In June 2025, a molecule called rentosertib showed up in Nature Medicine with a Phase 2a readout for idiopathic pulmonary fibrosis. Patients on the 60mg arm gained lung function: a mean +98.4 mL in forced vital capacity, against a 20.3 mL decline on placebo. That result matters less than how the molecule got there. An AI system identified the disease target and designed the compound that hits it. The whole discovery leg took under 18 months and cost roughly $2.6 million, against a conventional average north of $600 million.

It’s the most important proof-of-concept the industry has produced in a decade. In July 2026 it became the first drug with both an AI-identified target and an AI-designed molecule to initiate Phase 3. And it still faces roughly the same odds of failing that trial as any other molecule. Both of those things are the story. If you only hear the first one, you’ll build the wrong strategy.

What’s actually happening

The compression is real, and it’s not confined to discovery. Insilico has nominated over thirty preclinical candidates on 12-to-18-month timelines against a traditional average of about 4.5 years. More than 170 AI-originated drug programs are now in clinical development, with 15 to 20 expected to reach pivotal trials this year. In the clinic, AI-powered patient-matching platforms have compressed prescreening from roughly eighteen weeks to under one week in published case studies.

On the plant floor, AI visual inspection runs on commercial lines today, and AI-supported real-time release testing is moving from pilot into validated commercial use, though batch disposition remains human-gated. And the regulator itself now runs on AI: the FDA launched Elsa agency-wide in June 2025, and by May 2026 had upgraded it to run custom AI agents, with staff use climbing from 1 percent to over 80 percent.

The market has noticed. AI in drug discovery alone is a three-to-five-billion-dollar segment in 2026, on track to roughly double by 2030. Every stage of the lifecycle now has an AI layer. That’s the part everyone reports.

Here’s the part they skip. None of it has moved the wall that actually kills drugs.

AI-discovered molecules clear Phase 1 at 80 to 90 percent, well above the historical ~52, because Phase 1 is a safety gate and cleaner chemistry clears it. But Phase 2, where biology decides whether the drug works, still runs around 40 percent. Same as always.

No AI-discovered drug has reached approval yet. The closest, Takeda’s zasocitinib — an AI-assisted discovery that originated at Nimbus Therapeutics before Takeda acquired it — cleared Phase 3 and is heading for an FDA filing this year, and it still had to survive the same trials, on the same odds, as anything else. Eighty percent of trials still miss their enrollment targets. We got dramatically faster at the parts that were never the bottleneck.

Subscribe now

The model most orgs still run on

Ask anyone to draw the drug lifecycle and you get the same picture: discovery, then preclinical, then clinical, then CMC, then regulatory, then commercial. Discrete boxes. A baton passed left to right. Each function optimizes its own leg and hands off a clean package to the next.

That relay race is baked into everything. It’s how the org chart is drawn, how budgets are fenced, how decision rights are assigned. Discovery owns the molecule until IND, then it belongs to development. CMC inherits whatever chemistry it’s handed. Regulatory writes the story at the end. The whole system is built on the assumption that the stages are separable and the handoffs are clean.

What the loop actually is

That assumption is what AI is dissolving. The stages remain. The separation between them doesn’t.

When a discovery model is trained on why molecules failed in Phase 2, clinical failure data is now shaping the first sketch of a compound. When manufacturability constraints feed back into molecular design, the plant floor is in the room during discovery. When real-world evidence loops into label expansion, the commercial stage is talking to the clinical one. The baton isn’t being passed forward anymore. Every stage is feeding every other stage, continuously. The relay race is becoming a loop.

Why the wall didn’t move, and what did

If the loop is so powerful, why is Phase 2 still 40 percent? Because speed was never what stood between a molecule and a patient. Judgment was. The hard calls have always lived at the handoffs: which target is real, which signal is noise, which risk is worth carrying into a pivotal trial. AI compressed the work between the decisions. It didn’t make the decisions. It made more of them, faster, and moved them closer together.

And it added a decision-maker nobody’s org chart accounts for. When the FDA’s reviewer is running Elsa, your AI-assisted submission is being read by an AI-assisted regulator. The loop now includes the agency. That’s a preview of where accountability is heading. When AI is on both sides of the submission, the only thing that isn’t automated is the question of who’s willing to sign their name to the judgment. AI assists, humans decide stops being a slogan and becomes the entire job.

So the scarce resource changed. In a relay race, the constraint was the speed of each leg. In a loop, the constraint is decision rights. Who holds authority over the calls the AI is now surfacing at ten times the rate, at every stage at once, with the handoffs too blurred to tell you whose call it even is.

The call you’re actually being asked to make

Here’s the uncomfortable part for anyone running biotech operations. Your organization is almost certainly still structured as a relay race, with functional silos, staged budgets, and decision rights fenced by stage, while the work has become a loop. That gap is where good molecules die. The science clears. The organization stalls.

So don’t start with another AI pilot in each function. Start with one live program and trace it. Where does a discovery call now wait on manufacturing evidence? Where does a clinical decision depend on data the team can’t pull fast enough? Where does the work stall because the next owner needs the context rebuilt in a deck, a tracker, or a meeting? That’s your handoff map. Then ask the harder question: when one of those handoffs breaks, who can actually make the call? If the answer is a committee with no clear owner, a functional escalation, or “we’ll decide at the next governance meeting,” your loop is running on relay-race governance. That’s the thing to fix: who holds decision rights in a loop.

Share

The relay race rewarded the fastest leg. The loop rewards whoever owns the handoffs. That role, the person who can hold judgment across stages that no longer hand off cleanly, didn’t exist on the old org chart. It’s being created quietly, right now, in the gap between how the work runs and how the org is still drawn. Someone is going to step into it.

That role belongs to someone who's spent fifteen years watching where programs stall at every handoff. The vocabulary can be learned in a weekend. The scar tissue takes a career.

Thanks for being here,

Jizel

Next in this series: if the lifecycle is now a loop, what does a company built with no handoffs actually look like from discovery to delivery, and where are the pieces still missing? That gap is the business. Part two, “The Company With No Handoffs,” maps it.

Leave a comment

The case studies I went looking for did not exist where I expected them. I assumed the AI-implementation success stories in clinical-stage biotech would belong to the well-funded Series C and beyond. They mostly did not. Coefficient Bio was eight months old with fewer than ten people when Anthropic paid $400M to acquire it. Insilico Medicine moved 22 drug candidates from project initiation to preclinical nomination in 12 to 18 months. Relation Therapeutics, lean and London-based, won a $45 million upfront partnership from GSK on the strength of a lab-in-the-loop platform.

These are not pharma-scale orgs. They are the size of clinical-stage biotechs that most of you actually work inside.

The implementation pattern under their results is consistent. This issue is the deep dive on that pattern sequence.

The replies after Part 1 told me something. Most of the people who opened it are not at Merck. They are at clinical-stage orgs with one platform, a handful of CDMOs, and a data science hire who became a de facto AI program of one. That gap between the enterprise case study and the lean operational reality is exactly what this issue is for.

The hidden adoption gap

AI-first biotechs integrate AI into core drug discovery at 75% adoption. Traditional biotech sits five times lower. The reason most ops leaders cite is budget, headcount, compute, semantic data layer. All real. None of them is the binding constraint anymore.

