes.

Similarly, in our excitement seeing the early traction on our e-learning app, we forgot to ask what was the deeper pain point this was meant to solve, which would keep them coming back in the long term. The lack of good learning content was surely an issue, but the much more urgently felt pain point was a lack of formal recognition from the health system as primary care providers, which would only be solved with some sort of a government-issued accreditation. Then again, a one-time accreditation would not suffice in solving the problem of long-term engagement — you would require regular renewals of the accreditation based on units of learning or requalification exams, similar to CME requirements for doctors in advanced economies. This was not a technological problem but a regulatory one, and hence was way beyond the scope of a 30-person startup.

(As a tangential but important point, it did not help our cause at all that the fundamental technologies we were banking on (particularly rapid diagnostics toolkits) did not materialize on the timelines we needed them to. With today’s price-performance of these kits combined with AI, it may have been a very different story. As they say, success in innovation is as much about timing as it is about the idea or the execution.)

Incentives reign supreme

Having come from the experience of working only with government and NGO health workers, who are salaried employees that tend to do what their managers ask of them, I was completely oblivious to the complex incentive structures private sector players like drug shops operate within. Not only do they have direct financial interests and motives as a micro-business, they also care deeply about their reputation in the community, being perceived and even addressed by the local community as their “doctor”. As a result, for example, they were often reluctant to encourage their own patients to opt for a telemedicine consult, since it made them look less confident in their own skills.

On top of this were the explicit bonuses and pressures from pharma companies to sell the latest branded drugs, the significantly higher margins often associated with poor-quality medicines, and the strong implicit expectation from patients for the drugs that would most swiftly relieve their symptoms and allow them to get back to work. This was indeed the perfect storm of perverse incentives!

This remained a blind spot for us for a long time because with telemedicine, we were initially trying our best to optimize value for patients. We were late to realize that we needed to align with the pharmacists’ incentives more directly and solve their pain points, since they were actually our primary clients, not patients, even though we were trying to ultimately impact patient outcomes. This “aha” moment was partly the result of a really fortuitous 1:1 discussion with Bill Gates in 2016, where he insightfully pointed this out as a key flaw of our telemedicine model. (I wrote a piece reflecting on that meeting and imagining a follow-up conversation below). This was why we tried to address the pharmacy’s pain points more directly in all our subsequent models post- telemedicine.

But incentives were at play at not just the pharmacy. Patients had their own unique set of incentives due to their limited health education and the varied pressures on their limited time and resources. Even doctors in our telemedicine model responded strongly to simple patient feedback and satisfaction scores, especially when ranked on a public leaderboard against their peers. It opened my eyes to how much we as humans —experts or otherwise— operate mainly as a function of our surrounding incentive architecture.

As a corollary, it also continues to amaze me how little effort and sophistication we bring to shaping the incentive structures of our providers and systems, and still expect genuine effort and positive outcomes from them.

Aligning the 3 Ms — Mission, Market and Model — is critical

Our for-profit business model was inconsistent with the rural market we were trying to address, and our mission of delivering positive health outcomes at any cost. When we pivoted to e-learning and then e-commerce, however, the market and models were better aligned, but we were straying further from the mission. We felt this deeply in our sagging morale and creeping self-doubt. It was only when we chose to let go of our for-profit orientation and decided to go after COVID response work with a philanthropic mindset, that our mission, market and model were finally completely aligned, allowing us to reach national scale and impact virtually overnight.

If I were to do it again with the benefit of this lesson, I would probably have started as a social business in a peri-urban market or a captive audience such as a factory, where people had more disposable income and greater health awareness, and oriented our mission towards a more narrowly defined problem (like low-cost diagnostics, or integrated diabetes care). If it were successful, it could provide the platform to form a sister non-profit that targeted more rural and difficult markets with a philanthropic approach.

***

If you survived till the end of this long post, my sincere thanks. 🙏🏾 Please comment (on the Substack website) or email (replying to this mailer) with your reflections or pointing out blind spots I may still have. I would also appreciate if you shared this with anyone else who might find it interesting.

