Let’s explore the three questions sit at the heart of building an AI-native organization.

1. What constraints have changed?

An organization is, in part, a coordination mechanism built around constraints. We divide work because no one person has all the skills or time to do everything. We create functions because expertise is scarce, management layers to coordinate work, and approval processes because mistakes are costly.

So before redesigning an organization around AI, it is worth asking: What constraints shaped the current organization, and which of them are no longer as binding?

Marketing as a case study: For the past few decades, marketing organizations have been built around three constraints: (i) content is expensive to create, (ii) customer understanding is expensive to generate, and (iii) execution bandwidth is scarce.

AI weakens all three constraints. One marketer can now generate hundreds of ad variants or thousands of email variations. AI can synthesize customer interviews, sales calls, survey responses, reviews, and support tickets. Agents can increasingly build campaigns, analyze results, and optimize. When content becomes cheap, the new constraint is deciding what is worth publishing. When analysis becomes abundant, the harder question is deciding which insights matter. When execution gets easier, judgment about what to execute becomes more important.

2. Which silos no longer make sense?

Most work today happens through specialization. I know my piece well. You know your piece well. We coordinate across roles and functions to produce an outcome.

Consider software development. A product manager talks to customers and writes a feature spec, a UI/UX designer creates mocks, and front-end and back-end engineers build it. Every handoff carries a coordination cost (“That isn’t quite what I had in mind. Can we change this?”).

AI changes what one person can do. Someone who understands the customer problem may now be able to prototype, build interfaces, analyze data, and produce working code with tools such as Claude Code. If that person can own more of the outcome end to end, you need fewer handoffs and fewer rounds of "that's not quite what I meant." This may be AI's greatest organizational value: lower coordination costs.

Marketing case study: As the constraints in marketing are shifting, many of the previous silos don’t make sense anymore. Even within digital marketing, there may be separate people/teams for email, social, SEO, paid media, PR, affiliates, and analytics. Creative and brand teams often sit across these channels. Now a 2-person AI-enabled team can perform work that once required a dozen people across functions. The project no longer has to wait in a queue for another team to finish its piece. Human labor shifts away from coordinating with others to directing AI systems and exercising judgment. This also gives rise to new roles (heard of GTM engineers or marketing engineers?).

3. How should the transformation be staged?

You cannot rebuild an organization overnight. I see the transition happening in four stages.

Stage 1: AI as an individual productivity tool. People use AI and become more productive, but the organization changes little.

Stage 2: AI as an enabler of new workflows. Local productivity is not system productivity. You can make every employee 30 percent faster without making the company 30 percent faster. If the work still passes through seven functions, five approvals, and a quarterly planning process, the local gain disappears. In stage 2, work is redesigned around what humans and AI can each do well.

Stage 3: AI agents as part of the workforce. Persistent agents take responsibility for bounded outcomes, with access to tools and data and limited authority to act.

Stage 4: AI as the operating model. Teams, decision rights, planning cycles, and management structures begin to change around what AI now makes possible.

Marketing Transformation in Four Stages

I will again use Marketing as a case study to outline this transformation (mostly because I am immersed in it lately but you can apply the same ideas to any function).

1. AI as a Productivity Tool

Yes, your marketing team is using LLMs, but that doesn't mean they are using them well. In sessions with roughly 300 students in my Enabling Technologies class, I've seen stark differences in how people use the same models, and in the results they get. (check out part 1 and part 2 of my older posts on good prompting)

Good prompting is one piece of the puzzle. The next is to create persistent skills. A prompt is a one-time instruction but a skill is a reusable and evolving set of rules and preferences that AI applies every time you ask it to use a skill to complete a task.

Then comes context. Some companies call this a “company brain” or “customer brain”: a body of organizational context that AI systems can access when they work. AI performs much better when it has access to the information needed to understand the business. For example, can your AI access your brand guidelines, previous campaigns, customer research, product information, and examples of good work? Can employees share that context with each other rather than rebuilding it separately?

Finally, the AI needs access to tools. Once an AI system can work with your CRM, analytics tools, CMS, databases, and other software, it can move beyond giving advice and begin helping employees complete the work itself.

(If skills, connectors, and organizational context are new ideas for you and you/your team can benefit from tutorials on these topics, fill out this survey. If there is enough interest, I will organize a free four-part webinar to cover how to do these using Claude Code or Codex.)

2. AI as an Enabler of New Workflows

The next stage is to redesign the work itself. This is both a design and a change management problem. The design problem is simply that we know very little about how to organize work when everyone has access to capable AI systems. It will take a couple years of experimentation for us to get there. The change management problem is getting people to drop old ways and adopt new ways of doing work. I believe that AI-native workflows are often easier to build in net-new areas than in functions that have operated the same way for twenty years.

Consider Generative Engine Optimization (GEO), the emerging field of influencing how AI systems understand and recommend your products. A traditional workflow, inherited from SEO, might look like this. An analyst identifies gaps in how your brand appears in AI search. A GEO specialist turns those findings into recommendations, then works with a creative team on new content. Brand reviews it, Legal reviews it, and a content manager publishes it. At many firms, this can take a quarter.

Now consider a different workflow. AI continuously analyzes the data and identifies opportunities: a new article, a schema change, transcripts for YouTube videos, or more activity on Reddit. A marketer decides which ones matter. AI then develops several possible campaign directions. The marketer chooses one or two, bringing taste, organizational context, and knowledge of the brand. AI produces the full draft. In some companies, a human writer may still do this, particularly when the brand depends on highly curated and elevated creative work (think Hermès versus a B2B software company … no offense, SaaS readers). An AI review agent then performs a first pass. It can check brand voice, factual claims, prohibited language and serious issues get flagged. If there is no such issue, a marketer performs a final review and gives the green light. AI publishes the content through the CMS.

The human role in this workflow is more editorial. Humans set priorities, choose directions, evaluate quality, and handle exceptions. AI does the execution. Instead of having a brand expert or lawyer personally review every piece of content, they create the rules/guardrails for the system. They test whether the system is working, audit it periodically, and step in when exceptions arise.

3. AI agents as part of the workforce

In this stage, AI agents are persistent, have decision rights, are are evaluated regularly. Instead of waiting for an employee to give it a prompt, a persistent agent can monitor data, identify opportunities, recommend actions, carry out approved actions, measure results, and continue working. A GEO agent, for example, might detect a drop in AI visibility, diagnose likely causes, propose several interventions, test them, and execute low-risk changes within predefined limits.

Even though we hear the term agents everywhere, we are far from this stage. We don’t yet know where agents can go wrong or how to govern them. To reach stage 3, AI systems need persistent goals, access to the right data and tools, some degree of decision authority, and clear guardrails. Just as we define goals and assess employees, we will need to have strong evaluation frameworks for agents. Are their recommendations good? How often do humans override them? What types of errors do they make? Do they stay within their authority? Are they producing better business outcomes? Are they improving over time?

4. AI as an Operating Model/AI-Native Org

Once enough workflows have been redesigned around AI, the organization itself becomes the constraint. If you were building the marketing function today, would you hire the same roles, organize around the same channels, use the same agencies, and keep the same management layers? Let me reuse the GEO example. If AI can analyze the data, recommend actions, draft and review content, and publish it, do you still need separate analytics, SEO, content, brand, and publishing teams touching the same work? Do you still need quarterly campaign cycles or multiple layers of approval? At this point, AI transformation becomes an organization design problem. Roles, team structures, and decision rights were built for a world in which execution was slow and breadth of knowledge/skills was scarce. As those constraints weaken, small teams can own more of an outcome end to end, with fewer handoffs and shorter planning cycles.

Instead of routing every action through approvals, companies will need to define guardrails: what employees and AI systems can do on their own, where the limits are, and when they must escalate to other teams such as Legal and Brand. Performance measures should shift as well. When AI can produce campaigns, reports, and content cheaply, the ability to produce output matters less than decision quality, speed of learning, and business impact.

Key Traits of an AI-Native Marketing Organization

1. Experimentation mindset

The old question was, “What campaign should we run?” Now, everybody and their grandma can generate a campaign. The harder question is, “What hypothesis are we testing?” AI-native teams are learning systems that generate ideas quickly, test them, measure what happened, and feed the result into the next decision.

2. Small teams have more leverage

A team of five may be able to perform work that once required 25 people spread across several functions. These teams need two kinds of people: (i) a strategist who brings judgment, taste, domain knowledge, customer understanding, and organizational context to make decisions. (ii) a builder who knows how to connect AI models, company data, and software tools so the organization can act on those decisions quickly. A GTM engineer is an example of the latter; someone who can connect tools such as Claude Code, Clay, and HubSpot to execute in minutes work that once required several specialists and a series of handoffs.

3. Humans move from production toward judgment and alignment

AI can generate options, analyze data, draft content, and execute many routine tasks. Humans decide what deserves attention, set priorities, understand org context, resolve ambiguity, and take responsibility for the outcome. The strongest marketer may spend less time personally producing each artifact and more time directing a system that can produce many of them.

4. Agents become part of the workforce

AI-native orgs have AI systems with persistent responsibilities. They also need guardrails and clarity on who is accountable for how well they perform.

5. Work becomes closed-loop

If AI makes execution abundant, the advantage will increasingly belong to organizations that can learn what works, and act on that learning, faster. But much of marketing today still has weak feedback loops. A team creates something. It launches it. Someone reviews the results weeks later. Then the team begins another campaign. AI can connect decision, action, measurement, and correction to make closed-loop systems possible.

We are in very early days for this. This last idea, building closed-loop marketing systems, drives much of our work at my new startup, Bodhium Labs. We are currently focused on closed-loop GEO. Over time, we hope to build similar systems across many more parts of marketing.

Subscribe now

Share

Leave a comment

You are neither told which framework to use nor is there a formula to apply. Instead, you gather data, interview people, analyze competitors, build financial models, estimate market sizes, and eventually assemble a slide deck that argues for a course of action. Then your manager tears it apart. You revise. The partner tears it apart again. Over time, you learn not just how to analyze a business problem, but how to think like an experienced C-suite decision-maker.

Now replace McKinsey with Amazon, Goldman Sachs, Blackstone, General Mills, or any Fortune 500 company. The details differ based on whether you are in strategy or operations or marketing or another function, but the pattern is the same. Much of the early career of knowledge workers in business consists of taking an ambiguous problem, applying analytical tools, and producing a recommendation that someone more senior can evaluate, sharpen, and implement.

In new research, my coauthors and I raise a question: what happens when AI becomes remarkably good at the entry-level tasks that most college and MBA graduates perform at the workplace. And what does it mean for the future of work and business education?

We asked AI to do an MBA’s job

Most AI benchmarks tell us whether a model can recall facts, solve math problems, write code, or answer exam questions. Those are useful capabilities, but they are a poor reflection of what most white-collar professionals actually do.

Knowledge work is messy. Managers rarely receive neatly packaged questions with objectively correct answers. They receive incomplete information, conflicting objectives, and uncertain forecasts, and have to juggle organizational politics and pressure to make decisions fast.

In business schools, the case method has become the signature tool for training students for exactly this kind of knowledge work. Rather than teaching theory alone, instructors place students inside business situations (some real and others fictional, but all realistic). Students have to wrestle with the same uncertainty managers face when making decisions: information is incomplete, stakeholders disagree, and every option carries trade-offs. There is usually no single correct answer.