Cloud GPU on H100 runs $2.50 to $3.20 per hour. A typical large-molecule discovery campaign clocks 50,000 to 200,000 GPU-hours, which puts the compute envelope at $125K to $640K for the campaign. Lab-as-a-Service from Emerald Cloud Lab and Strateos lets a clinical-stage team run hundreds of instruments 24/7 without owning any of them. Enterprise platform licensing from Schrödinger, Atomwise, or BenevolentAI runs $500K to $5M annually, scaled to user count. These are Series A and Series B-stage cost envelopes, not big-pharma ones.

The constraint at clinical-stage biotech is not “we cannot afford the stack.” It is “we have not sequenced the stack.” The 22% who scale are running the same three waves. So are the AI-native lean biotechs.

What most clinical-stage teams are still doing

The conventional play at a sub-200-employee biotech is reasonable on paper. Hire one data scientist or Head of IT. Buy one platform. See what sticks. The strategy off-site approves it. The CFO signs off on it.

It is also why the small-biotech AI adoption rate is five times below the AI-native peer set. A single hire and a single model do not produce production AI. They produce a curiosity project that lives in one person’s GitHub for nine months, then quietly stops.

The pattern the AI-native lean biotechs run instead is sequenced, gated, and small. Three waves. Each one cheap. Each one a gate to the next.

The 90-Day Implementation Arc

The same three-phase rhythm shows up across every successful lean-biotech case study I pulled. It maps cleanly to a Tier 1 entry point from the Three-Tier Risk Model.

Wave 1. Days 1 to 30. Triage to one bottleneck. Not three. One. The single operational chokepoint where AI plausibly compresses cycle time by 30% or more inside your GxP envelope.

The shortlist should come from the Tier 1 / Green-light territory I mapped in the Agentic AI in CMC issue. That is the Three-Tier Risk Model from Issue #14 applied to your real workflow inventory: Tier 1 is non-GxP work or GxP-adjacent first-draft work where a human owns the final review, no full computer system validation required, and a clean checkpoint exists before anything leaves the building.

The selection question is this: which workflow is currently eating two to four hours of manual work per week that a well-designed AI workflow could return? Time saved is the right ROI signal for Wave 1. Not accuracy improvement. Not quality scores. Time. Your CFO will ask about it first, and it is the data that makes Wave 2 fundable.

Tier 1 examples that have produced measurable wins at clinical-stage scale:

• Process Development / MSAT. Tech transfer document gap analysis. Stability data trend analysis. Method transfer equivalence summaries. PPQ completion tracking against acceptance criteria.

• Quality. Batch record pre-screening. Environmental monitoring trend analysis across cleanroom suites. Deviation investigation pre-drafting (with rigorous human review before anything enters the QMS).

• Supply Chain. Supplier delivery and quality metric aggregation across CDMOs. Cold chain excursion monitoring against predefined protocols.

• Regulatory. Regulatory intelligence scanning across FDA / EMA / agency Q&A. eCTD module completeness checks. First-draft Module 2 summaries built from Module 3 data, human-reviewed before file.

• Clinical Ops. Investigator brochure section drafting. Protocol consistency checks across versions. Site-selection predictive screening against historical enrollment data.

Each is a structured workflow with a clean human checkpoint and a finite exception set. Each maps directly to Green-light territory in the Agentic AI in CMC framework. Pick one. Resist the temptation to pick three. A second bottleneck inside Wave 1 is the most common failure mode at sub-200-employee orgs.

Wave 2. Days 31 to 90. One pilot, built GxP-grade from Day 1. Skip the sandbox-then-rebuild path. Build the pilot directly inside a production-grade environment, on a model platform that already meets enterprise security and compliance baselines. Four credible stacks for clinical-stage biotech:

• Anthropic Claude. Available directly via the Anthropic API, via Claude Enterprise, or via AWS Bedrock. Strong for long-context regulatory and operational documents in published case studies (Module 3 sections, deviation packages, tech transfer files).

• AWS Bedrock + AgentCore. Used in production by Pfizer (PACT), Sanofi, and Bristol Myers Squibb. Native access to multiple frontier models, including Claude. GxP-defensible cloud baseline.

• Azure OpenAI. The platform behind Boehringer Ingelheim’s iQNow (150,000 work-hours saved in 70 days). Strong fit if your org is already standardized on Microsoft 365 and Power Automate.

• Google Vertex AI / Gemini Enterprise. The Merck-Google Cloud architecture, scaled-down. Strong for orgs with heavy BigQuery or Google Cloud presence.

Form a pod of three to five people for the pilot: one operations SME with domain authority, one quality reviewer with regulatory context, one engineer or vendor partner who can build inside the chosen stack. Validate against real-world data on Day 60. Run a Day 75 go/no-go gate, decision in writing. The lean biotechs that scale built the gate before the pilot started. The ones that stall never kill one.

Wave 3. Days 91 to 365. Stack the wins, not the tools. Every subsequent pilot reuses the three pieces of scaffolding the first pilot built. In plain English:

• The validation layer. The written rules and automated checks that determine whether an AI output is allowed to move forward in your workflow. Ginkgo Bioworks built this as a Pydantic gate that checks reagent availability, volume constraints, and synthetic viability before any AI-proposed experiment touches the bench. At your scale it can be simpler. A reviewer checklist, a structured prompt template, and a deterministic check against your acceptance criteria are enough for the first pilot. Every subsequent pilot inherits and extends the same layer instead of rebuilding it.

• The semantic data scaffolding. The structured definitions of your company-specific terms. Your batch record fields. Your deviation taxonomies. Your CDMO naming conventions. Your validated method codes. AI agents do not inherently understand any of this. Without the scaffolding the agent guesses. With it, the agent operates inside your actual data model. You build this once, in the first pilot, and reuse it across every workflow after.

• The embedded-team governance. The written rules for what a partner engineer or vendor employee inside your pod can and cannot decide. Who signs when. Who reviews when. Who is accountable when the agent surfaces an exception. These get codified in the first pilot’s RACI and inherited by every subsequent pod.

Lab-as-a-Service from Emerald Cloud Lab or Strateos absorbs the experimental capex on the wet-lab side. Platform-Plus-Asset framing ties each new pilot to a specific clinical program, which is what de-risks the spend for your board and your Series B investors. By month twelve, the disciplined lean biotech has three or four production workflows running off one shared stack, not four disconnected pilots.

Why this is more accessible than it looks

Your lean team is more constrained than enterprise pharma. It is also more naturally suited to this work.

Coefficient Bio scaled to an acquisition exit with fewer than ten people because the team did not have the option to spend nine months negotiating a $1B alliance. They built into production from Day 1 because there was no other path. Insilico’s 12-to-18-month target-to-preclinical timeline came from sequencing discipline, not from a bigger budget. The AI-native lean biotechs are hitting 80-90% Phase I success rates versus an industry average of 40-65% because the production discipline is structural, not aspirational.

The GxP layer is unchanged through all of this. Even in a three-person pod with embedded vendor engineers, regulatory accountability stays inside the licensee. The agent doesn’t sign. You do.