I started off on that journey first with ClickDiagnostics Inc., a startup founded out of Harvard/MIT in 2008 that envisioned using the first generation of smartphones as a platform to extend telemedicine services to remote and underserved areas of the world. Fresh out of college, having a background in both technology and public health, and inspired by the call to young people to build “Social Businesses” by Prof. Muhammad Yunus, I joined the stellar founding team and designed/led some pioneering projects in the as-yet unnamed field of mHealth, mainly with BRAC in Bangladesh but also for a period with the Egyptian Ministry of Health. Yet, as the projects and team grew and we rebranded (to mPower Social Enterprises) and established ourselves as a trusted name in digital health, a feeling in my gut grew stronger and stronger that however scalable and impactful the model of working through large government and non-profit organizations was, the way we were approaching it was not going to be the full and complete answer to that North Star question. Yes, I was frustrated with the slow pace at which large bureaucracies moved. I was certainly disillusioned by the politics, power dynamics and top-down nature of the global health industrial complex. I spoke about these frustrations at length subsequently as in the Keynote video below (and this 2015 Guardian column and a recent Aspen Institute book on Decolonizing Development), but I still did not fully understand the reason behind that disconnect as I had only seen things from the system’s perspective.

It was around this time of a nagging but subliminal discomfort that I met Matt Dalio for the first time in 2012, in a small buzzing cafe in Cambridge, MA. It was an energizing conversation to say the least - we were both 20-something year old budding entrepreneurs from vastly different backgrounds; yet in that very first meeting, we left with a sense that our life missions were aligned at some very deep level. Despite our formal relationship since evolving through many stages (through funder-researcher, investor-investee, and most recently employer-employee), that deep comradery and friendship we formed in that first meeting has luckily continued to this day.

Matt supported our first research project at mPower to do an anthropological “deep dive” (using then cutting-edge HCD methodologies) into rural healthcare in Bangladesh to understand it from the people’s perspective. It was during those two weeks of intimate living in the homes of rural Bangladeshis, hearing their deeply human dreams and aspirations and frustrations, and accompanying them on their trips to pharmacies and hospitals, that I first realized the source of the disconnect I had been feeling all along.

The short answer was this - our health systems simplify poor people’s health as a narrowly defined set of diseases to solve, and completely loses view of the complex and ever-changing tapestry of physical and mental health needs, aspirations, constraints, sufferings, etc. that every human (rich or poor) inevitably goes through in the course of their life.

Often, this focus on “sick-care” is further limited to those conditions that can kill people. Indeed, in mPower’s projects with BRAC and the Ministry of Health in Bangladesh, we were technologically upgrading programs exclusively focused on pregnancies and childbirth, one of the major drivers of mortality and hence a key component of the Millennium Development Goals (MDGs). Yet, just like you and me, poor people don’t just have occasional episodes of Malaria, Tuberculosis and Pregnancies and gratefully thrive through the intervening periods. It’s the banal but painful daily health issues - the back pains and persistent itches, the insomnias and panic attacks, the vertigo and light-headedness preventing one from going to work, and the sick 2-year olds at 2 AM in the morning - that cause an inordinate amount of suffering without due recourse from the health system.

(No doubt this was originally an intentional choice, since there are a lot of good reasons to simplify and focus - resource constraints, clear metrics for impact, lack of data and infrastructure to handle the complexity, etc. We have also certainly made a lot of progress, such as with maternal mortality and HIV, due to this “vertical integration”. However, if my experiences in global health over the past 15 years is any guide, we have lost sight of the original reasons why we made this choice, got stuck in our jargons and processes, and completely lost touch with people’s lived realities. These are also deeply interlinked with the colonial roots of global public health, the political economy of power and voice, the lack of easy ways to capture people’s needs, preferences and feedback, and the lack of incentives for providers and system designers to even consider it important to listen to patients’ voices. Each of these issues merit their own deep dives, and I will hopefully come back to these topics in future posts.)

In that and subsequent visits, I met a man who had broken his shin in a construction job, was forced to spend his life savings and sell family land to get 3 poorly performed surgeries at shady and exploitative private clinics, and still limped after 2 years without being able to get back to work. I met a bright 13 year old girl who stopped going to school after a psoriasis on her scalp caused her to lose all her hair. I encountered an old lady whose internal organ systems were all but decimated after getting administered powerful steroids by her local drug shop continuously for 12+ years to manage her chronic back pain (she died soon afterwards). These, like countless others, were the stories that were lost when you obsessed over the number of maternal and child deaths, and patted yourself on the back for the cases of HIV and TB treated. No doubt deaths from such preventable causes are tragic, and urgent needs we had to address as global health practitioners to begin with. But we are fooling ourselves if we believe we are building a real “HEALTH-care” system.