Together with my coauthors, we built BusinessCaseBench, a benchmark based on 238 business school cases spanning 18 disciplines, from finance and accounting to strategy, leadership, operations, marketing, entrepreneurship, ethics, and international business. The benchmark contains 615 open-ended questions, each graded against expert-written solutions using detailed scoring rubrics. Access to these cases is gated and they are unlikely to be in LLM training datasets. In particular, the sample solutions are only accessible to verified instructors and even more unlikely to be in LLM training data

The results surprised us. Today’s leading AI models score above 87% against instructor rubrics. Within one model family (OpenAI), performance improved by roughly 23.3 percentage points in just two years. AI is no longer merely answering factual finance questions or calculating discounted cash flows. It is synthesizing complex situations, weighing alternatives, and producing strong business recommendations across almost every business discipline we examined.

But there is one nuance.

We evaluated responses in two ways. The first awards partial credit based on the fraction of rubric criteria satisfied. The second asked a stricter question: did the model produce a complete answer, satisfying every important criterion identified by the instructor? Here, performance dropped to under 50%. AI often generated strong overall analyses while overlooking an important stakeholder, omitting a key assumption, failing to discuss an implementation challenge, or stopping just short of a truly decision-ready recommendation.

Fig 1. In standard scoring, scores for OpenAI family of models went up by 23.3%

In short, AI is increasingly good at producing exceptional first drafts. It is not yet consistently good at producing gold-standard final drafts i.e. matching instructor solutions along all dimensions (To be clear, I doubt MBAs at top business schools will score much higher, on average, against this stricter criterion). And that has big implications as I explore next.

If AI can do the cases, what is the MBA for?

For decades, much of business education has focused on teaching students to generate sound analyses to support decision-making under uncertainty. Increasingly, AI will do that in seconds. The scarce skill now is verifying whether a draft — completed by someone else — is actually good enough to support an important decision. That is a harder skill. And one that has historically been developed by doing the work oneself over and over. As my colleague Michael Roberts put it in our email exchange, “[it] used to require years of building DCF models to gain sufficient experience so one could pick up someone else’s model and assess it.” If AI systems perform more of the production work, firms may need new apprenticeship models that develop verification and model supervision earlier in career paths.

The other question is where will value shift at the workplace. BusinessCaseBench measures a central part of knowledge work: the production of structured analysis, but it does so under relatively clean, single-turn conditions. In organizations, the same decision problem often comes with differing opinions, organizational politics, and implementation consequences that more experienced managers have to wrestle. If AI produces strong first drafts, I believe value will likely migrate upstream toward the selection of which question is worth answering, and downstream toward building consensus, adapting recommendations to organizational context, and owning the consequences of decisions. Expecting entry-level workers to be able to do this is a tall ask but that may in fact be what they have to do to add value over and above what AI can do.

The educational problem for professors like me shifts from whether students can generate acceptable analysis to whether they can evaluate, stress-test, and improve one created by someone else. This shift does not make foundational analytical skills obsolete. Effective verification requires domain knowledge to identify faulty assumptions, omitted stakeholders, weak causal claims, numerical errors, and incomplete recommendations. The instructional burden for professors in fact increases: students must learn both how to perform the work and how to evaluate AI-generated work. Business schools need to add repeated AI-in-the-loop exercises in which students audit flawed or incomplete analyses, compare competing recommendations, and recognize what a strong answer looks like.

There are some interesting implications for model training as well. Many of today’s most famous benchmarks are approaching saturation. Measuring future progress will increasingly require benchmarks grounded in economically meaningful work rather than narrow factual recall or math reasoning. AI labs have been spending big $$$ buying data on typical work performed by consultants, bankers, marketers, etc through post-training data providers such as Turing, Mercor, and AfterQuery. Clearly, Fig 1 shows this has paid off. But we do find a nice subset of open-ended analytical reasoning problems that remains challenging for frontier models. These harder cases provide a natural target for future model development.

Fig 2. Across disciplines, there remain questions that are hard for frontier models (especially so under complete answer scoring but also under partial scoring)

If you are an experienced exec, let me know how some of the above findings and general trends are impacting hiring at your firm. If you are a recent graduate, share your thoughts on what kinds of questions and concerns this raises for you.

Links to the paper and the project website.

In other news, How do you market to an AI customer is a recent HBR article of mine in which I explore how AI is transforming the customer buying journey and why it would be a mistake for marketers to assume that AI is just another channel through which your customers are finding you. Instead, the new problem for marketers is how to inform and influence AI agents as they do research or make decisions on behalf of your customers. My recent keynote at the just-concluded CommerceNext conference also explored the same theme. More updates on that research in a future newsletter.

Subscribe now

Share

Leave a comment

Citadel responded with a counterargument: Firms don’t replace workers overnight. Regulation and firing costs will slow things down. Further, in many cases, workers become more productive rather than fully displaced. And low costs will drive up demand and fuel the economy.

Both arguments are coherent and can be supported with economic theory. I have previously written about some of the relevant mechanisms, such as this WSJ op-ed on how AI will help bring down costs.

But there is another perspective that I think is missing in much of the debate on AI and the labor market. Managerial decision-making is almost always under constraints. In Management Science and Operations Research, we model firms attempting to maximize an objective. In this case, firms choose their labor capacity and AI investments to maximize profits. AI can both augment and displace labor, and we can model how those choices evolve as the technology improves. But a big part of that modeling focuses on real-world supply and demand constraints that shape what firms can actually do. The Citrini report largely abstracts away from these constraints. An MS/OR perspective brings them back in and I believe that changes the answer.

The goal of a system is not the sum of its parts — Eliyahu Goldratt

Adoption curve: The fact that a model can perform a task does not mean a firm can deploy it inside a real workflow. Work is organized around approval chains, incentives, accountability. Automating or augmenting one or more tasks often means redesigning the system of work. That creates coordination problems and raises trust issues. At my previous startup, we designed an AI-based system for processing incoming screenplays and books for Hollywood agencies and studios. AI can do the work but most of these companies have still not been able to redesign their work around such capability, and big piles of unread scripts remain at these agencies. In short, the constraint is often not intelligence or technological capability. Instead, "Business Transformation is frequently a change maangement problem

Energy: AI compute requires energy. At small scale, this shows up as cost and is fungible through reallocation from labor to AI. At large scale, energy itself is a constraint. If we imagine AI doing 20% of mankind’s work, where is that energy going to come from? We don’t have the ability to ramp up capacity that fast (recall Citrini report makes a 2028 projection). The speed at which we can ramp up energy capacity is a bound on how fast AI substitution can happen, even if the technology is ready. Further, as demand for electricity rises, so does the marginal cost of energy production. This makes AI less attractive relative to labor unless you slow things down a bit.

Reliability: Real-world deployment does not care about average performance but about tail risk. A system that is right 98% of the time may still be unusable if the remaining errors are hard to detect. Small error rates compound across thousands of decisions. Look at Klarna’s reversal of its 2024 strategy of replacing most customer service agents with AI. Now they use a hybrid human-AI system where AI handles simpler issues and humans handle more complex issues. An AI system that works most of the time makes for a good demo but can still not be deployed for a lot of work.

Trust and liability: If an AI executes a bad trade, approves a fraudulent payment, or gives harmful advice, someone needs to own the liability. OpenAI simply says washes its hands off such liabilities. So do all AI vendors. Until responsibility is clear and insurable, firms will limit autonomy and keep humans in the loop even when they are not strictly needed. That slows the transition.

That said, while I disagree with Citrini’s timeline, I don’t doubt that the transition will be disruptive for many. Automation in the UK’s cotton mills was met with the Luddite revolt and the burning of factories. Even though new jobs were eventually created within the same industry (retailing, fashion, and others), many workers suffered during the transition. With AI, new demand and new jobs will also take time to form, and the effects will likely be uneven across sectors. So the tension in the Citrini report is real, even if the timeline is exaggerated.

There is also a deeper issue. Even if firms understand that widespread automation may reduce aggregate demand and hurt their revenues, they do not internalize the full cost of that effect. My former colleague, Gerry Tsoukalas, has a nice paper that frames this as a prisoner’s dilemma: each firm captures its own cost savings, but the demand loss is shared. Automation becomes individually rational even if the collective outcome is worse. You do not need irrational behavior to get a bad equilibrium. Local optimization is enough.

In conclusion, the debate on AI and labor is often: will AI replace workers or not. That is the wrong question. The more useful question is simpler. Which constraints bind, where, and for how long. For practitioners, I think the right operating rule is process alongside AI: start with a narrow workflow that is already measured, where handoffs and approval points are visible, rather than with the most impressive demo. Solve the workflow and acceptance/trust constraints before expecting any results from AI.

Opportunity at an exciting, new startup

I’m working on a new venture. Strong, early traction and a founding team from Deepmind and Wharton. It is an applied AI lab at the intersection of AI and marketing, with a focus on agentic commerce. Much of what I wrote above shows up in practice: what agents can do in real workflows is often less constrained than what firms can deploy, and the interesting challenges sits in that gap. More on what we do later this summer. We’re hiring for two early roles.

First, a Product Manager with 3–7 years of experience at a company with a strong product culture, comfortable operating in ambiguity, and building from scratch with AI. Second, a forward-deployed marketing specialist/engineer; someone with strong communication skills, strategic instincts, technical background, and an AI-native way of working. This could be a consultant with a tech background, someone who has handled CX or accounts at a SaaS company, or client-facing roles at a marketing agency. If we should be chatting with you or someone you know, reach out at kartik@bodhiumlabs.com and CC katie@bodhiumlabs.com (she’ll help coordinate while I’m offline over the next few days).

Subscribe now

Share

Leave a comment

Two months before my search, Walmart had announced a partnership with OpenAI around Agentic Commerce Protocol (ACP). ACP makes it easy for retailers to share their product feed with OpenAI, making them accessible inside ChatGPT; with Stripe powering Instant Checkout, purchases can soon be completed directly in the chat. Two weeks back, Walmart decided to partner with Google for its Universal Commerce Protocol, an end-to-end protocol to allow AI agents to buy and sell.

Walmart’s move surprised me. If I do my product research, discovery, and purchase on Google or OpenAI’s chat interface, they’re giving away the customer relationship and the brand experience. Walmart has spent decades owning the customer relationship through stores, websites, and loyalty programs. Why hand over the interface so early, before consumer behavior has shifted, and especially when competitors like Amazon are clearly holding back?

What looks like a routine retail partnership is in my opinion a sign that commerce is about to change. Not an incremental shift in channels or interfaces. But an “Internet changed retail forever“ type of change. In today’s article, I’ll explore this shift.

What Is Agentic Commerce?

Think of agentic commerce as AI that can autonomously shop and complete purchases, not just help you find products. There are two modes of AI support in shopping today.

First, the AI assistant. You explain your context. “I am training for a marathon. I had ankle surgery. I run on trails. Please suggest shoes.” The LLM searches the web, reads various pages and suggests options and links. You still evaluate, decide, and complete the transaction. This is how most of us use ChatGPT/Gemini today for shopping research.