From the Field

The pattern I keep watching is this: the teams that make Wave 1 work are not the ones who picked the most sophisticated use case. They are the ones who picked the narrowest one. A regulatory team I have been following spent their first thirty days mapping one workflow end to end: eCTD completeness checking against a predefined review standard. Single application. Single checkpoint before anything moved forward. They validated on Day 60, ran the go/no-go gate on Day 75, and called it in writing. Six months later, that same scaffolding was running two additional workflows without adding a single new hire. They did not add headcount between pilots one and three. They added pods.

Subscribe now

Your Move

If you are a senior ops leader at a clinical-stage biotech, the comparison set is not Merck. It is Coefficient Bio, Relation, and Formation Bio. Four moves that scale at your size.

Pick one bottleneck. Not three. Triage first, pilot second. A second bottleneck in Wave 1 is the most common failure mode at sub-200-employee orgs.

Build inside production from Day 1. Pick the platform stack you identified in Wave 2 and commit to it now. The cost of designing for the sandbox and rebuilding for GxP later is higher than the cost of building inside production once.

Form a 3-to-5 person pod, not a function. One ops SME, one quality reviewer, one engineer or vendor partner. Your hyperscaler’s partner program or a CRO with embedded AI capabilities is the fastest path to that third seat. End-to-end ownership for one workflow. Scale the pod model, not the headcount. That distinction matters. Scaling the pod model means that when you take on your second workflow, you do not hire a department of fifteen. You spin up a second 3-to-5 person pod, identical in composition, working off the same validation layer, the same semantic scaffolding, and the same governance the first pod built. Roles can rotate across pods (your senior quality reviewer can sit on two; your vendor partner can support three). You add pods horizontally, not headcount vertically. By the end of Year One, the lean biotech runs three or four pods off a single shared scaffold. The big-pharma alternative would be a 40-person digital department, twice as expensive, half as fast.

Time-bound the Day-75 gate. Decide go or no-go in writing. The lean biotechs that scale are the ones that kill pilots on cadence and free the pod for the next one. The judgment gap closes inside the gate review, not inside the model.

That is leadership work. And it is the work the M-shaped operator builds before the alliance arrives, not after.

What this comes back to

You do not need a $1 billion hyperscaler partnership to implement AI inside a clinical-stage biotech. You need to copy the discipline of an eight-month-old AI-native team at smaller scale. Pick one bottleneck. Build inside production. Pod the work. Gate the result. Stack the wins. That is the whole playbook.

If this issue mapped to something you are already working through, I want to hear it. Hit reply and tell me which Wave 1 bottleneck your team would start with. I read every response.

Thank you for being here,

Jizel

That closes Issue 16, the two-part series. Part 1 walked the enterprise pattern. This issue walked the lean version. The two are not opposing strategies. They are the same operating logic at different scales. Forward this to one peer at a clinical-stage biotech who is staring at a $1B alliance headline and wondering what their version of it looks like.

Leave a comment

Pfizer saved scientists up to 16,000 hours of search time a year. Boehringer Ingelheim saved almost 150,000 work-hours in 70 business days. Sanofi cut advanced analytics from six months to one. Moderna merged HR and IT under one executive and runs more than 3,000 custom GPT agents.

I went into the 2026 case studies expecting to find a model preference. A favorite vendor. Some clever prompting the winners had figured out. What I found instead: the same five things under every success story.

This is what the 22% built. The 78% are still arguing about which model to implement.

The hidden layer under the alliance headlines

We have spent 2026 watching billion-dollar alliance announcements. Lilly and NVIDIA opened the year at J.P. Morgan with a co-innovation lab and up to $1 billion over five years. Merck followed in April with up to $1 billion into Google Cloud for agentic AI. The headlines are mostly accurate.

Underneath them is a more useful story. A small group of companies has actually scaled AI to measurable productivity. Deloitte’s 2026 Life Sciences Outlook puts it at 22% of leaders who say they have scaled AI, and just 9% reporting significant returns.

Their playbook is not about which model. It is about a stack. Five layers, each one a different ops decision, each one solvable inside a clinical or commercial-stage biotech without a billion-dollar partner.

Takeda’s chief data and technology officer Gabriele Ricci, in the same outlook, called this the period where “discipline and innovation must coexist as the industry matures beyond hype toward measurable productivity from AI and data.”

That mentality is the entry ticket. The stack is what you do once you have it.

What most boards are still solving for

Most board conversations land in the same place. Pick the right model. Pick the right vendor. Pick the right pilot use case.

Fair enough. That framing made sense when the constraint was access to a frontier model. The constraint is no longer access. The constraint is the operating system the model lands inside of.

The Builder’s Stack

The high-performing companies run the same five layers. Each has a named exemplar. Each is auditable.

Layer 1. The Operating Model. AI is not a department in the 22%. It is a wiring change. Moderna put HR and IT under one Chief People and Digital Technology Officer, Tracey Franklin, to put AI fluency next to talent strategy. Lilly and NVIDIA are colocating biology and medicine experts with AI model builders in a single Bay Area lab. The giveaway is org-chart geometry. Decision rights for tool deployment, hiring, and process redesign move out of any one function and into a shared operating layer.

Layer 2. The Semantic Data Layer. AI agents do not inherently understand your ontology. Your batch records, your deviation taxonomies, your CDMO naming conventions, your validated method codes. Until that semantic layer exists, no agent gets past PoC. This wall ends more pilots than a wrong model selection ever has. AbbVie’s ARCH platform connects more than 2 billion points of knowledge across 200-plus internal and external sources, so researchers query one connected graph instead of two hundred silos. The 22% built the data layer first. The 78% trained models on data their own agents could not interpret.

Layer 3. Production Discipline. Design for the production-grade environment first. Demo second. Pfizer’s PACT collaboration with AWS prototyped generative tools inside production infrastructure across 14 projects, saving scientists up to 16,000 hours of search time a year per AWS’s published case study. Sanofi’s Digital Accelerator cut advanced analytics from six months to one. Boehringer’s iQNow, embedded in Word and PowerPoint where drug development documents live, saved almost 150,000 work-hours in 70 business days by Microsoft’s accounting. Vendor-published numbers, so read them as directional. The pattern holds either way: none of these started life as a sandbox demo needing retroactive validation.

Layer 4. Democratization With Guardrails. Sanofi’s plai app puts real-time decision data in the hands of 20,000 Sanofians every day, backed by a 10-week training program. Moderna’s 3,000-plus custom GPTs exist because ordinary employees were trained and trusted to build them inside a bounded platform. The 22% give a lot of people access to a tightly-bounded sandbox. The 78% give a few people access to a wide-open one. That is the difference between fluency at scale and a curiosity project.

Layer 5. Regulatory Credibility. Explainability and FAIR data principles are not corporate optics in the 22%. They are operating requirements. FDA’s January 2025 draft guidance laid out a 7-step, risk-based credibility assessment framework, and in January 2026 FDA and EMA jointly published ten Guiding Principles of Good AI Practice covering human oversight, data governance, and lifecycle controls. Black-box outputs do not survive an inspector. The 22% built explainability in from Layer 1. The 78% are bolting it on six months before their first AI-touched submission. That is the judgment gap, expressed as architecture.

The 22% built five layers. The 78% bought a model.

What this means for your ops org

Two things are true at the same time. None of these five layers requires a billion-dollar partner. And none of them get easier when you skip the earlier one. Skip Layer 2 and you hit the semantic-data wall at month nine, when production agents start asking questions your ontology cannot answer. Skip Layer 4 and you build a fluent digital team next to an analog operating org. Skip Layer 5 and you discover your scaled deployment is unsubmittable.