That same visit also made it very clear to me that health systems, however well-meaning, often can’t replace the deeply trusting and human connections people have with their existing health providers and caregivers, whether that is the local midwife, chemist/druggist, traditional healer or the “village doctor”. This is due to a variety of reasons - both push factors (e.g. distance and cost of reaching care, the uncertainty that services and medicines will be available, the inherent power dynamic and social hierarchy between doctors and patients in a country like Bangladesh, the lack of incentives of public sector providers to do their best job, the lack of accountability for providing poor quality care, etc.), and pull factors (the longstanding trust and comfort with local providers, the convenience and proximity, their flexibility with payments, the range of services and products offered, etc.). No wonder therefore, that 70% of the population in Bangladesh still visited pharmacies as their first-point-of-care despite 18,000 government “Community Clinics” that had been built all over the country. By completely discounting existing providers as “quacks that will soon be irrelevant” and blithely building new facilities and cadres of providers, we were at best wasting resources and at worst relegating a vast section of the populace to unmonitored, substandard care.

This realization that we need to build solutions that meet people where they are, is what led me to start Jeeon, with the mission of figuring out how to leverage the vast existing infrastructure of informal community pharmacies across Bangladesh to provide high quality care to rural and underserved communities, and what convinced Matt to join our journey as an investor.

(Next up: I will write about our journey with Jeeon over the next 10 years and our main takeaways in Part Two of this Origin Story series)

***

If you have enjoyed this post, I only request that you leave a comment sharing your own stories, or why you disagree with any specific points. My intention from this is as much to reflect on my own learnings as to learn from my community. Please also forward to anyone else who might enjoy this post.

You’re receiving this email because you decided to subscribe to my new Substack (warm welcome!), were subscribed to my monthly Jeeon newsletter that ran from 2016 to 2022 (welcome back!), or in our recent 1:1 conversations demonstrated genuine interest in my/Endless’s work and pursuits (🙏🏾). I humbly invite you along on my latest journey towards deeper understanding and impact.

If you ended up here by mistake or for whatever reason do not want to receive my emails, my sincere apologies - please feel free to unsubscribe, or send me a brief note so that I can take you off.

Although you aren’t likely to be here unless you already know me, a brief refresher nevertheless - I’m a health innovator, data scientist and tech entrepreneur from Bangladesh, helped build two social enterprises (mPower Social Enterprises and Jeeon) and learned a $h!t ton from their successes as well as failures, and in 2023 joined Endless Network, a family foundation focused on transformative impact through a people-centered approach to technology in health, education and livelihoods, where I lead its global health work. I live in Baltimore with my wonderful life partner of 20+ years, Sarah, and our two lovely kids, Tahrir and Aarya. Here’s a recent selfie:

As Director of Health Programs, my main job is to source innovative organizations and projects for grant-making at the intersection of global health and technology. But I have a secondary title of entrepreneur-in-residence, which means I also have the rare opportunity to pursue my passions to design and build new projects or products (most recent one was a RAG-based chatbot to organize global health knowledge), and play out my inner activist in pushing for system-level reforms and paradigm shifts at the global stage (this is by no small measure facilitated by the fact that I now wear a funding hat, an unfair privilege I feel both deeply grateful and uncomfortable about!).

Over the past 15 months, I’ve come to realize what a rare and unearned privilege it is to get to work with some of the best and most innovative organizations around the planet, learn from the leading experts in any sub-field of global health, and after 15 years of non-stop firefighting (which I still very much miss, however weird that sounds!) finally have some space and time to read, reflect and write on topics that I deeply care about.

Just a small sample of the biggest questions and interests on my mind these days:

• What would it take to build a truly people-driven (going even beyond people-centered) health system that puts them on the seat of active agents and shapers, not passive recipients, of their own health and wellbeing journeys?

• What role can emerging technologies - especially AI and health sensors/rapid diagnostics - play in democratizing access to high-quality healthcare even in low-resource settings?

• What are the fundamental information asymmetries, blind spots and incentive misalignments that prevent people from taking optimal care of their own health, providers from doing their best and treating their clients as people not patients, organizations from integrating their services across practice areas and with each other, and health funders from financing long-term holistic approaches to health and wellbeing - and what can we do to jolt them to more positive equilibria?