Second, the shopping agent. You say “Order a good Hoka, Brooks, or New Balance running shoe with strong ankle support for running a half-marathon. Shoe size 10. Price under $100.” The agent researches products, filters options, and picks one. In the “Human Present” mode, you approve the purchase and the agent uses your credit card to make the purchase. In the fully autonomous mode, the system completes the purchase without even asking for approval.

Three technology layers enable AI agents to shop autonomously:

1. Agents Need to Talk to Each Other (The Protocol Layer)

AI agents must communicate across platforms. If there is no common standard, every merchant will have their own approach to sharing product information and shopping agents will need to integrate with thousands of them. That fragmentation will make agentic shopping impractical. To avoid this, Google and OpenAI are developing UCP and ACP respectively.

2. Agents Need Places to Actually Shop (The Commerce Layer)

Retailers must open “machine-accessible” back doors so agents can query and transact. Shopify, Etsy, and Salesforce are already enabling this for merchants. Some integrations are already live:

ChatGPT + Instant Checkout: You can now buy certain products directly in a ChatGPT conversation. Works with major platforms like Shopify, Stripe, and Etsy.

Perplexity: Partnered with Firmly.ai to enable in-chat purchasing across 150+ retailers since launch in late 2024, with shopping queries up 5x.

Google Gemini: Rolling out features with the UCP announcement.

3. Agents need to trust each other (The Governance and Payments Layer)

As AI agents begin to transact, the existing payment system loses a line of accountability. Issuers need to know whether a user actually authorized a purchase to an agent, merchants need to understand liability when an agent misbehaves, and networks need a way to distinguish compliant agents from bots. While Google’s Agent Payments Protocol (AP2) references “mandates,” there is still no industrial-grade system for issuing, verifying, and enforcing user authorization across autonomous agents.

Why Agentic Commerce Changes Everything

Most retailers I chat with are falling into three critical misconceptions that could undermine their ability to respond.

Misconception 1: “There will always be a human at the other end.”

I am not suggesting we will outsource all our shopping to AI agents but some of it offers no joy and involves a fair amount of research. For example, I would be happy to outsource my travel booking to a competent AI agent.

With agentic shopping, users can set spending rules. “Reorder my usual groceries if the total is under $200.” “Buy Coldplay concert tickets under $150.” (OK, that one is infeasible), “buy a gift for a birthday.” Sometimes, we may choose a “human present” mode in which we make the final approval. But even then, the agent – and not the human – is choosing. And when that happens, the agent never sees the website experience.

Misconception 2: “We just need AI crawlers to read our content.”

AI Engine Optimization (AEO) assumes visibility is enough: make content structured and crawlable. But AI agents are not just a filtering tool; they can decide.

Over decades, marketers have mastered human persuasion. We know that $19.99 converts better than $20. We understand social proof, scarcity, authority. These techniques work because we understand human psychology. But artificial neural networks (ANNs) have different biases and framing effects than biological neural networks (BNNs). Companies will have to decode “AI persuasion” to succeed.

Misconception 3: “AI Is Another Channel”

This is the mistake many studios made with Netflix’s entry: mistaking a system-level shift for another distribution channel. Soon, the entire content market changed and most studios are now at the mercy of Netflix and Amazon, which control what gets made, by whom, and who gets to see it.

Similarly, a new platform (AI models and agents) will likely control big chunks of commerce. If retailers are not careful with their next moves, AI will control what people buy and from whom. Retailers risk becoming back-end fulfillment centers to AI models that shape demand and control customer relationships.

What Should Companies Do?

Agentic commerce opens up many tactical questions and two big strategic ones for me. At a tactical level, companies face familiar questions:

Infrastructure: Is your product data structured, and machine-readable?

Partnerships: Which AI ecosystems should you integrate with: ChatGPT, Google, Perplexity, or all?

Measurement: If visits vanish, what replaces website analytics as the measure of success?

Branding: How do you build preference when AI makes the choice?

Customer Relationships: When AI owns the interaction, what does “customer loyalty” look like for retailers?

Strategically, I see two big questions.

First, What is AI, if it’s not a channel? My view is that for retailers, AI is best understood as a new class of customer. You have spent decades figuring out how to influence humans (BNNs). The same approaches may no longer work. ANNs are like a new species and you therefore need a new science on how this species makes (shopping) decisions. You need to get to work to figure out this science.

Second, how can you play in agentic commerce and still control branding and customer relationships. Amazon’s strategy seems to be that they need to own the I agent that consumers will use for shopping. This is their strategy with their shopping agent, Rufus. It’s another story that Rufus is mostly underwhelming. Walmart is hedging its bets. On the one hand, Sparky is their response to Rufus. They’d love for everyone to use Rufus. But they are also (rightly) worried that shopping gateways might be ChatGPT or Gemini so they are also providing their catalogs to them. Should other retailers follow suit? Should they resist like Amazon and try to build their own shopping agents? That’s the big strategic question.

How are you thinking about it?

Subscribe now

Share

Leave a comment

Looking back, this past year felt full and yet marked by an ease that stands in contrast to the many years before that felt equally full but not as light.

Our family had ended 2024 with a last-minute trip to the Galápagos. My wife once again proved that even improbable plans can come together if you simply decide that they will.

This year was no different. We closed 2025 with another last-minute decision, this time to Australia. Snorkeling in the Great Barrier Reef and seeing koalas and kangaroos in the wild was a perfect way to mark the year-end for a lifelong wildlife lover. Earlier in the year, we had a spring visit to Santa Fe and a summer trip to France, where the highlight was an entire day cycling through the hilly villages of Provence culminating in an excellent French winery. There is something satisfying about exploring a new place on feet or one pedal stroke at a time.

Early in 2025, I also ran a 10K and then a half marathon. On paper, these are just races. In reality, they marked, at least for me, the closing of a long chapter that began with a misdiagnosed ankle fracture from a volleyball game, followed by surgery, and nearly a year of rehab. I wasn’t much of a runner before the injury and had never even run a 5K. Running these races was mostly about telling myself that this chapter was now complete and that I no longer needed to allocate mental energy to the injury or recovery.

After years of skipping live music post-Covid, we also returned to concerts. Coldplay in 2025. Taylor Swift the year before. These were my first concerts with my kids, an entirely different experience introducing them to the power of live music.

On Endings and New Beginnings

Professionally, the year carried a significant closing. We completed the sale of my previous startup, Jumpcut Media, an AI workflow platform used by creative executives at nearly 400 companies across Hollywood. I’ve written before about how difficult those years were, and about the internal battles that accompanied them. When the sale finally wrapped, I was convinced of one thing: this would be my last startup as a founder.

At the very least, I told myself, I would take a two-year break to slow down and be more present. Invest in health, mindfulness, and the parts of life that tend to get neglected in the fast lane.

And then, sometime in the summer of 2025, I started another company!

Not because I was restless. And not because I felt compelled to prove anything. I simply couldn’t see the wisdom in forcing distance to honor an old promise especially when the opportunity felt immediate, energizing, and aligned, and when I had found cofounders I genuinely looked forward to working with.

The new company is a research lab focused on building the future of AI-enabled commerce. At a high level, we’re working to make it easier for companies to market to and transact with AI, from research-backed approaches to improving visibility in AI search to preparing for a world of agentic commerce where AI agents shop on behalf of customers. I’ll write more about this soon.

For now, what matters most to me is something simpler. I’m satisfied that I can return to the kind of entrepreneurial work that once stressed me. That I can approach it with curiosity instead of fear. Ultimately, that is the growth we all seek.

Being aligned with my cofounders on our approach to entrepreneurship was very important to me. One of my cofounders, with whom I first began brainstorming this idea, comes from a deep tradition of Indian Vedanta practice. There’s a line from that tradition that emphasizes the ability to meet challenging work without agitation and acting fully while remaining inwardly free:

“Established in yoga, perform actions, abandoning attachment, remaining even-minded in success and failure. Evenness of mind is yoga.”
— from the Bhagavad Gita

That increasingly feels like the real goal: facing challenges without making the work your identity; staying engaged without becoming entangled. If I had to name one goal for the new year, it would be deepen this practice.

Looking Ahead

I hope you all had a good year. And for those of you who faced significant challenges in 2025, I wish you a gentler and better year ahead. I leave you with the Pali word anicca (pronounced anicha) —the concept that all things, from physical bodies to mental states, are impermanent, and that understanding this impermanence is the key to liberation.

My best wishes for a wonderful 2026 to all of you.

Subscribe now

Share

Leave a comment

The Productivity Paradox: How AI is De-skilling You

A few weeks ago, I was driving home with my wife. We were an exit away when she casually ended the GPS directions. I immediately turned them back on. It startled me how disoriented I felt. This was our own neighborhood and I should know the way. In fact, I did know the way. But I’ve grown so accustomed to GPS guiding every turn that its absence made me uncomfortable.

It wasn’t always like this. Growing up in Bangalore, my father built a mental map of the city the old-fashioned way: through curiosity, attention, and plenty of wrong turns. I never had to develop that skill.

In London, aspiring “black cab” drivers spend years mastering The Knowledge of London: thousands of streets and landmarks, memorized through intense study. Today, an Uber driver can navigate the same city without knowing a single turn.

We celebrate this convenience. But as we use AI to drive efficiency in more parts of life, it makes me wonder: What parts of our minds are going quiet without us noticing?

AI and Skill Fade

AI makes task completion easier and faster, but it creates a paradox of help and harm: the more the tool does for us, the less capable we become without it. This de-skilling process isn’t new; tools and automation have long eroded physical skills. But with modern AI, are we outsourcing cognitive judgment itself?

• The Early Warning: Autopilot systems made flying safer but also reduced pilots’ manual practice. The tragic crash of Air France 447 in 2009, triggered in part by the pilots’ inability to manually fly the plane after automation disengaged, remains a stark reminder of how quickly manual proficiency and situational awareness can erode through overreliance on automation.

  
  A 2019 study found that long-haul pilots are particularly vulnerable to skill erosion due to prolonged exposure to highly automated flights. The degradation goes beyond “stick and rudder” skills; it affects situational awareness, cognitive agility, and workload management. Their recommendation? More scheduled manual flying and unpredictable simulator scenarios. The FAA now recommends that pilots do more manual flying to maintain core skills, a recognition that automation, left unchecked, hollows out the very expertise it was designed to support.

• The Current Reality: A recent study found that Endoscopists who regularly used AI for polyp detection became worse at finding polyps when the AI was turned off, with adenoma detection rates dropping from 28% to 22%. Continuous reliance on AI subtly altered clinician behavior in ways that reduced performance when the technology wasn’t available. My colleagues found a similar pattern in education: students who practiced math with unrestricted access to GPT-4 initially performed better, but once access was removed, they underperformed compared with peers who had never used AI. Generative AI without guardrails can hinder skill development when learners begin treating it as a cognitive crutch.

For tech teams, everyday tools like coding assistants speed up development but consistently offloading routine tasks can make developers less capable of starting from scratch.

Subscribe now

Building Mindful Friction Into AI Workflows

AI-enabled workflows are designed to remove friction. But we underestimate how valuable friction is. Struggling through a task and doing the reps is what helps us develop skills.