The GxP layer is unchanged through all of this. The human-in-the-loop stays. The licensee accountability stays. The agent doesn’t sign. You do.

Your Move

If you lead ops at a clinical or commercial-stage biotech, the question is not whether you can replicate Merck. The question is which of the five layers your org is furthest along on, and which one is your real bottleneck.

Four moves a peer would not dismiss as theater.

• Audit your operating model. Where does AI sit on the org chart? If the answer is “we have a digital team,” you are still in the old model. The 22% reshape the wiring, not the org column.

• Audit your data layer. How long does a cross-system query take inside your org? If the answer is “weeks,” your semantic layer is the bottleneck, not your model selection.

• Audit your production discipline. Pick one current AI pilot. Was it designed inside a GxP-grade environment from day one, or was it designed in a sandbox and now needs retroactive validation? The answer predicts whether it scales.

• Audit your democratization. How many operational team members have self-service access to a guardrail-bounded AI sandbox today? If the answer is “the digital team,” scale is years away. If the answer is “most of the function,” you are in the 22% pattern, whether your CFO has noticed yet or not.

That is leadership work. And it is the work the M-shaped operator builds before the alliance arrives.

What this comes back to

The 22% built five layers. The 78% kept arguing about the model. The gap compounds inside the ops org, not inside the data science team. The stack is the work. The model is the easy part.

Hit reply with the layer your org is stuck on. The bottleneck is almost never the one the strategy deck names.

Thank you for being here,

Jizel

Next issue: Issue 16 Part 2: The Builder’s Stack is the enterprise-pharma pattern. Most of you do not work at Merck or Lilly. The real question is what this looks like inside a sub-200-employee clinical-stage biotech with a Series A and a 12-month runway. The Lean Biotech Implementation Playbook. How the lean builders actually do it.

Leave a comment

The AI produced what it was asked to produce. The firm never built the judgment to govern it.

I spent last weekend finishing Anthropic’s AI Fluency certification, and the thing that stayed with me wasn’t a technique. It was a reframe: what the course calls fluency is really the judgment discipline you’ve been building your entire career in regulated biotech CMC ops, now applied to a new kind of collaborator.

The policy will name the walls. Your job is to know which rooms to work in. That’s the gap the Purolea firm couldn’t close.

The firm didn’t lack access to AI tools. They lacked the judgment to know what AI can do and where it must never operate alone in a GMP environment. Learning to draw that line clearly is exactly what this 30 days is designed to build.

You already have the foundational instincts. You’ve been navigating regulated environments where a wrong call has consequences. You evaluate claims, verify data, and escalate when something doesn’t add up. Those instincts translate directly to working well with AI. What’s missing isn’t aptitude. It’s a framework that names what you’re already doing and sharpens it deliberately.

This is that framework.

Most AI training treats this as a technique problem. Learn the right prompts, learn the tool, and you’re done. That’s familiarity, not fluency. The gap Purolea couldn’t close wasn’t about access to AI. It was about not having a judgment structure to govern it.

The 4D Framework: Delegation, Description, Discernment, Diligence

The research on AI fluency, including Anthropic’s own curriculum for working effectively with AI, identifies four core competencies. Not technical skills. Judgment capabilities.

Delegation: knowing what to hand off, and what to hold
Description: communicating precisely enough for AI to be useful
Discernment: evaluating outputs critically before you act on them
Diligence: verifying, documenting, and maintaining accountability at the end

These aren’t sequential steps. They’re a set of interlocking habits. The four weeks below build one at a time, but you’ll find they compound.

Week 1: Delegation : What to Hand Off to AI, and What to Hold

The Purolea warning letter is a case study in failed delegation. The firm delegated to AI something AI cannot hold: regulatory accountability. Under 21 CFR 211.22(c), the quality unit is responsible for approving all procedures and specifications affecting identity, strength, quality, and purity. That responsibility cannot be offloaded to a system that can’t sign, can’t be held accountable, and doesn’t know what it doesn’t know.

Delegation fluency means mapping your work to three zones before you assign anything to AI. (For a deeper look at agentic AI in CMC specifically, see below)

Think of it the way an experienced auditor thinks about any new process your team introduces. They want to know: what is this doing, who decided it was appropriate for this use, how was it tested, and who is accountable if it produces an error? FDA is now asking exactly those questions when it sees AI in a GMP context.

Before we get to the three zones, one sentence that anchors everything: The agent doesn’t sign. You do. That accountability is what makes your judgment more valuable as these systems scale.

Green Zone: AI leads, you verify
Drafting, summarizing, structuring, synthesizing. AI is fast and thorough on these. Your job is to review the output, not produce it from scratch. Examples in biotech ops: first-pass meeting summaries, deviation narrative drafts, literature summaries for regulatory background sections, batch record exception compilations, agenda building for cross-functional reviews.

Yellow Zone: AI assists, you decide
Risk-informed judgment calls. AI surfaces the information; you make the call. Examples: gap analysis against a regulatory guidance, preliminary risk stratification for a FMEA, draft CDMO audit finding summaries (you assign severity), APR/PQR trend flagging where your data is in a format AI can read (you interpret significance). AI can lay out the landscape. Signing off on what it means is yours.

Red Zone: You decide. AI does not touch.
Anything where the regulatory accountability is yours by law, not by preference. Under 21 CFR 211.22(c): approving procedures or specifications affecting identity, strength, quality, and purity. Under 21 CFR 211.100: written procedures for production and process control must exist, be QU-approved, and be followed in execution. Change control decisions with GMP impact. Deviation investigation root cause conclusions. Batch disposition. These tasks aren’t candidates for “AI-assisted.” They’re yours.

The specific skill to build this week: task decomposition. Break a complex workflow into smaller steps and decide which ones AI holds and which ones you do. Same instinct as a project manager assigning work: you have to understand the task well enough to know who can safely hold it.

Week 2: Description : How to Communicate With AI So It Can Actually Help

Most people treat prompting as a word-choice problem. The actual lever is context: giving AI enough specificity about the situation, audience, and constraints to produce something worth using. I wrote more about context libraries here: The Agentic AI Playbook for Regulated Biotech:

In a GMP environment, you already know how to write unambiguous instructions. A vague procedure is a deviation waiting to happen. A vague prompt produces the same result. Same discipline, different medium.

Anthropic calls this the Description-Discernment loop: describe the task clearly, evaluate the AI’s response, refine based on what was missing or off, and repeat. You’re calibrating AI to your specific context, the same way you’d calibrate a new team member to your process requirements.

What good description looks like in biotech ops:

Weak: “Summarize this deviation report.”
Strong: “Summarize this deviation investigation report for a cross-functional review. Audience is QA, manufacturing, and the CMO. Focus on root cause, CAPA status, and any open questions about recurrence potential. Keep it under 200 words. Flag anything that requires a regulatory notification decision.”

The gap between these two prompts comes down to three things: who’s reading it, what decision follows from it, and what constraints matter.

Other high-leverage description patterns for biotech ops:

• MDF narratives and Manufacturing Deviations: specify regulatory filing context, product development stage, and any known agency sensitivities before asking AI to draft

• Stability deviation analyses: include the study design, specification limits, and what “acceptable” means before asking AI to assess. Without that context, you’ll get generalities.