• How can we pivot health systems from their historically passive “sick-care” approach to a proactive “health and well-being” approach that will sufficiently address 21st century health problems like chronic diseases and antibiotic resistance? In other words, what should a modern, 21st century, people-centered, community-based, integrated, holistic, proactive primary care system look like?

In this newsletter, I intend to write about some of these frontier issues on a bi-weekly cadence — to honestly, systematically and vulnerably reflect on my own candid (and likely unpolished) thoughts and learnings, with the goal of fostering more conversation about these topics and learn from the hive-mind to refine my own hypotheses. I will therefore greatly appreciate if you leave a comment on my posts with your thoughts (even when, and especially if, they are critical of my view), and point me to relevant resources and/or experts to learn more from.

I truly hope and wish that you get a lot of value from reading along and reflecting together with me and this community. Welcome once again!

Disclaimer: While some of these same issues form the basis of Endless Health’s strategy, this newsletter will be a personal one and should not be assumed to represent Endless or Dalio Foundation’s views on these issues.

Lastly, some housekeeping…

You will start receiving updates right in your inbox, every couple of Fridays. You can also access the full archives anytime on my Substack webpage, or a clean ad-free reading experience on the Substack iPhone/Android app.

If you can’t find the newsletter, check your spam or Promotions/Updates folders, and move this email to your primary inbox (or mark “not spam”).

Central to a Duo’s efficacy is its ability to access and analyze all of a Prime’s data, encompassing sensitive information such as medical history and financial records. With this all-encompassing knowledge, a Duo can make informed, personalized decisions on its Prime’s behalf. However, the centralization of such vast personal data raises significant security and privacy concerns.

We’ve seen how disastrous data breaches can be with incidents like the Facebook data breach, where the personal data of over 530 million users was exposed. The risk of similar breaches in a world with Duos is exponentially higher, threatening not just privacy but our collective future. Imagine such data falling into unscrupulous hands, enabling dictators to conduct mass surveillance or manipulate elections.

Furthermore, big tech companies whose business models revolve around data monetization might not be the most suitable architects for our Duos. These companies have strong incentives to use a Duo’s access to personal data for advertising or influencing a Prime’s behavior.

Most importantly, we should emphasize data ownership. In a future with Duos, we must insist on maintaining full control over our personal data, and not allowing a Google or Microsoft to monetize our digital clones.

The solution lies in leveraging Web3 technologies. A blockchain-based architecture allows the Prime to maintain control over their encrypted data, permitting their Duo access in a traceable manner. Data decentralization ensures no single entity can access the complete database, providing a robust safeguard against data breaches.

Additionally, for Duos to replace our need to visit websites or platforms, they must securely interact with specialized AIs. Utilizing smart contracts on the blockchain, Duos can transfer necessary information to these AIs for the task’s duration and revoke access once completed, preventing data misuse.

In conclusion, while personal AI agents hold transformative power, it’s essential to proactively regulate and design these systems with privacy and security as primary considerations. By placing human beings, the “Primes,” at the center of this process, we can unlock the full potential of this technology and ensure our AI works for us, not a centralized tech giant.

Often when I visited a doctor in Bangladesh and posed questions about my illness, demonstrating some prior research, the irritated response I often received was, “Did you hear that from Dr. Google?” For better or for worse, Google (and later social media) has transformed the way people seek health information over the last 20 years across the world. But as AI systems like GPT-4 begin to ace medical exams like the USMLE (and also the Japanese MLE), we’re entering a whole new era where accurate, up-to-date and personalized medical advice is soon going to be universally accessible.

When I started my career in global health ~15 years ago, medical knowledge was doubling roughly every ~4 years, which (if the rate held constant) meant that a doctor graduating in 2010 and practicing for 40 years would have to absorb 2¹⁰ or 1000x the medical knowledge over their career than what they learned in medical school if they were to keep their know-how completely up-to-date.

But that rate of growth is not constant. In 2020, it was estimated that medical knowledge was doubling every 71 days. Just for comparison with the previous number, a medical student graduating in 2020 would need to learn 2²⁰⁰ (an exorbitant number) times the information they learned in medical school over the next 40 years to remain an all-knowing super-doctor. And that rate is going to get even faster due to the nature of exponential growth, so in reality, the amount of knowledge to be absorbed would actually be many times more than that.