Be Mindful of AI Use

Not every skill deserves preservation. I don’t need to chop wood to heat my home but some capabilities are too fundamental to lose. Are we outsourcing foundational capabilities like ideation, critical thinking, and the ability to organize our thoughts while struggling with a messy first draft?

Being more mindful about our AI use is a first step. When I ask AI to design something for me, I am not losing a core skill. I never had that skill in the first place. AI simply reduces my dependency on others but does not erode a skill for me. But when I use AI to write the first draft of a paper or figure out how best to communicate an idea, I’m giving up something central to my work as a professor and entrepreneur.

Require Human First Rounds

Before handing a task to AI, take an initial pass: a sketch, a list of assumptions, a hypothesis, a rough direction. This simple constraint ensures the model amplifies a human starting point rather than defining it outright. When a designer produces three rough concepts before asking AI to refine them or generate a few more concepts, the final output reflects human imagination expanded by machine capability and not machine output lightly edited by human taste.

In recent research, my coauthor and I studied how different human–AI collaboration models shape the diversity of stories writers produce. When writers ceded early creative control to AI, outputs became strikingly similar. When humans led the early stages—ideation, outlining—before bringing in AI for drafting, the resulting stories preserved far more diversity, in fact close to fully human work.

Build in Deliberate Review Points

As MIT’s Renée Richardson Gosline notes, “Sometimes, slowing people down is the best way to serve them.” Short, intentional pauses before accepting an AI output or moving to the next stage force us to re-engage judgment rather than sleepwalk through automation.

Organizational Solutions: Processes That Build Resilience

While individuals have agency in retaining their skills, organizations must implement structural safeguards to counteract systemic over-reliance on AI. Two process-level tools help: Reliance Drills and Analog Practice.

Reliance Drills

These are routine stress tests that simulate what happens when AI guidance is wrong, missing, or unavailable. In practice, this might mean running failure scenarios where teams must reach decisions without AI support, or instituting periodic “off-AI days” that expose where knowledge has quietly eroded. The goal isn’t to eliminate AI but to prevent passive human oversight and keep expertise alive. When a team discovers during a drill that they can no longer evaluate a financial model without the model’s assistance, they’ve surfaced a critical vulnerability before it turns into a crisis.

Analog Practice

Just as pilots are required to hand-fly periodically to maintain touch, analog practice ensures your team can still think independently, create from zero, and reason across ambiguity. These aren’t nostalgic exercises but are insurance policies.

Every generation gains tools that free it from old forms of effort. But every generation also risks forgetting something essential in the process. AI will take on more of our tasks, but it shouldn’t take over our thinking. The work ahead is to remain awake. To protect the skills that shape our ideas and anchor our judgment.

Recommended Readings on the Slow Life

Two recent posts I enjoyed (and which nominally connect with the theme of this post):

Vizi Andrei’s essay in praise of the slow life and why meaning and purpose are to be found there.

A post by PG Bhat, who writes about living lightly and purposefully, sustained by teaching and simple self-reliant routines (with plenty of friction). (Mr Bhat is my uncle and a reader of this Substack)

Leave a comment

Share

Subscribe now

“For a conscious creature, there is something it is like to be that creature,”

— Thomas Nagel, What Is It Like to Be a Bat? (1974)

One of my students once asked me: “Will AI ever be conscious?”

It’s a profound question at the intersection of philosophy, neuroscience, and technology. In this episode of Creative Intelligence, I speak with Anil Seth, professor of cognitive and computational neuroscience at the University of Sussex and author of Being You, about the nature of consciousness, whether it’s something machines could ever truly have, and and what it would mean if machines ever crossed the line into sentience.

There’s No Checklist for Consciousness — And That’s a Problem

I asked Anil for a checklist to determine if an entity (an AI system) is conscious. Anil was very clear: even today, we don’t have a scientifically agreed-upon test to determine whether something (someone?) is conscious. This creates major ethical and philosophical uncertainty as AI systems become more human-like in their behavior.

In fact, not only do we not have a checklist, we also don’t know why and how physical processes in the brain give rise to subjective experience. Philosopher David Chalmers has referred to this challenge as the “hard problem of consciousness” (in contrast to “easy” problems such as explaining behavior, memory, etc).

“It would be lovely to have a checklist... but we don’t. That’s one of the most critical challenges in consciousness research.”

🧬 The Brain Is Not Just a “Meat-Based Computer”

One view that I have heard before is that consciousness is a type of information processing architecture involving self-modeling, narrative generation, predictive modeling, etc. Joscha Bach, among others, has advocated for this view that “we are all software.” I have always found it hard to accept this view that consciousness can emerge from similar information processing systems. One striking moment in our conversation was when Anil challenged the viewpoint that consciousness is just advanced information processing. Just because a computer can simulate the brain really well doesn’t mean you can instantiate consciousness. We use computers to simulate all kinds of things without expecting the computer to give rise to the phenomenon they're simulating. The problem is that when it comes to AI we sometimes forget that it’s a simulation.

“You know a simulation of a weather system doesn't get wet. A simulation of a black hole doesn't destroy the universe. For brains, because we think they are not only usefully simulated by computers, but actually are computational at their most important level, we slip into this temptation and, and mistake the map for the territory.”

This analogy cuts deep. Even the most powerful AI systems today simulate intelligence, understanding, even empathy. But that doesn’t mean they feel anything. Anil isn’t asking us to believe consciousness is magic either. According to Anil, consciousness might depend on biological processes like metabolism and that means simulating a brain isn't the same as being one.

🪞 Conscious Illusions Are Dangerous

Even if AI systems aren’t conscious, they often look like they are. Anil warns that this illusion could be one of the more psychologically and socially disruptive aspects of AI.

“It’s dangerous to build systems that give the illusion of being conscious — it can be a pretty dangerous illusion.”

When we’re emotionally vulnerable, we’re more likely to attribute feelings and intent to machines. That creates a real risk of misplaced trust, ethical confusion, and societal distraction especially as these tools become more personalized and persuasive.

My Reflection

One of my main realizations from this episode is despite millennia of interest in consciousness, we are still in the early innings of understanding consciousness itself. Let alone building it (which is a source of much relief to all of us).

As AI systems grow more powerful, we’ll be tempted to ask, “Is this system conscious?” For now, I feel it may be the wrong question (for the sake of completeness, many smart people believe otherwise and it’s possible I am wrong and they are right). Perhaps a better question from my perspective is: Does it give the illusion of consciousness? And if so, what should we do about that?

We humans are wired to anthropomorphize — to project intention, emotion, and sentience onto things that behave like us. That tendency is now being engineered into the software we use. Many readers have shared recent news stories about people who treat ChatGPT as a confidante, friend, or even a romantic partner. There are also darker accounts: troubling relationships with chatbots, and in some heartbreaking cases, suicides following interactions with a next-token predicting large language model. That is exactly what can go wrong with an illusion of consciousness. We should be aware of it and, in my opinion, be wary of it.

Timestamps

• 0:00 – Introduction

• 2:02 – What Is Consciousness?

• 5:00 – Is Reality a Construction of the Mind?

• 7:16 – Can AI Be Conscious? A Checklist for Consciousness

• 12:07 – The Hard Problem of Consciousness

• 17:16 – Are We All “Meat-based Computers”

• 25:27 – Building vs. Emerging Consciousness: Biology, Brains, and AI

• 35:56 – Identifying Consciousness: Beyond Intelligence and Behavior

• 40:35 – Anil Seth’s Research & Ethical Reflections

• 45:53 – Closing Remarks

This will be the penultimate episode of this season. As I gear up for the start of my teaching semester, I hope to bring to you one final episode of Season 1.

If you have thoughts, reactions, or disagreements, I’d love to hear them. You can reply to this post here or on LinkedIn or Twitter.

Share

Subscribe now

Leave a comment

Sound familiar? This concern is older than you imagine. Back in 1987, economist Robert Solow famously observed, “You can see the computer age everywhere but in the productivity statistics.” It kicked off years of debate about the so-called IT productivity paradox. A similar debate is emerging about Gen AI.

In this episode of Creative Intelligence, I bring to you a live conversation from the 2025 Wharton Global Forum that was held in San Francisco. I sat down with Aparna Chennapragada (CPO, Microsoft), Sajjad Jaffar (Partner at GrowthCurve Capital), Xing Liu (HongShan Capital Group/formerly Sequoia China), Ramon Richards (CTO, T. Rowe Price), and Lin Yuan (CAIO, Newfront) to explore many questions that executives are now quietly asking, including the big one: Where is the ROI from Gen AI?

The Productivity Payoff Takes Time—and Transformation

The IT productivity paradox was a big research question for technology researchers in the 80s and 90s. By the late 90s, there was ample evidence of the productivity impact of IT and those findings are important today: the gains from IT investments came with a lag. The lag was partly because investments in IT alone weren't enough. One needed complementary resources like skilled labor and workflow redesign. In short, firms needed time to adapt, restructure, and learn. (there were other factors like measurement issues which is not my focus here) Fast forward to Gen AI in 2025. We are seeing interest and adoption at a rapid pace but ROI and productivity are still elusive.

If you're a CEO or CIO wondering where the payoff is—look beyond the tools alone. Invest in reskilling, rethink processes and workflows, and give your teams time to adapt.

Product Strategy and Moats in the AI Age

In a world where every competitor has access to the same foundation models, what’s defensible? Several of our guests converged on the same answer: data.

• Lin emphasized the value of proprietary data in generating moats and also argued that vertical-specific models and domain-tuned data are a powerful edge, not just in quality but also in cost-efficiency.

• Sajjad described data as a “latent off-balance-sheet asset” that is underutilized and underappreciated, but should be at the center of every successful AI strategy.

• Aparna and I chatted about the broader product strategy. She warned: “Every few months, the model eats the product.” The best product strategy in her opinion is a “Thin product, thick model.” and building continuous feedback loop between the two. This is a big theme in my opinion, one that I hope to explore more deeply in a later episode.

• We also discussed how can technology and product teams adapt to constant changes and model releases. Ramon emphasized the need for a modern architecture, sound data infra, and automated delivery practices in the tech org as important preconditions to react to this rapidly evolving space.

One-Person Unicorns—and the End of the Traditional Org Chart

Sam Altman predicted that “one-person unicorns” will emerge from agentic AI; tiny teams powered by massive intelligence. That future is already starting to show up in leaner early-stage AI startups.

“We’ll have to rethink everything we teach in business school about org behavior, leadership, and people management.”
— Xing Liu, on the implications of agentic AI

If AI reshapes the firm, then business education must adapt with it. Leadership might become less about managing massive teams. Management theory may have to grapple with humans managing human and non-human teams.

My Reflection

We often overestimate short-term impact and underestimate long-term transformation.

Just as IT reshaped firms but the impact was seen in decades, Gen AI is unlikely to move the productivity needle until other pieces such as redesigned workflows and upskilled talent are in place.

Too bad, there is no plug-and-play ROI.

Preview of the next episode: I just recorded an episode with Prof Anil Seth who studies the biological basis for consciousness. He and I discussed whether AI can be conscious, the defining traits of sentient machines, and the ethical implications. That was a fun conversation and I’m excited to bring it to you in two weeks.