• Regulatory response drafts: specify the agency, the concern, your firm’s position, and what the response needs to accomplish before asking AI to build the structure

• CDMO audit gap reports: give AI the audit standard (ICH Q10, ISO 15378, etc.), your firm’s baseline, and the specific observations before asking it to classify or prioritize gaps

The specific skill to build this week: context engineering. Choose what background information to give AI before you ask it anything. Prompting is what you ask. Context engineering is the briefing you give before you ask. Think of the difference between handing a new contractor a job title and handing them a full onboarding packet.

I’m writing a dedicated prompting guide for biotech ops contexts. Please subscribe so you’ll get notified when it’s live.

Week 3: Discernment : Evaluating What AI Gives You

AI outputs are confident. That’s a feature when the output is correct, and a real problem when it isn’t. The Purolea warning letter is partially a discernment failure: the firm accepted AI-generated documents without applying the expert judgment required to catch errors, omissions, and compliance gaps. FDA’s direct language: “If you use AI as an aid in document creation, you must review the AI generated documents to ensure they were accurate and actually compliant with CGMP.”

Discernment is the habit of not taking AI at its word.

In regulated environments, you already have a version of this. You don’t accept analytical data without reviewing the methodology. You don’t release a batch without reviewing the batch records. Apply the same pattern: AI produces, you review, you verify, and only then does it move forward.

Three discernment questions to ask about any AI output:

1. Is this factually correct in my specific context? AI is trained on broad data, not your regulatory history, your CDMO relationships, your product development stage, or your firm’s quality standards. Generic accuracy is not the same as contextual accuracy.

2. Is this complete, or did it miss something critical? AI outputs often omit qualifications, edge cases, and open questions. Those are precisely the things that matter most in regulated decision-making. A gap analysis that doesn’t surface the gaps it doesn’t know about is worse than no gap analysis.

3. Is this mine to sign? Before using any AI output in a document that carries your name, your function’s name, or your firm’s regulatory standing, answer this question explicitly. If you’re not certain, it goes in Yellow or Red, not Green.

The Description-Discernment loop in practice: If the output doesn’t pass discernment, go back to description. Was the task context incomplete? Was the constraint missing? Iteration is not failure. It’s the workflow.

The specific skill to build this week: hallucination auditing. Actively check AI claims against primary sources rather than reading for plausibility. The same instinct as a QC reviewer checking analytical data against the method spec rather than assuming the instrument was right.

Week 4: Diligence: Verifying, Documenting, and Staying Accountable

Diligence is where the Delegation-Diligence loop closes. You delegated a task. AI produced output. You applied discernment. Now: you verify the final product, document the process, and ensure accountability is clear.

In GMP contexts, ALCOA+ applies to every GMP record AI helps create. Who reviewed it, when, what changed, and whether the final version reflects what was actually approved.

• Attributable: The record should read “Person X reviewed and approved this, using AI as a drafting tool,” not simply “AI generated this.”

• Legible and Contemporaneous: The review happened when it happened. Not backdated, not reconstructed.

• Original and Accurate: The approved version reflects what was actually reviewed and corrected, not the AI first draft, and not a hybrid that no one fully reviewed.

If AI drafted a document that enters your quality system, the review trail needs to be clear: who reviewed it, what standard it was reviewed against, and what was changed before approval. “AI wrote it, I looked at it” is not a QU approval. 21 CFR 211.22(c) sets the bar. That review has to be real.

What diligence looks like in biotech ops:

• AI-drafted SOPs: document the review as you would any authored procedure: reviewer, standard referenced, changes made, approval basis

• AI-assisted deviation narratives: the investigator signs the root cause conclusion, not the AI summary

• AI-generated regulatory response drafts: regulatory affairs lead reviews for strategic accuracy and compliance posture, not just grammar

The Delegation-Diligence Loop: Anthropic’s curriculum frames this as a paired cycle. When you delegate, you commit to the corresponding diligence. Delegation without diligence is the pattern FDA cited in Warning Letter 320-26-58. The loop doesn’t close without the second half.

The specific skill to build this week: output versioning. Keep the AI-generated draft visibly separate from your reviewed and approved version so the review trail is unambiguous. Same principle as maintaining raw data and the analyzed dataset as separate files in your study records.

The Judgment Call

AI fluency is a judgment discipline. The biotech ops leaders who build it now, before their organizations mandate it, before FDA guidance tightens further, before the agentic wave hits GMP workflows, will be the ones defining how it works in their organizations rather than inheriting someone else’s framework.

The Purolea firm had the tools. What they lacked was the judgment to know where the tools stop and where the human begins.

That line is yours to define. The four weeks above are how you start.

The experience you’ve built in CMC and biotech ops gives you a real advantage here. Precision, verification, judgment under regulatory pressure: you’ve been practicing this for years. The 4D framework isn’t asking you to learn something new. It’s asking you to apply what you already know to a new kind of collaborator.

Over the next several weeks, I’ll be going deeper on each of these competencies: the specific skills inside each pillar that make the difference between a framework you understand and one you actually use. Prompting, context engineering, hallucination auditing and audit trail design. The 30 days above will get you started. The arc ahead will build the fluency.

And if you want to go further right now, Anthropic’s AI Fluency certification is free, model-agnostic, and worth every minute: anthropic.skilljar.com/ai-fluency-framework-foundations

Thank you for being here,
Jizel

Sources

• FDA Warning Letter 320-26-58, Purolea Cosmetics Lab, issued April 2, 2026: LINK

• 21 CFR 211.22(c): Quality unit responsibilities for approving procedures and specifications

• 21 CFR 211.100: Written procedures for production and process control

• Anthropic AI Fluency Certification: anthropic.skilljar.com/ai-fluency-framework-foundations

Leave a comment

Subscribe now

Last week I named seven emerging biotech ops roles being shaped by the AI lag arriving from tech: Director of Clinical AI Operations, Director of Scientific Intelligence (Medical Affairs), Director of Regulatory AI Strategy, Head of Digital Quality, Director of AI-Enabled External Manufacturing, Director of AI-Enabled Supply Operations, and MSAT AI Operations Lead. Each follows the same pattern. The senior role stays. The second and third roles backstopping it get absorbed by agentic workflows. The judgment remains. The structured work around it does not.

That is the function-by-function picture. This week I want to talk about what sits above it.

The Role Nobody Has Hired Yet

There is one role no clinical-stage biotech I’ve seen has hired yet but every one will need: Head of AI-Enabled Operations.

Pfizer appointed Berta Rodriguez-Hervas as its Chief AI and Analytics Officer in 2024, after senior AI roles at Stellantis, Nvidia, and Tesla, joining Pfizer’s digital leadership team. AstraZeneca established an Executive Director, AI for Clinical Intelligence and Evidence, a senior enterprise role responsible for AI-driven evidence capabilities across the medicine lifecycle. Big Pharma is roughly 18 months ahead of clinical-stage on this curve, which is exactly the lag I have been pointing at.

The clinical-stage version of this role will not carry C-suite weight on day one, and it will not need to. It will sit in COO org or report directly to the CEO, with a budget line and cross-functional authority over every agentic workflow touching GxP. The title will vary. The mandate will not.

What This Role Actually Owns

The scope is concrete, and it is bigger than most senior leaders realize until they map it.