Clearly, no doctor will ever stay on top of this deluge of information, and it’s hard not to feel sorry for them for even being expected to stay on the cutting edge of their craft. Indeed, this is the underlying reason we are seeing hyper-specialization in medicine — the human brain hasn’t evolved for thousands of years, and it just can’t process so much information. And the hyper-specialization leads to silos of knowledge, which often comes back to hurt patients by not being evaluated and treated holistically.

AI, on the other hand, excels in exactly this. GPT-4 was trained on (simplistically) all of the Internet, which it can draw from as it forms its responses. Even without any specialized training, it already passes the medical licensing exams in US and Japan, and outperforms human doctors in responding to common medical queries. Already emerging are more specialized instances of GPT-4 such as BioGPT which has been trained on millions of medical journal papers, and HeyPi which will act as your mental health companion (it’s quite good actually, based on an hour of engaging and thought-provoking conversations!).

As soon as someone figures out how to train GPT-4 on millions of actual patient medical records plus all of the medical literature (which is not a question of if, but when — my bet is this happens before the end of this year), it will suddenly be the best doctor the world has ever seen. And that same technology will be available at the touch of a button in the pocket of a teenage girl in Kibera as much as it will be in the hand of Bill Gates.

Most discussions about AI replacing doctors are situated in Western contexts, where doctor shortages are less severe. But in places like rural Bangladesh, where there’s only one doctor for every 10,000 people, a scalable AI is not actually replacing doctors, it suddenly and magically solves the perennial doctor scarcity, and that too with a super-human clinician who provides better and more accurate medical advice than even the best human doctor in the world.

Let us also not beat around the bush. Human doctors ain’t always that great! As a global health practitioner, I’ve often faced the heart-wrenching reality of human doctors making lethal mistakes or acting against the best interest of patients due to misaligned incentives. The know-do gap is a persistent issue in many low- and middle-income countries. Imagine a world where AI-driven medical diagnoses and treatments eliminate these challenges, providing consistent and optimal advice based on a patient’s comprehensive health history. No longer will a patient get a RESERVE class antibiotic just because it was the most profitable thing for the doctor to prescribe! This could be a game-changer for global health.

So far, it may sound like I’m a blue-eyed, naive optimist without a sense of the limitations of AI. When I ran Jeeon, I also had high hopes for AI-driven healthcare solutions like Babylon and IBM Watson, then grew disillusioned as they failed to deliver the expected impact. That is why my first reaction to GPT’s potential in medicine was one of skepticism — I was very concerned about “hallucinations” causing egregious errors which could be fatal in a medical context.

However, if you notice the rate of improvement (GPT3.5 scored 58% on USMLE, while GPT-4 scored 88%), it is clear these errors will become much rarer with time. After playing around with and comparing GPT 3.5 and now GPT-4, I am now convinced that a new round of disruption — this time, a genuine one — is coming to medical practice within a very short period of time.

Black box AIs like LLMs may also be problematic from a regulatory perspective, but once they empirically prove their superiority over human doctors, the paradigm will shift swiftly. Like driverless cars, as soon as AI proves to be more accurate and safer than humans (even though they are not entirely foolproof), there will be a strong case to be made to ban humans from driving cars, and yes, treating patients unassisted by AI!

So am I saying there will be no role for human caregivers? Absolutely not! I envision a future where AI and human healthcare providers work in tandem to offer accurate and empathetic care, with AI ensuring technical correctness and the human ensuring a great experience. Picture a village in rural Bangladesh where a community health worker, armed with an AI-powered diagnostic tool, can detect and treat a child’s illness early, ensuring they grow up healthy and strong. Or imagine the elderly man who can consult his AI personal health assistant (connected to his $5 activity tracker) to manage his diabetes, enabling him to spend more quality time with his grandchildren. Or the depressed adolescent who has a mental health companion 24/7 to listen to them and help them process their suicidal thoughts. There will certainly be a role for highly specialized surgeons and ER physicians (at least until robots have their own ChatGPT moment), but that should be a far smaller sliver of the health system than self-driven and community-based care that is continuous, personalized, intuitive, perceptive, and proactive.

When I envision global health in 2050, I therefore see an army of community-based providers offering personalized, AI-supported preventive and primary care. As for self-care, individuals will have access to an evolved version of Dr. Google, an AI-powered personal health assistant, enabling them to maintain their physical and mental health at a fraction of today’s costs.