Timestamps

0:00 – Introduction
3:05 – Beyond AGI: Intelligence Overhang and the Infrastructure Gap
9:32 – ROI, Pattern Matching, and What Makes AI Valuable Today
15:45 – From Products to Systems: Rethinking Software in a Model-First World
23:28 – Data as Differentiator: Memory, Moats, and Domain-Specific AI
30:52 – Organizational Shifts: Data Culture, Change Management, and the New Stack 36:18 – One-Person Unicorns, Global Teams, and Agent Collaboration
39:50 – Competing in the Age of Giants: How Startups Can Win
43:04 – Regional Dynamics: Global Networks and the Open-Source Moment
46:05 – The Talent Revolution: Skills That Will Define the Next Decade
50:46 – Interoperability and the Cost of Lock-In
53:20 – Value, Risk, and Society: Who Gains, Who Loses?
57:48 – Closing Remarks & Conclusion

As always, I’d love to hear your thoughts. Share them below.

Since joining Grammarly after its acquisition of his previous company Coda, Shishir has been reshaping the company's vision—and, more broadly, rethinking the future of productivity tools, agents, and work itself.

This conversation was full of insightful analogies and practical frameworks for business leaders. Below, I share some themes that stuck with me but let me first advocate you really need to listen to the whole episode. Shishir is someone whose clear thinking I have always admired and in this episode he covered a lot of ground: the future of workplace productivity, how to transform companies around AI, setting bold AI visions, and how a leader should think about company culture.

Every Worker Will Start Their Job as a Manager

Imagine you show up to your first job and already have a team of 50 direct reports. They’re not people but AI agents. They work 24/7, don’t argue, and know everything in your knowledge base. Shishir thinks it’s where we’re headed.

“Imagine you showed for your first job and you had a team of 50 that worked for you … and your job is to lead this group. The skill that humans are going to need more of is management … Every employee starts off managing a team of agents. That’s the new baseline.”
(Timestamp: ~18:00)

I like this thought because it reframes the question from “Will AI replace me?” to “How do I orchestrate and lead this new team I’ve just inherited?”

Grammarly Is a Superhighway, Not a Grammar App

Grammarly is misunderstood by many including me. What people see is grammar suggestions. Shishir argues that what’s actually under the hood is a distributed AI agent framework that operates invisibly across thousands of apps, from Google Docs to Apple Notes to Slack.

“Think of Grammarly as the AI superhighway. So far, only one car has been on it—your high school English teacher. But now we’re opening up the road to all kinds of agents.”
(Timestamp: ~4:00–12:00)

This shift in framing offers a glimpse of what AI-native software might actually look like. But it also offers a framework for how to reinvent companies around AI (Timestamp ~ 34:00). We covered several enablers of this business transformation ranging from how to manage employee mindset to platform choices and company culture.

Culture Is a Product Too—And Rituals Are How You Ship It

Shishir has a forthcoming book on team rituals and we spent some time talking about it. He explains an idea he first heard from Dharmesh Shah of Hubspot: a company doesn’t just build one product. It builds two: one for customers, and one for employees. And culture is that second product. Shishir goes on to say that company rituals are how you put them in practice.

“If values are what you believe, rituals are how you behave.”
(Timestamp: ~38:00-46:00)

He shares a standout example from Coda and Grammarly: a ritual called Dory (ranked Q&A so everyone can equally influence the agenda of team meetings) and Pulse (removing groupthink so you get everyone’s inputs. It’s worth asking what rituals you are using in your company and why.

My Reflections

The conversation shows how a company that could have been “disrupted” by AI chose instead to reimagine itself around it—not just defensively, but with a bold, forward-looking vision. That’s something many leaders can learn from. Overall, there is a lot to take away and implement here.

Finally, a preview of the next episode: I had a great conversation with several AI leaders at the recently concluded Wharton Global Forum. My next episode will be based on that live conversation and will explore a range of themes from AI-centric product strategies to measuring RoI from AI investments.

Timestamps

• 0:00 Introduction – Why ChatGPT Didn’t Kill Grammarly

• 7:00 The AI Superhighway: Why Grammar Was Just the First Car

• 13:30 Agents and Productivity;

• 18:00 Why Everyone Will Be Managers (of Agents)

• 22:00 Which AI tools to use today?

• 32:00 Reimagining Education

• 34:30 Playbook For Business Transformation Around AI

• 38: 30 Building culture with rituals

• 46:00 Rapid-Fire Productivity Confessions

• 50:00 Final Reflections

Share

Leave a comment

Subscribe now

In a world where consumers increasingly consult LLMs before making purchase decisions and, soon, where AI agents will make choices on their behalf, you need to rethink who you need to influence, and how.

Having a Gen AI strategy that doesn’t account for this B2AI wave feels like you are sharpening your pencils during a printing press revolution.

1. AI as a new front for marketing persuasion

Digital platforms popularized terms like B2B2C—where companies sell to customers through powerful platforms that curate and matchmake. Instacart, for example, partners with grocery retailers and CPG brands, but owns the user experience, data, and customer relationships. Search engines similarly intermediate between brands and shoppers. Now many of us are outsourcing our entire shopping research to AI for everything from consumer use cases like which wallet to buy as a father’s day gift to enterprise use cases such as which consent management platform to license for your business. Tools like ChatGPT are becoming trusted concierges: instead of giving you search results, they search, synthesize, and recommend an option.

And it may not stop there. Today, buyers consult ChatGPT and make the final decisions. Tomorrow, their personalized agents might ask an enterprise’s AI agent for product information, negotiate pricing, compare personalized bundles in real time, and make purchase decisions. The better those agents understand the user and the world, the more they will dominate as the interface between supply and demand. (Yes, this means loss of agency, but the march to that future began long ago with newsfeed personalization and recommendation systems. I discuss this in detail in my book).

2. AI’s impact by the numbers

Adobe reports that 39 percent of consumers in the US have used generative AI for online shopping and Forrester reports that 89% of B2B buyers now use Gen AI in the purchase process for everything from vendor discovery to evaluation. To be clear, the actual volume of traffic coming from LLMs is still a tiny fraction of web traffic for most businesses but the numbers are growing fast. For example, the same Adobe report indicates that “traffic from generative AI sources increased by 1,200 percent” in the 7 months from Jul ‘24 to Feb ‘25.

LLM usage patterns are different from search. Only 30% of LLM questions match traditional search while “the other 70% consisted of unique queries rarely seen in standard search engines” [1] Further, top cited domains are different for different LLMs. Google favors big domains but ChatGPT is comfortable recommending smaller brands and domains [1].

AI Overviews (AIO) in search results is also reshaping what it means to “rank” in search results. Bain reports that 60% of all searches now end without a user clicking through to a website, because the answer is consumed directly from the AIO. AIOs are showing up more often in search results. In Healthcare and education, 90% of search results pages have AI overviews [2]. B2B tech industry saw increase in AIO from 36% to 70%. AIO is rarer for ecommerce today, which likely helps protect Google’s monetization. When Google’s AI Overviews (AIO) is present, click-through rates on the top organic link drop by 34.5%, according to Ahrefs. For paid search as well, CTR drops by 15-20% when AIO is present [3]. For one B2B SaaS startup I invested in, AIO shows up in 74% of their top keyword clusters and this has lowered CTR in both SEO (-37%) and paid search (-9%).

Competing for AI’s “attention,” whether through tools like ChatGPT or AI Overviews in search, is the new marketing battleground. This is variously referred to as Generative Engine Optimization (GEO) or AI Engine Optimization (AEO), where the goal is to be cited or recommended by AI systems. To succeed, brands must understand how they are represented in AI training data, how clearly their value propositions are expressed in structured formats, how they are recommended by search tools used by LLMs, and how their products are interpreted in LLM inference steps.

3. A new problem statement

As AI agents take over product research, evaluation, and decisioning layers of the customer purchase journey, traditional assumptions about influence, messaging, and brand visibility no longer apply. We must ask:

How do we rewire marketing and CX for a future where our primary audience is equally human and AI?

How do we ensure AI systems are reliable and accurate ambassadors for our brand?

How do we prepare for an AI-to-AI world in which customers’ AI agents will want to talk to our enterprise AI agents about our products and services?

These are not purely technical challenges, nor are they purely creative ones. They touch everything from how we describe our products and structure our data to how we anticipate customers will use AI in the future. And while we don’t yet have a complete playbook for answering these questions, some elements of a solution are emerging.

4. Elements of a playbook

Your presence in AI results is shaped by many factors: (i) Whether LLM crawlers have discovered your content (ii) how your brand is represented in the LLM training dataset (iii) Which search tools are used by the LLMs and how you show up in the results (iv) The actual content of your pages and how it impacts LLM inference.

• Discovery by LLM crawlers: Many websites unintentionally block AI crawlers through their robots.txt or CDN settings. For instance, Cloudflare recently began blocking LLM crawlers by default. That may make sense if you're a publisher monetizing your content (like a news outlet), but for most businesses aiming to be discoverable, it’s worth reviewing those settings. To explicitly make your site accessible to LLMs, implement a llms.txt file; it is a new standard that helps LLMs find and understand your content. There are two main formats: (i) llms.txt: A lightweight index of your pages, with titles and descriptions. (ii) llms-full.txt: A more comprehensive version that includes full page content. You can try out a generator like Firecrawl’s tool to get started.

  
  Beyond access, focus on machine readability. It's not enough for AI crawlers to read your content; they need to understand it. That’s where structured data (or schema markup) comes in. Structured data adds a machine-readable layer to your site, by adding tags to clarify whether a piece of content is a product, a review, a FAQ, or something else entirely. While structured data has long played a role in SEO, it’s just as important in the Gen AI era.

• Training Data: Just like humans have emotional reactions to brand names, LLMs are predisposed to view brands one way or another based on how those brands up in AI’s training data. This is not merely about your first-party data being crawled by LLMs but also includes your brand shows up on Reddit, Youtube, Substack, and more. OpenAI, for example, has started licensing data from Reddit. Similarly, many LLMs have started licensing data from news outlets, and how you show up in news media matters as well. While search engines will show a list of results to a user and therefore present a range of perspectives on your brand, a Neural Network’s weights reflect the net impact of all the training data. As a result, it is less likely to present a range of perspectives but a single viewpoint informed by all of its sources.

• Search engines: For many queries, tools like ChatGPT use real-time search capabilities to retrieve the latest information and then synthesize a response. The good news? Your SEO efforts still count. If your site ranks well in traditional search engines, it’s more likely to surface when an LLM uses search tools. But the catch is that different models use different search backends. Gemini uses Google, Anthropic’s Claude uses Brave Search, and ChatGPT defaults to Bing (though there’s growing speculation that it may be incorporating Google as well). That means you can’t optimize solely for Google anymore. In the short term, improving your search rankings is the fastest path to visibility in LLM-powered experiences. That’s because LLM training data are only updated every 6–9 months, whereas search-based responses reflect real-time content. What’s more important — being cited or just showing up in the results (yes, I just used an em-dash there and no, it’s not AI)? Citation is harder to achieve and may not reliably drive clicks, since many users don't even view the citation details. So for now, focus on strong visibility in the results.