Own the enterprise AI roadmap and the prioritization framework that decides which function gets agentic capacity next. Own the validation playbook that translates FDA’s January 2025 draft guidance on AI in regulatory decision-making and the January 14, 2026 FDA-EMA Guiding Principles of Good AI Practice in Drug Development into deployable SOPs every function reuses. Own the central inventory of AI tools, models, and agents in production, the change control on each, and the inspection-readiness package that holds it all together. Sit on the joint steering committee with Quality and Regulatory leadership. Brief the board on AI risk and governance posture.

This is the role that turns seven function-level AI bets into one defensible enterprise capability. Without it, every function reinvents validation, every CMO conversation duplicates governance, and the inspection trail fragments. With it, the company’s AI posture becomes a competitive moat instead of a compliance liability.

Why Internal-First

Here is what makes this role different from the other seven on the function list. It will most likely be filled internally first.

The job requires two things external AI hires from Big Tech do not have. The first is regulated-environment fluency: GxP, change control, validation gravity, FDA inspection logic, how a Form 483 actually gets written and how an AI-assisted decision gets defended in front of an auditor. That is a multi-year build, not a course. The second is trust capital. The Head of AI-Enabled Operations sits on a steering committee with Quality and Regulatory leadership and has cross-functional authority over agentic workflows. Quality and Regulatory leaders do not extend that trust to someone who arrived last quarter from a self-driving car company, regardless of how sophisticated the model architecture they built.

External AI hires can run the technical layer. They cannot, in 2026, run the governance layer in a biotech preparing for a BLA.

The candidate pool is the senior ops leaders already inside the company. The ones who have been quietly building real agents inside their function. The ones who have documented validation as part of the artifact, not the footnote. The ones who have written one-page proposals for roles that do not exist yet. The ones who have read the FDA AI guidance closely enough to ask sharp questions about it in cross-functional meetings.

That is the talent pipeline. It is not posted anywhere. It is being built right now, mostly invisibly, by the people doing the work I described in last week’s issue.

The Strategic Implication

Two things follow.

For senior ops leaders, this is the role to position toward, even before the title exists. The path is not a certification. The path is operational evidence inside your current function plus regulatory literacy plus a written articulation of what the enterprise needs. The leaders who get the role will be the ones who described it before HR wrote the JD.

For the company, the implication is harder. Most clinical-stage biotechs are about 24 months from realizing they need this role and 36 months from being able to fill it. The runway is short. The companies that move first will have a coordinated AI posture by the time their BLAs hit the FDA. The companies that move late will have seven function-level AI strategies that do not talk to each other, validation evidence scattered across functions, and a CMO partnership stack the agency has to piece together at inspection.

That is not a hypothetical risk. That is the cost of treating AI strategy as a function-by-function initiative instead of an enterprise architecture.

What I Am Watching For

Last week I said the signal to watch is the first wave of job postings with the new function-specific titles. The signal for this role is different. It will not show up in job postings first. It will show up in three quieter places.

First, a senior ops leader at a clinical-stage biotech being given an “AI initiative” carve-out from their existing role, with a small budget and cross-functional convening authority. That is the role taking shape before the org chart catches up.

Second, a new line on an org chart sitting in COO org or reporting to the CEO, even without an external posting. Most of these roles will be promotions, not hires.

Third, the first Series C or pre-IPO biotech that names a Head of AI-Enabled Operations on their executive team slide. Once that happens once, the rest will follow in the same way the Chief Digital Officer title cascaded through pharma between 2018 and 2022.

The window is 18 to 24 months. The work is operational evidence, regulatory literacy, and the articulation of an enterprise AI posture before anyone asks for it. The reward is the role that sits above the seven I named last week.

Thank you for being here,

Jizel

Last week’s issue named the seven function-specific roles emerging across biotech ops. If you missed it, read it here before this one fully clicks. In the coming weeks: a working playbook for the operational evidence I keep pointing at. What it actually looks like to build the first real agent inside your function, document it for inspection-readiness, and turn it into the artifact a senior leader can repeat.

Leave a comment

I watch open AI job postings at Big Tech the way an ops leader reads FDA guidance. The roles being cut tell you what’s already gone. The roles staying open tell you what’s coming.

Tech layoffs ran over 80,000 in Q1 2026 alone, with AI a major driver. Customer support, QA, content moderation, mid-level engineering, layers of middle coordination. Those are the roles being compressed. Meanwhile, Google, Amazon, Meta, and Microsoft are spending $725 billion on AI infrastructure this year, up 77% year over year. They have roughly 275,000 AI-related roles sitting open. AI governance, AI operations, machine learning, domain-specific AI auditors. Those they can’t fill fast enough.

The workers being eliminated are not the workers being hired. That distinction is the entire story.

Why Biotech Ops Feels Insulated Right Now

If you lead operations at a clinical-stage biotech, the picture looks different. CMC and external manufacturing talent is structurally scarce. Regulatory expertise is harder to find than three years ago. Programs heading toward 2027 and 2028 BLA filings are competing for the same small pool of GMP and tech transfer experience.

The growth cycle is here. The talent is scarce. Senior ops roles feel safe.

That read is correct about today. It is wrong about the next 24 months. Gartner forecasts 40% of enterprise applications will have task-specific AI agents embedded by end-2026, up from less than 5% in 2025. Major pharma is already piloting agentic workflows in adjacent areas: temperature-critical logistics, predictive quality monitoring, regulatory document drafting, first-pass deviation triage. The 18 to 24 month lag is not an exemption. It is a window.

.

Subscribe now

The Roles That Disappear Around You

AI is not coming for the senior ops role. It is coming for the second and third roles that used to make the senior role possible.

The judgment calls are real. But they sit on top of a layer of structured work that agentic systems handle well, even when human-in-the-loop sign-off remains mandatory under GxP. The judgment remains. The validation gravity remains. The role around it gets lighter. And so does the headcount needed to do it.

The pattern repeats across every function in a clinical-stage biotech.

Clinical Operations

Agents already draft protocol synopses, summarize investigator queries, and pre-screen adverse event narratives. The senior leader stays. The coordinators who used to backstop her do not. AbbVie and AstraZeneca have already posted clinical AI and enterprise AI leadership roles.

Emerging role pattern: Clinical Operations Manager → Director of Clinical AI Operations.
Scope: Governs AI-generated study artifacts from synopsis through ICF and CSR. Aligns the function’s AI roadmap to the broader R&D AI framework.

Medical Affairs

AI is reshaping medical information response, MSL field intelligence synthesis, and publication planning. The Director who interprets real-world evidence faster than peers retains decision rights. The associates who compiled the inputs do not.

Emerging role pattern: Medical Affairs Director → Director of Scientific Intelligence (or AI-Medical Strategy Lead).
Scope: Owns the evidence-to-strategy pipeline. Fact-checks AI-generated content against label language. Governs how the function uses AI responsibly.

Regulatory and Quality

The same pattern plays out across submission management, deviation triage, and trend reporting. The Director who can defend an AI-assisted decision in front of an auditor becomes more valuable. The reviewers who used to do the first pass become harder to justify. FDA’s January 2025 AI guidance and the 2026 FDA-EMA joint principles both emphasize context of use, lifecycle management, and human oversight.