In other words, for the first time in history, this may allow global health to move from sick care to true holistic and wholesome “health care”.

In this new world, the role of humans will evolve, focusing on providing empathetic care and the personal touch that AI cannot replicate, at least not yet. Medical training and education will need to be reinvented, emphasizing the importance of community care providers, nurses, midwives, mental health counselors, and elderly care providers in the community, and focusing on soft skills instead of algorithms and protocols.

I know that this still sounds a bit like science fiction, or worse. But it is important that we wake up to the opportunities this moment opens up for global health. Adapting to this new reality, and truly making the most of it, will be complicated by vested interests, regulatory sluggishness, financing challenges, digital divide, and more. But the quicker we can recognize that a different kind of future is not only possible but desirable, the better off the world’s poor will be for it.

I recently had the opportunity to participate in the World Health Summit regional meeting 2023 in Washington, D.C. As a gathering of health experts and policymakers, the event aimed to discuss the global health agenda and pressing issues in the field. Yet, I couldn’t help but notice the irony in the title “Bridging the Science to Policy Gap in Global Health”, as the conference seemed to overlook the even wider gap between both science and policy on the one hand and practice and innovation on the other. The conversations at the summit largely tended to hover at a 30,000-foot level, with very little in the way of real challenges people across the world face in seeking healthcare, and concrete strategies or thoughts about overcoming them.

This experience was jarring because it made me realize a few things:

• Global health decision-making often occurs at a very high level, spearheaded by medical researchers and policymakers who may have been removed from treating real patients or implementing projects hands-on by decades.

• If the fact that I was one of very few implementers at the panels was any reliable indicator (or that my talk failed to elicit any questions from the audience), there also doesn’t seem to be a lot of meaningful interaction between entrepreneurs/innovators/ implementers on the ground and the policy-making crowd at such conferences. This exclusion of voices like mine, I suspect, leads to a significant gap in the relevance and fidelity of policies, as we are the ones who can provide valuable insights based on our hands-on experiences. A simple example is the fact that for most people in the audience, it came as a surprise that 70% of first-point-of-care interactions happen in the informal sector. How can we make policies when we don’t even know what the ground reality is?

• Given the average age of panelists was certainly above 50 (not that I’m too far away from that anymore!), and how little talk there was about the recent quantum advances in generative AI like ChatGPT (aside from the sole panel on technology itself — which was pretty great!), I also wondered if there is a real appreciation (or apprehension) in policymaking circles about the possible disruption that’s coming as an entire generation grows up trusting their GPT-Doc more than any real doctor.

In the past, I have often noticed the same issue in academia, where a lack of direct collaboration with practitioners often results in a disconnect between theory and reality. There is also a disconcerting anathema in academia about studying and publishing failed interventions, while the majority of interventions are bound to fail and hence constitute the majority of potential practical learning opportunities.

As Simone Weil, the French practitioner-philosopher once said, some people live their lives exclusively in the abstract, rather than the real. Marx hypothesized about the working class leading a revolution against the bourgeois, without ever truly experiencing the dehumanization the working class faces in the course of their life experiences which robs them of any and all revolutionary agency. Hiding behind abstract concepts and engaging in philosophical discussions can become a crutch for us if we are not careful, and can prevent us from truly understanding the world and the problems we face. In global health, this disconnect seems to be a significant barrier to progress.

It is also a wake-up call for me personally. Having moved to the US recently to take up a role in a family foundation, there is a real risk that I may lose my connection to ground reality rather quickly. Indeed, I spent the last six months zooming out into a 30,000-foot abstraction to come up with a clever summary of key gaps that need to be addressed! Ah, how comfortable it would be to continue to live in the clouds, go to these fancy policy conferences and share my hard-earned wisdom, and not ever again experience the blood, sweat, and tears that are involved with trying and failing and trying again!

But isn’t that the only way to make meaningful progress in the real world? Despite my new role as an ecosystem enabler, I must ground those theories in real practice on the ground, and work alongside innovators and practitioners to develop, learn from, and iterate towards better and better models of delivering care to those who need it most. By aligning myself with those who are actively working to improve healthcare for the poor and marginalized, I can probably hope to have a more tangible impact.

And perhaps, in doing so, I can also contribute to bridging the gap between abstract discussions and real-world solutions in this space.