• Website Content: Just as certain messages are more emotionally persuasive to human brains, some words on your website are likely to activate certain parts of the LLM’s Neural network and likely to push your website higher in the AI results. Unfortunately (or should I say fortunately), there is no golden rule here on which types of phrases help and it requires significant experimentation. There are a couple early “best practices” I have heard but no formal academic study has proven these two strategies: (i) Prioritize your first 100 words. LLMs like ChatGPT often quote or cite the first meaningful paragraph they find. Placing key insights or summary content near the top of your page might help. (ii) Experiment with FAQ formats. LLMs are trained heavily on question–answer pairs, so that format may align more naturally. Overall, this is a very promising AEO approach but is also the one that is least understood and can benefit from some formal A/B tests.

• Agents: This part is science fiction as of today but it is not set in a distant future. Imagine a future where a buyer’s AI agent knows their budget, health goals, preferences, and schedule, and finds the right travel package without the human ever intervening. Or picture a business AI agent that autonomously researches, selects, and purchases a SaaS or AI tool on behalf of a team. In this world, AI agents for buyers and sellers trade information in real time not unlike how real-time ad buying works today. In this world, it won’t be enough to just make your content machine-readable. Your business will need its own AI agent to interpret, synthesize, and respond to questions from a buyer’s AI agent using your entire body of content. That will be the start of AI-to-AI (i.e. agentic) commerce.

A recent study showed that that the biggest driver of your rank/visibility in LLM answers when search tools are used by the LLM is your rank in the search results. Next, it is how you show up in the training data. A brand that is well represented in the training data may be AI’s top recommendation even if it is consistently ranked lower in search results. The third — and still quite important — factor is your website content/wording and how that activates the Neural network (I am starting to get nervous about my use of em-dash … what a sad day for writers).

5. Next steps

How to kickstart your GEO journey:

• Start tracking AI-driven traffic to your site: Even if the numbers are low today, they’re likely undercounting. Why? Because many AI tools don’t link directly to your site. Instead, users see your brand in an AI-generated answer, then type it into their browser. Attribution is fuzzy but the trend line is worth tracking regardless.

• Use a basic measurement platform: Start tracking how visible your brand (and your competitors) are in AI-generated answers across key topic areas. While no one knows exactly what people are searching for on LLMs, you can use proxies such as which topics get more search volume on Google. I’m not endorsing a specific tool here, but feel free to DM or email me for suggestions.

• Make sure LLM crawlers can access your site: Check your robots.txt and CDN settings to ensure you’re not accidentally blocking AI crawlers. Also, think beyond Google. For example, verify your site is being indexed by Bing. Here is an example of a site that has unknowingly blocked Bing.

  

• Some websites are repeatedly cited by LLMs. If those sites link to you, you're more likely to appear in AI-generated answers. GEO measurement platforms can help you find your top citation domains. These are the partners you should prioritize for affiliate marketing.

How do you measure success?

Your immediate goal over the next quarter should be education. Try, test, and learn. For the next couple quarters, focus on:

• Visibility: Is your brand showing up in LLM answers and how high do you show up for high-volume topic areas

• Content citation rates: How often is your website being cited (helpful but indirect measure of success)

Performance marketers will not love the above two metrics because they don’t reflect impact on your website traffic or revenues. But I believe that attribution will improve soon. Today, most LLMs don’t show clickable links in their responses. But that’s likely to change especially as these tools move toward monetization and need to track outbound traffic. When that shift happens, brands that have been investing early will be far ahead.

Finall

REFERENCES/REPORTS

1. SEMRush report

2. Google AI Overview and search clicks

3. Google AI Overview and Paid Search

Subscribe now

Share

Leave a comment

For Episode 5, I sat down with with David Droga, a well known creative exec from the advertising world. David is the current (and outgoing) CEO of Accenture’s creative agency called Song. Our conversation was originally recorded for the Where AI Works podcast from The Wharton School. I felt our discussion was particularly relevant for Creative Intelligence and am excited to bring it to you.

We touched upon many themes centered around the future of human creativity and creative jobs in a world where AI can write, design, animate, compose music, and more. Here are the themes that stayed with me:

1. “Not All Creativity Is Worth Saving”

Timestamp: ~4:45–6:30

David argues that much of what passes for creative output—especially in marketing and media—is formulaic, safe, and driven by conformity. AI, rather than threatening creativity, might actually help us shed the "mediocre middle." The ability of AI to increase content production — especially mediocre content — was not lost on David. He has an interesting take on how AI’s ability to industrialize imagination could in fact help us to get to the exceptional works faster.

2. The Industrialization of Taste and the Risk of Sameness

Timestamp: ~10:00–13:30

I brought up the concern that when everyone uses the same AI tools and best practices, “taste gets industrialized.” My own research set in the context of creative writing shows that even when AI improves writing quality and efficiency, it reduces the aggregate diversity of ideas. This is because everyone is using similar AI tools and generating similar ideas. Where your team of 10 people may have generated a diverse set of ideas in the absence of AI, their collective range might reduce when they are using similar tools. David agrees and emphasizes the uniquely human role in infusing taste.

3. Speed, Cost, and Quality—Now You Can Have All Three

Timestamp: ~23:30–26:00

David notes that AI has collapsed the classic creative tradeoff: “Speed, cost, quality—you used to have to pick two. Now you can get all three.” The tools let creatives get ideas out of their heads faster, prototype quicker, and operate more autonomously. But the real unlock, he argues, comes when creatives stop fearing AI and start using it to do things they previously couldn’t.

If you liked this conversation, you may also enjoy listening to the Where AI Works podcast from The Wharton School. I hosted that podcast in season 1 where I focused primarily on AI’s implications for marketing. My colleague Serguei Netessine is hosting Season 2 and he’s focusing on new business models for AI.

A preview of our next episode: I’m going to confess that I am not yet sure what episode 6 will focus on. Wish me luck as I scramble to get an episode ready in 2 weeks.

Leave a comment

Share

Subscribe now

It’s the most quoted line in the future-of-work playbook. But what if it’s wrong — or at least, dangerously incomplete?

I sat down with Sangeet Paul Choudary — author of the bestselling book Platform Revolution and the upcoming release Reshuffle — to unpack how AI isn’t just shifting individual tasks. It’s reshaping entire systems of work.

Sangeet makes an important point that the most quoted line about AI not taking our jobs is "seductive precisely because it feels empowering. It makes you feel like you’ve figured something out ... it gives you just enough conceptual clarity to stop asking the harder questions that really matter: How does AI change the structure of work? How does it restructure workflows?"

Subscribe now

A few themes that stuck with me:

• AI is Changing the System, Not Just the Tasks: Sangeet compares our current framing of AI’s impact to France’s misplaced faith in the Maginot Line during WWII. Just as France failed to anticipate Germany’s blitzkrieg (a new system of warfare), many today assume AI is just about replacing or augmenting existing tasks. But the real shift is at the system level: how organizations are structured, how competition works, and what creates value. (Timestamp: ~4:00 – 6:30)

• Specialized Skills Becoming Generalist Capabilities: Sangeet reminds us that typists weren’t displaced by other typists using word processors; they were displaced because the system of work changed and their specialized skill became a generalist requirement with the advent of word processors. The same is now happening in coding, design, and more. (timestamp: ~20:10 – 24:30)

• You Might Be Training AI to Replace You: If you're in consulting, law, or pharma and using 3rd-party AI tools without proprietary safeguards, you may be feeding sensitive industry knowledge into the very systems that will eventually automate your role. (Timestamp: ~35:50 – 36:35)

• Follow the constraint: As AI commoditizes certain capabilities—answering questions, generating code, drafting content—the source of value shifts. He argues that instead of clinging to skills that are becoming abundant, we should identify where scarcity (and thus economic value) is emerging. (Timestamp: ~13:00 – 20:00)

Let me share my reflection on this last point. In my job as a business school professor, the constraint is no longer knowledge transfer. AI can handle that exceedingly well. The real value now lies in motivation, inspiration, and helping students understand why something matters and feel something for a subject. That shift—from being a transmitter of knowledge to a cultivator of curiosity—is how I now think of my job. To me, it is also a blueprint for many knowledge workers and “experts.” The value isn’t just in what you know, but in the emotional and motivational leverage you can create. You’re not just a source of answers—you’re a cultivator of meaningful inquiry. A curator and tastemaker.

Talking about curation and tastemaking, here’s a preview of my next episode. In Episode 5, I will be chatting with David Droga, current (and outgoing) CEO of Accenture Song. David is among the most well known names in advertising. We will discuss whether AI is the best creative partner for creative workers or a threat to them? We will also explore what will be the role for humans when AI can excel at creativity. And how creative jobs such as that of a CMO or a creative director will evolve with AI. That should be a fun discussion.

Timestamps

0:00 Introduction – AI Won’t Replace You? Rethinking the Premise

2:40 Systemic Change vs Task Replacement: How AI Reshapes Work

6:58 AI changes entire systems of work: Shipping Containers to Word Processors

13:36 Economic value depends on the constraint:: The case of vibe coding

18:17 Teaching, Knowledge, and Scarcity in the Age of AI

21:24 Power Shifts Across Industries

29:48 Structure, Strategy, and Survival

35:59 How to Personally Adapt

40:03 Takeaways – Follow the Constraints, Not Just the Tasks

Subscribe now

Share

Leave a comment

My own personal framework for the durable skills that matter post AI is inspired by the 6Cs framework. The 6Cs in education are content, collaboration, communication, critical thinking, creativity, and confidence. AI is getting incredibly good at content skills (reading, writing, math, science). This makes the other 5C skills more differentiated for humans. Does that mean content skills are not as important? AI can do those well but we still need to know them so we can stress test AI’s answers and evaluate them more critically (so AI is more an accelerator for our content skills). That tells me that 6Cs matter and we should know the content skills well and use AI to accelerate such content work.

A preview of our next episode: At one point in our conversation, Keith mentioned the oft-used phrase “You won't lose your job to AI, but to someone using AI.” I have often wondered about this statement. In some ways, the statement is true. Not having AI skills will make workers less competitive against those who know how to leverage AI in their work. But, in many ways, the statement can be misleading IMO. It can lull us into a false sense of security. In Episode 4, I’ll be discussing this theme with author Sangeet Paul Chaudhry.

Timestamps

Introduction (0:00 - 1:30)

Reskilling Knowledge and Frontline Workers in an AI-Enabled World (1:30 - 8:00)

Redesigning Workflows Around AI (8:00 - 17:30)

Why Are We Not Seeing Redesigned Workflows Yet? (17:30 - 24:10)

Managing Worker Fears: Leadership Advice & Personal Reflections (24:10 - 28:35)

Call Center Workers & Broader Implications (or How do I adapt)? (28:35 - 34:45)

Upskilling Strategies & Feedback Loop (34:45 - 43:15)

One Piece of Advise for Listeners (43:15 - 46:50)

Closing Segment & Kartik’s Reflections (47:05 - End)

We covered:
🔹 The infamous auction where Geoffrey Hinton (who many consider as one of the godfathers of moderns AI) “sold” himself and his students to Google for over $30M
🔹 What early AI cultures at Google/Deepmind and OpenAI reveal about where we’re headed next
🔹 Who is taking AI ethics seriously versus paying lip service to it

The part that stayed with me the most was our discussion of the “brain-stretching paradox [that] the people who say they are most worried about A.I. are among the most determined to create it and enjoy its riches. They have justified their ambition with their strong belief that they alone can keep A.I. from endangering Earth.”