Emerging role pattern: Regulatory Affairs Director → Director of Regulatory AI Strategy. QA Director → Head of Digital Quality.
Scope: Owns the validation and monitoring logic for AI-assisted submissions and decisions. Defends those decisions in front of auditors and regulators.

External Manufacturing

The role is shifting from manual coordination to orchestrating an AI-assisted oversight stack. CMO correspondence triage, capacity scenario modeling, deviation pre-classification, tech transfer document compilation. All running under a validated change-control trail.

Emerging role pattern: External Manufacturing Lead → Director of AI-Enabled External Manufacturing.
Scope: Owns the CMO oversight system. Decides what gets escalated to humans. Makes AI outputs legible enough to survive audit.

Supply Chain

The control tower is becoming the operating model. AI-driven demand sensing, predictive risk alerts, and scenario planning are moving from pilot to production. The important work is making those systems auditable, trusted, and tied to ERP and change control.

Emerging role pattern: Supply Chain Director → Director of AI-Enabled Supply Operations.
Scope: Owns the control tower stack. Translates AI-generated alerts into operator decisions. Closes the trust gap by making the reasoning visible.

MSAT

AI is reshaping tech transfer, comparability, and PPQ support more than it replaces the function outright. The function least exposed to disappearance. Most exposed to redefinition.

Emerging role pattern: MSAT Senior Director → MSAT AI Operations Lead.
Scope: Owns AI-augmented tech transfer. Designs the model maintenance plan. Connects AI-supported process work to lifecycle documentation and submission readiness.

I’m watching this play out at half a dozen clinical-stage biotechs. The same compression every time. And the same emergence. Industry pattern, not a single company’s plan.

This is what happened in tech. The senior engineer kept her job. The three mid-level engineers who used to support her did not. Then her budget conversations got harder, because her team was now “more efficient.”

Subscribe now

What This Means for Your Career, Specifically

Two things follow from this pattern, and senior ops leaders need to hold both at once.

Your role is more durable than you think, and more exposed than you realize. The judgment is durable. The team structure that made the judgment possible is not. When the second and third roles around you get absorbed, your delivery expectations stay the same. You become the person doing both the judgment work and the structured work that AI did not fully handle. That is not a promotion. That is a squeeze.

The premium is real, but only for evidence. Not credentials. PwC’s 2025 Global AI Jobs Barometer, drawn from nearly a billion job ads across six continents, found a 56% wage premium for workers with AI skills over peers in the same role, up from 25% the year before. That premium more than doubled in twelve months. It will land in biotech CMC, Quality, Regulatory, and external manufacturing too. But it lands for the people who can actually do the work.

There is a real difference between exposure and capability, and senior leaders can detect it in five minutes of conversation.

What Actually Compounds in This Window

Three things look like preparation but are not. Stacking certificates. Generic AI courses. LinkedIn rebrands.

What compounds is harder and quieter:

• Build one real agent that automates one real pain point in your actual workflow. Not a demo. Something a peer would use after watching you run it.

• Read FDA’s evolving AI/ML guidance closely enough to ask sharp questions about validation in agentic systems. The person who can frame the regulatory exposure cleanly becomes the de facto governance lead.

• Document the outcome of an AI-assisted process change in language an executive can repeat. Treat the validation work as part of the artifact, not a footnote. The change control trail is what makes the win defensible.

• Write a one-page proposal for the role that does not exist on your org chart yet but should. You just read the titles. The people who get them are the ones who described the role before it existed.

The signal worth watching for is not layoffs. It is the first wave of these new titles appearing in biotech job postings. When they show up in earnest, the lag has closed and the market has already moved.

The window is open now. The question is whether you build evidence inside it, or wait until the postings tell you it closed.

Thanks for being here,

Jizel

Next week: the eighth role. The one above all seven. The role no clinical-stage biotech has hired yet, but every one will need. Why it will most likely be filled internally first, and what that means if you are the candidate pool.

Leave a comment

Not “which model should we use?” Not “is AI ready for biotech and pharma?” The real question is: how do you deploy AI agents in an environment where a single wrong output could trigger a 483 observation, delay a filing, or compromise patient safety?

Three things happened in Q1 2026 that shifted the conversation.

The regulatory ground shifted

On January 14, the FDA and EMA jointly released ten Guiding Principles for Good AI Practice in Drug Development. A coordinated, transatlantic framework covering AI use across the entire medicines lifecycle.

The principles emphasize a risk-based approach, strong data governance, human oversight, multidisciplinary expertise, and lifecycle management. They’re not legally binding. But if you’ve worked in biotech/ pharma for more than a year, you know that “guiding principles” from FDA/EMA today become inspection expectations tomorrow.

Around the same time, the EU’s draft Annex 22 landed. First structured regulatory framework for AI in GMP manufacturing. The operational message: static, deterministic models are acceptable in critical processes. Adaptive and learning models, the ones that modify their behavior during use, are restricted to non-critical applications with documented human oversight.

And the FDA’s guidance pipeline signals additional AI/ML manufacturing-specific guidance is in development.

The regulators aren’t saying “don’t use AI.” They’re saying “use it, but here’s how we expect you to govern it.” The companies that figure out governance now have a window. Everyone else will be writing SOPs under pressure in eighteen months.

The bottleneck moved from installation to specification

Here’s the part nobody talks about at conferences. Deploying an AI agent takes minutes. Configuring it to actually do useful work in a regulated environment takes weeks. The bottleneck has shifted from “can AI do this?” to “can we tell AI exactly what we need it to do, with enough context to do it well?”

Nate B. Jones calls this the “installation vs. specification gap.” The technology works. What’s missing is the structured knowledge that agents need to operate with domain expertise.

In biotech, that knowledge is painfully specific: which FDA precedents apply to your therapeutic area, how your CDMO communicates about deviations versus routine updates, what your quality team’s risk tolerance looks like for different parameter excursions, which cross-references in your CMC package need to stay consistent across dozens of interconnected documents.

This expertise lives inside the heads of senior team members. It’s been compressed through years of experience into intuitive judgment. The MSAT director doesn’t think through why one deviation reads as routine versus systemic. She just knows.

That compressed knowledge is exactly what AI agents need. And exactly what most AI deployments fail to capture.

Building the Context Library

The solution isn’t better prompts. It’s better knowledge architecture.

I came to this the long way. My first serious agent build wasn’t a workflow tool. It was a writing assistant and thinking partner for this newsletter. The model was capable. The prompts were polished. The output read like generic AI sludge, because the agent had no idea what “sounds like Jizel” actually meant or which arguments I’d already made in earlier issues. The fix wasn’t a better prompt. It was weeks of structured work building out my own Context Library, a “second brain” for this newsletter, until the agent had voice patterns and prior arguments to draw from.

That’s when the architecture clicked.

The same pattern holds in regulated biotech, with much higher stakes. None of that compressed expertise lives in prompts, however polished. It only works once the judgment is structured into a Context Library the agent can draw from.

Before you deploy a single workflow agent, you need to create a Context Library: a structured set of documents that function as the operating system for every agent in your organization.

The architecture follows an approach emerging across enterprise AI deployments:

soul.md holds company mission, values, and GxP non-negotiables. The principles no agent should ever violate.

identity.md defines team roles, expertise domains, and escalation paths. Who owns which call.

knowledge.md is the core expertise layer: process rationale, deviation patterns, regulatory precedents, lessons learned. This is where tacit knowledge gets externalized.

preferences.md captures how leadership receives information: report formats, submission language standards, communication norms.

heartbeat.md tracks task cadences, review schedules, filing milestones. Agents use this to surface what matters when it matters.