This episode is full of stories that show how who builds AI matters just as much as what AI can do. Hope you enjoy this episode. Comments as well as suggestions for future topics and guests are always welcome.

Episode Segments:

0:00 Introduction

2:44 Cade’s Approach to Writing Genius Makers

8:15 Hinton: A Godfather of AI

15:00 AI versus Human Creativity?

19:25 DeepMind’s Founder and the Contrast with Hinton

24:05 Hinton’s Persistence

27:36 Companies behind AI

33:55 Google’s AI Leadership and OpenAI Threat

38:55 Writing and AI

45:00 Audience Q/A

59:40 Last Question

1:03:33 Kartik’s Closing Thoughts

Share

Leave a comment

1. How is AI changing education? What are the new things Khan Academy can do with AI that they could only have dreamed about in the past?

2. AI can become a crutch for students and magnify lazy intent? How does Sal think about this issue?

3. As someone involved in high school education for nearly two decades, what skills does Sal think will matter going forward? Does AI change the skills we should teach or seek to acquire?

Timestamps

0:00 Intro

2:12 Founding story

5:28 Non-profit status of Khan Academy

11:25 Meeting Sam Altman

17:25 How can AI transform education

26:07 KhanMigo: Khan Academy’s AI Tutor

29:36 Will AI be a crutch for students?

40:16 Re-Skilling for the future

47:14 What skills will matter?

51:05 Optimism about AI

52:48 Takeaways

If you prefer to watch/listen on a different platform, here is the full list of platforms on which you can access Creative Intelligence https://linktr.ee/creativeintelligence .

But when thinking machines can now outthink us, what does that mean for how we learn, create, and grow to meet our potential?

I’ve been working on something new — a podcast (also called Creative Intelligence). It’s a series of conversations about AI, creativity, and the future of work, drawing on my experience in AI education, entrepreneurship, and mindfulness. The first episode drops May 20, featuring Sal Khan, founder of Khan Academy. We talk about:

• Why Sal is optimistic about AI’s impact on education

• How assessments must evolve

• Will AI become a crutch for students

• Why purpose still matters in an AI-driven world

Please save or bookmark wherever you listen to your podcasts:

• Spotify

• Apple Podcasts

• YouTube

• Subscribe on Substack

Thanks for being an early supporter of this Substack. As a token of appreciation, I’ll be sharing exclusive content with subscribers that won’t be available anywhere else. I hope you will enjoy the podcast and I appreciate your help in spreading the word by sharing my LinkedIn or Twitter posts (or simply sharing the podcast with your friends).

It was October 2022. I had enrolled in a 10-day silent meditation program called Vipassana to take a break from the day-to-day stress of running a startup. I launched Jumpcut in early 2020 with the vision of building the first AI-powered movie studio. Finding it hard to operate on venture timelines—the time from initial story development to releasing a movie was simply too long—we pivoted to integrating blockchain technology into Film/TV. Within six months, the Web3 market had collapsed. I went to my board with a proposal for yet another pivot, asking them to bet on our new direction rather than simply take their money back. The stress of pivots, not finding PM fit, and explaining all this to investors was starting to show. This is when I remembered a conversation with a former student, Chris, who had told me about the wonders of a long, multi-day, Vipassana meditation course.

Upon arrival at the Vipassana venue in Palm Springs on that day in October, I surrendered all devices and distractions to focus on inner calm. We were not supposed to make eye contact with anyone or read books or take notes. Daily sessions began at 4:30 a.m. and lasted until about 9 p.m., with only brief meal and restroom breaks (roughly 12 hours of meditation per day). While I expected calm, my mind was instead besieged by anxious thoughts every day—what if the company failed, if the next pivot cost us more money, or if I had to lay off employees, or if investors felt I had let them down? With nothing to do and no external stimuli, a mind under stress will progressively gravitate towards more problems. Sure enough, the questions got progressively worse. If I failed, would people respect me? (and more such questions)

There was a daily emergency office hour with Tim, our teacher, and some attendees came by on an occasional day. I was the “most loyal customer” at these office hours. When I mentioned that I was struggling to sleep, Tim said, “It’s OK. Buddha only slept four hours.” I looked at him in disbelief, wondering how the enlightened one had become a benchmark for a mere mortal just trying to get through the night. By the third day, I began having what I now realize were panic attacks but at the time felt like deathly terror. I had no past experience with that kind of fear, which made it all the more disorienting. “What’s the worst that can happen?” Tim asked me another day while recapping his (far worse) ordeal the first time he did Vipassana. He survived and so would I, he insisted. I was honestly worried I was going insane–a strange sense of panic had gripped me. I begged Tim to let me leave. “Would you walk away in the middle of a surgery just because it is uncomfortable? This is brain surgery in progress,” he said another day while convincing me to persist. Somehow, I soldiered on. By the time I was done, I was completely shaken. For months after that, I struggled to connect with day-to-day reality. Here I had thought I’d find calm and clarity and answers for questions facing my startup and instead I came out with bigger questions. How could I find myself feeling so much panic sitting all alone in a meditation retreat with no obvious threat? Was something wrong with my mind? What got me here? Would I come out of it?

Figure: We were not allowed to take notes but I found a napkin in the room and scribbled some notes of encouragement for myself

Subscribe now

“The wound is the place where the Light enters you.” — Rumi

Returning home, I began to reflect on my journey as an academic entrepreneur. My very first venture was founded with a fellow CMU graduate student who had started a web business selling baby and maternal products. I joined later as her technical cofounder. We were early in the days of e-commerce, and while we managed to attract customers and our startup was even profiled in the New York Times, we soon encountered operational challenges and eventually shut down the business. The idea itself was sound; Diapers.com launched by two Wharton MBA students five years later, had a similar focus but a better outcome (acquired by Amazon). We were perhaps a bit early. More importantly, as first-time entrepreneurs, we erred by not surrounding ourselves with great advisors who could guide us through the journey. And to be honest, I wasn’t sufficiently committed to stick it out long enough for the venture to make it.

My next major venture, Yodle, was founded in 2005 with some brilliant undergrads at Penn. The startup did incredibly well and we managed advertising for nearly 40,000 SMBs by the time of our acquisition in 2016. I was active during the first five years having built our core ML and optimization algorithms for bidding and keyword generation. Key to our early success was assembling the right founding team: a founder, Nate, with relentless hustle and intimate knowledge of customer needs (his dad ran an SMB), a strong data-science-based product moat (which I ran), and exceptionally talented early salespeople. I recollect my first meeting with cofounder Ben, our first salesperson. He was probably around 20 years old then but looked more like 15. I doubted someone who looked so young could do B2B sales but Ben turned out to be one of the best salespersons I have met. Sometimes, magic happens when the right people serendipitously run into each other. Our CEO went on to build a great sales and BD org. But an important driver of success wasn’t something we did but simply that the market opportunity and our timing were just right. SMBs were moving their ad budgets from Yellow Pages to online at a rapid rate. We didn’t need to be perfect. We had to ensure that we were among the top 3 players in a fast-growing space. The rising tide would lift all boats including ours.

With Jumpcut, I hadn’t chosen the problem because the market opportunity and growth were right. It was because I was fundamentally excited about making movies and working with creative professionals. I started it to enjoy the process of building a startup and pursue a lifelong desire to make movies. But if my motivation was so pure, why did I feel stressed when things weren’t going as per plan? Along the way, I raised money and hired a team and struggled to hold on to the lightness and joy while meeting the expectations of a venture-funded business. Hence the series of pivots and desperation to find PM fit.

A New Me: The Mindful Entrepreneur

“Pain is inevitable. Suffering is optional.” — Buddhist teaching sometimes attributed to Haruki Murakami

The first few months after the meditation retreat felt like free fall. I felt like I was staring into an abyss with nowhere to go. I couldn’t go back to where I came from, but it was not clear where to go from here. Over time, I began to realize that what I needed to fix wasn’t the operational aspects of my company but the internal dialog. One thing the meditation made clear was that I had ignored – bottled up – a lot of my worries while trying to hunker down and stay focused on the mission. What I was calling (and many of us call) stress was a series of subconscious worries that I was not willing to consciously hear. “We’re the only species who will force ourselves to do something for years that we don’t enjoy by telling ourselves, ‘We have to do this,’” says author Dr. Martha Beck. The other issue was that I was simply not even aware of the internal dialog going on in my head until this meditation. It was through meditation that I learned to tune in and listen to my own self-chatter.

Something else happened. Previously, I’d put up a bold front in front of my employees and investors, worrying that fear can be infectious . That was a big load to carry. I started to be more transparent with my employees. I encouraged them to take on a bigger role in the Org and discover hidden superpowers within themselves. I encouraged my younger cofounder Dilip to be more of a leader. Slowly, things started to feel lighter. Dilip developed a prototype of Scriptsense, our AI tool to help Producers and agents process the massive volume of screenplays coming in (that were often lying in the unread pile). After a brief experiment to bring in an outside CEO, I returned to an active leadership role at the company. This time, I had all the tough conversations I needed to have – with myself, with investors, employees, and partners. We soon found product-market fit and started hearing from many of our users how our tool was like a godsend to them. The startup recently got acquired by another player in the space. It was not a massive financial outcome. But the story of a startup is multidimensional. We landed the plane in the middle of a storm. We built a great product that was not only considered best in class but was also the first AI product to find meaningful adoption in Hollywood, an industry that is slow to embrace tech. In 2024, the tool processed nearly 10,000 scripts for over 400 companies in the industry.

But most of all, the greatest personal growth was that I met many of my fears and worries. Initially, I wasn't even aware of them. Then, when I met them during Vipassana, I was afraid of the fear itself (“anxiety about anxiety” is a thing). Tbh, I wanted to run away from them, and briefly I did. Next I learned to recognize they were toothless tigers. And after being patient with them, I found a way to hug them, learn from them, and grow as a founder and as a person. So, whatever my next adventure is, it will be about discovering new fears and self-defeating beliefs and finding ways to grow.

Ultimately, the biggest battles are not the ones we fight on the outside. The battle is always within.

PS: Some people have emailed me and asked about meditation and Vipassana. Meditation is a great tool to cultivate presence and manage mental clutter. If you are in a high stress job (or otherwise), I highly recommend it (and other mindfulness practices). That said, Vipassana is extremely intense for most people. In fact, I have since been informed that it’s not a good idea when one is under a lot of stress. My analogy is it’s like an ultra-marathon. You don’t want to try it without sufficient training.

Subscribe now

Leave a comment

Share

The Limits of LLM Accuracy

A recent OpenAI study tested LLMs on over 4,000 difficult, fact-based questions with only one correct answer. Example: “What day, month, and year was Carrie Underwood’s album ‘Cry Pretty’ certified Gold by the RIAA?” The results were eye-opening:

These questions were intentionally hard—selected because at least one model got them wrong. So, such a low rate of correct answers is expected and not where I want to focus your attention. The real problem lies in the low factuality rate when models attempted an answer (last column). Given that LLMs often give incorrect responses with confidence—and people tend to trust confident-sounding LLM answers—there is an issue here.

The LLM-as-AGI Illusion

Our brains have multiple specialized centers each of which perform different functions: the frontal lobe, temporal lobe, limbic system, etc. But LLMs are being overloaded as monolithic AGI system—one that undeniably excels at language but thrust with responsibilities far beyond their capabilities.

Take reasoning, for example. Reasoning models like O3 are fine-tuned on reasoning datasets and do well at what we can think of as pattern-based reasoning; map new problems to known ones and apply similar approaches. But they don’t understand logic so they struggle with novel problems.

Prof. Vince Conitzer highlighted this in a recent post on linkedIn, showing how LLMs falter when reasoning through unfamiliar problems—and often bluff their way through:

Many reasoning steps alter:

A Better Approach: LLMs as Language Hubs

Instead of forcing LLMs to perform tasks they aren’t suited for, why not leverage them as a language hub? For a reasoning problem such as the one above, imagine using an LLM to:

1. Take a brand new reasoning problem in (say) English and extract structured logical statements/premises in the syntax of a logic solver (i.e translate from one language to another).

2. Pass these statements to a logic solver which then does the solving.

3. Convert the solver’s output/conclusion back into natural language.

I suspect this approach would significantly improve reasoning performance while giving us logical explanations and also reducing energy costs. In short, rather than treating LLMs as AGI capable of everything, we use them as language processors that interface with other computational systems (search engines, logic solvers, databases) to generate reliable results.

To be clear, I think there is room in the long run for both the agentic and AGI visions that OpenAI and others are promoting. May be, it even looks like a Transformer (which I doubt). But I think many real-world problems are better served by LLMs acting as intelligent intermediaries rather than an all-in-one AGI.

Other News

• I promised to cover Deepseek but it’s already been over a month since my last post. So I’ll instead share this LinkedIn post of mine where I argued that calling Deepseek (or Llama) as open source is being too generous with that term. Interestingly, since then, Deepseek has open sourced multiple low-level packages and libraries so it is definitely a step forward for open source models.

• Upcoming posts: (i) NYTimes says chatbots are defeating doctors at diagnosis. How seriously should we take that statement? (ii) Mindfulness, intelligence, and AI as a tool to understand human consciousness, (iii) How exactly do AI critics go from a system like ChatGPT to “AI can be an existential threat for humanity”?
  

  Leave a comment

  Share

  Subscribe now

We recently conducted one of the first studies to examine how human-AI collaboration design affects human productivity, satisfaction, quality of work, and diversity of creative output.

Subscribe now

Designing Creative Work

The creative process is commonly divided into two phases: a design/pre-production phase involving ideation and planning and a production phase which involves implementing the plan and review/refinement. We study human-AI collaboration in the context of creative writing. With creative writing in particular, the pre-production/design phase consists of idea generation and producing a story outline whereas the production phase consists of drafting and editing.

Study Design

On Day 1, participants completed an initial survey assessing their English proficiency, creative writing skills and experience. They then wrote a 1,000-word story without any AI assistance following a seven-sequence structure commonly used in fiction writing. On Day 2, participants wrote a new 1,000-word story using ChatGPT 3.5, following their assigned collaboration model. Participants were randomly assigned to one of three AI collaboration models which reflect some of the common ways in which we see AI being integrated into work in practice:

• Human Confirmation, where the AI handles all tasks and humans confirm or reject the output

• Human Creativity, where humans perform pre-production (i.e., ideation and outlining) without AI assistance, while AI executes production phase (i.e., drafting and editing) with humans confirming or rejecting AI output like in Human Confirmation design.

• Copilot, where humans and AI collaborate throughout the entire process from ideation all the way to editing.

Completion Time

We observed no statistically significant variations in total completion time across the different collaboration designs. Total completion time went down by 36.2%, with the most significant gains during the drafting stage (68.6% reduction). So all models have the potential to deliver big productivity gains.

Writing Quality

Without AI assistance (Day 1), overall story quality was not statistically different across groups (see figure below). With AI collaboration (Day 2), the Human Confirmation group, where AI managed creative tasks, produced lower-quality stories compared to the Human Creativity and Copilot groups, where humans performed or collaborated in creative tasks. No significant differences between the latter two groups. In short, human involvement in early creative aspects of writing (ideation and structuring) enhances overall quality when using AI assistance.

In Figure B below, the X-axis represents the quality of stories on Day 1. In other words, it’s a measure of writer skill level. The Y-axis is the story quality on day 2 with AI assistance. AI reduces quality differential between high versus low skilled writers, as seen from Day 2 versus Day 1 plot (slopes << 1). This aligns with other studies that suggest that AI bridges skill-based performance gaps. Human Confirmation design, which minimized human creative input, shows the lowest slope (not significantly different from zero implying that skill-based differences in story quality almost disappeared). Notably, the lower slope with Human Confirmation was driven by higher-skilled writers being less effective in this design (high-skilled writers fared worse with this design than the other two designs) and not because lower-skilled writers benefited more from it. These results suggest that allowing active human involvement in creative tasks provides more opportunity for high-skilled writers to excel.

User Satisfaction

To measure participant satisfaction, we measured overall satisfaction with the collaboration model, whether the design offered them enough flexibility, whether they felt the process was effective for them, and whether they would reuse the process again,. The Human Creativity and Copilot groups reported higher overall satisfaction, flexibility, process effectiveness, and intent to reuse the process compared to the Human Confirmation group. When comparing AI-assisted to human-only writing (i.e. Day 2 vs Day 1), the Human Creativity and Copilot groups experienced increased satisfaction, effectiveness, and intent to reuse the process relative to human-only modes of writing; the Human Confirmation group showed no improvement in satisfaction.

Creative Diversity
Next we measured the diversity of stories written by participants. In the absence of AI support, our participants gravitated primarily towards writing Young Adult stories, which simply reflected their age (students). In contrast, when they had AI support, we saw an increase in the frequency of other genres such as Fantasy, Mystery, and Sci-Fi. In short, AI helped individual writers explore new writing styles and genres which is a big plus.

An analysis of story similarity shows that AI-assisted writing leads to increased semantic similarity among stories within treatment groups. The Human Confirmation group, where AI had the greatest role, showed the highest similarity, followed by Copilot, Human Creativity, and writing without AI (see figure below). Even though we found that individuals are exploring more with AI, different users using AI will explore similarly thereby reducing aggregate creative diversity. In short, excessive use of AI by workers in creative tasks will reduce creative diversity in organizations. But human involvement in creative tasks mitigates this effect, with the Human Creativity model producing the most diverse stories among AI-assisted groups.

Participant Feedback

Participants in the Human Confirmation group expressed feeling detached from their final stories, highlighted struggles with getting AI to produce content fully aligned with their vision, and found AI-generated stories lacking personal touch. For instance, one participant mentioned feeling "super removed from the content of the story," which led to a final product they were "far from proud of."

End Notes

Many recent studies have asked “Does AI help increase human productivity?” I think that answer is now clear and it’s time to switch our focus towards a different question: “How should we design the future of work with humans and AI to improve productivity and work quality while preserving work's meaning and value?” A central insight of our research is that the human creative role remains essential, even as AI capabilities advance. As you think about AI integration, please keep in mind that workflow design will have a big impact. The full paper is here.

Community Notes

Many of you have written to me about Deepseek. There also appears to be a lot of misunderstanding about it. My next post (which I hope will not take more than a few days) will be a Deepseek FAQ and its implications for AI.

Subscribe now

Leave a comment

AAt a recent conference, I shared some tips on how to reduce the carbon footprint of AI use, and several attendees expressed interest. Their reactions inspired me to write this post: How can we keep using these powerful tools while minimizing their environmental impact?

What Affects the Energy Usage of LLMs?

The energy consumption of LLMs depends on several factors:

1. Model Size: Larger models with more parameters require more computational resources, both for training and inference (i.e. generating responses).

2. Input and Output Tokens: The length of your input (what you type or files you upload) and the model’s output (its response) directly influence energy usage. Longer interactions require more processing power.

3. Model Optimization: Newer models are often optimized to deliver comparable results with fewer computational resources. For example, newer models use advanced pruning techniques and fine-tuning strategies to reduce the number of active parameters and streamline computations.

OpenAI’s Models: A Look at Energy Usage

OpenAI offers several models tailored to different tasks. Here’s a breakdown:

• GPT-4: OpenAI’s model with the highest number of parameters. It excels in complex, nuanced tasks but comes at a significant energy cost.

• GPT-4o: Suitable for general-purpose tasks. It has a reduced number of parameters than GPT-4, making it more computationally efficient.

• GPT-4o Mini: A smaller, more energy-efficient variant of GPT-4o, ideal for quick and simple queries.

• GPT-o1: Focuses on text-based reasoning using a Chain-of-thought approach, similar to human reasoning. The model is slower (the cost of "thinking") but it’s much smaller and likely consumes less energy than GPT-4 and GPT-4o.

• GPT-o1 Mini: A lightweight version of GPT-o1. Consumes less energy while maintaining decent performance for simple reasoning tasks.

• GPT-o3: Successor to o1 and is capable of advanced reasoning. Not available to most people. Amazing reasoning capabilities but massive carbon footprint.

Choosing the Right Model for Your Task

To minimize your carbon footprint while using ChatGPT, it's important to choose the right model based on the task at hand. Here’s how to do it:

• Summarization or Quick Information Retrieval: For these tasks, smaller models like GPT-4o Mini or GPT-o1 Mini are sufficient. They don't require the full power of larger models, making them more energy-efficient.

• Creative Writing or Complex Analysis: For tasks requiring nuance, opt for GPT-4o. However, consider whether splitting the task into smaller, simpler components might allow you to use a smaller model.

• Testing and Experimentation: If you’re experimenting, start with a smaller model (GPT-4o Mini or GPT-o1 Mini). Upgrade only if the results are insufficient.

To make these choices, simply select the model from the dropdown menu in your ChatGPT window. For developers accessing models through the API, keep in mind that smaller models are not only more energy-efficient but also more cost-effective. So, always start with smaller models and only use larger ones when necessary.

For those of you who use Anthropic's Claude, they have three models: Haiku, Sonnet, and Opus. Haiku is the smallest model (i.e. least number of parameters), Opus is the largest and most intelligent model, and Sonnet is in between. I generally tend to use Haiku for simpler tasks such as summarization and use Sonnet for more advanced tasks that require some reasoning. Sonnet works so well that I have so far not found a compelling use case justifying the higher cost and energy consumption of Opus.

By thoughtfully selecting the appropriate model, you can achieve your goals while keeping your energy usage and carbon footprint in check.

Food for Thought

While the o3 energy usage stat I start this post with is alarming, let's keep in mind that "the cost of GPT4 quality results has declined by more than 99% in the last two years." So I fully expect cost per task to come down dramatically.

But don't let that lull you into complacency. Economist William Jevons demonstrated that increased efficiency in steam engines led to higher coal consumption overall. In the same way, as the cost per task drops, aggregate usage and energy consumption are likely to rise Beware Jevons paradox.

PS: If you work in the area of AI and sustainability, I'd love to hear more about your work. Please share details below.

Leave a comment

Subscribe now