The insight that changed my thinking: the first agent you deploy shouldn’t automate a workflow. It should be an Interviewer Agent. An AI whose sole purpose is to help your experts externalize their knowledge through structured elicitation. Ask the regulatory lead which precedent she’d defend in a Type C meeting, even if the agency pushed back. Ask the CMC lead what judgment lives only in his head. Ask the manufacturing director where his CDMO’s incentives are diverging from theirs. The questions are the point. They surface the reasoning that no SOP captures.

That interview process builds the Context Library. The Context Library is what makes every subsequent agent worth deploying.

There’s a second benefit worth naming. When a senior leader leaves, their judgment leaves with them. The Context Library is where the company keeps it. People move on. The library stays.

A risk tiering model for GxP environments

Not all AI use cases carry the same regulatory risk. The mistake most companies make is treating AI deployment as binary: either everything needs full computer system validation, or nothing does. Both extremes are wrong.

Start with your lowest-risk work. Non-GxP, human-reviewed outputs: communications tracking, meeting summaries, regulatory intelligence scanning, first-draft document generation. No CSV/CSA validation required. These can go live this quarter.

The middle tier is GxP-adjacent work where AI processes GMP data or supports quality decisions, but outputs require human approval at defined checkpoints. Batch record pre-screening, deviation triage suggestions, CAPA trending analysis. Standard software validation methodology applies.

The highest tier is GxP-critical: AI directly influencing product quality decisions or patient safety outcomes. Release testing automation, process control optimization, clinical supply forecasting with patient-level data. These require full computer system validation, documented performance qualification, and continuous monitoring. They should come last, after your organization has built confidence at the lower tiers.

The FDA’s own principles draw this line explicitly. AI that informs decisions versus AI that makes them. Start low. Build up.

The non-negotiable principle

Every framework I’ve seen that works in regulated environments comes back to one rule: AI assists, humans decide.

When an inspector asks “how do you ensure the quality of AI-generated CMC documentation?”, your answer is: “Every AI-generated draft is reviewed, edited, and formally approved by a qualified subject matter expert. The AI accelerates the drafting. The human ensures the quality. Here’s the audit trail.”

That answer works with every regulator in the world.

Where this is heading

For clinical-stage biotech, especially companies approaching major filing milestones, the math is straightforward. McKinsey’s 2025 analysis of 270 pharma workflows found agentic AI could free up 25 to 40 percent of an organization’s capacity. A lean team augmented with AI agents can operate with the throughput of a team twice its size. That’s not a technology story. That’s a capital efficiency story.

The companies building their Context Libraries now and deploying Tier 1 agents this quarter aren’t doing it because the technology is perfect. They’re doing it because the regulatory window is open, the cost pressure is real, and their competitors are already moving.

I don’t have a clean prediction for how fast this goes. The honest answer is that the organizations I’m watching are still figuring out the specification problem in real time. But the ones who are at least building the knowledge architecture are weeks ahead of the ones still debating whether to start.

That gap only gets wider.

Want the full Implementation Guide? I put together a longer-form playbook for clinical-stage biotech leaders who want the full architecture: the top five pilot use cases with timelines and team effort estimates, the GxP tier classification for every workflow, the Context Library template, and the cost model versus hiring equivalent FTEs. It’s FREE for subscribers.

Subscribe now

Which tier is your team starting with? And what’s holding you back from Tier 1 today? Hit reply. I read every one.

Thanks for being here,

Jizel

This article draws on publicly available regulatory guidance including the FDA/EMA Guiding Principles of Good AI Practice in Drug Development (January 2026), EU draft GMP Annex 22: Artificial Intelligence (July 2025), emerging best practices from Tiago Forte’s Second Brain methodology and Andrej Karpathy’s knowledge architecture concepts, and enterprise context engineering frameworks. No proprietary or company-specific information is referenced.

Leave a comment

It is real. It’s arriving in biotech operations faster than most expected. The leaders who understand what it actually is, not the hype version but the operational reality, won’t be playing catch-up when their organizations start asking questions.

So this issue is a plain-language decode. What agentic AI actually is. Where it fits in CMC right now. How to evaluate whether a workflow is safe for it. What FDA is already signaling. And what to do if your organization hasn’t started yet.

Every section comes back to one question: where does human judgment stay in the loop?

So What Is Agentic AI, Actually?

To place it properly, it helps to see where it sits in the broader progression of AI development.

(If you already know the difference between an LLM and an agentic system, skip to Where Agentic AI Actually Fits Across CMC Functions. That’s where this gets operational.)

At the base: LLMs and chatbots. ChatGPT, Claude, Copilot. Tools that respond to a single prompt. You ask, it answers. You paste in a document, it summarizes. The human drives every interaction. Most biotech teams are somewhere in this layer right now.

One level up: Agentic AI. This is the inflection point. Agentic AI does not wait to be asked. It can plan, sequence, and execute a series of tasks on its own, using tools and data across multiple systems, with minimal human input between steps.

Above that: AGI, or Artificial General Intelligence. A system that matches human cognitive ability across all domains. Worth understanding, not your operational concern today.

At the top: ASI, or Artificial Superintelligence. A system that exceeds human cognitive ability across all domains. Theoretical. Not your immediate problem.

Think of it this way. An LLM is like a highly capable research assistant sitting at your desk. You hand them a task, they complete it, they hand it back. Every task starts with you.

Agentic AI is like that same assistant except now they have a desk of their own, access to your filing system and tools, and a checklist they work through on their own initiative. You check in at defined points. They handle everything in between. The question is whether you’ve designed those check-in points well.

In CMC, here’s what that looks like concretely. Your team receives stability data from a CDMO. Instead of someone pulling the data, formatting it, running trend analysis, flagging deviations, and drafting a summary for team review, an agentic system does all of that in sequence without being prompted for each step. By the time your scientist opens their laptop, the groundwork is already done.

That’s agentic AI. And it’s already being piloted in regulatory writing, batch record review, supply chain monitoring, and tech transfer documentation across the industry.

Your organization is either already piloting agentic tools in operations, or someone is about to propose one. The question is whether you’re in that conversation early enough to shape it.

You don’t need to know how these systems work at a technical level. You need to know what questions to ask about them. That’s the judgment that protects your function.

The rest of this issue gives you the map for those questions.

Subscribe now

Where Agentic AI Actually Fits Across CMC Functions

This is the question most vendor conversations skip. Not “can AI do this?” but “which workflows in your function are actually suited for it?”

The honest answer is that agentic AI is not equally useful everywhere. It thrives in workflows that are structured and repeatable. Steps are known, inputs are defined, and a human can review the output at a clear checkpoint. It struggles anywhere the work requires reading context that lives outside the data.

That applies differently depending on where you sit. Here’s some examples of what I’m seeing across five functional areas.

Process Development & MSAT

This is where some of the strongest near-term fits live.

Tech transfer documentation is the clearest example. Think of it like a very detailed moving checklist. Every item needs to be packed, labeled, and confirmed before the truck leaves. An agent can systematically work through source documents, check each requirement against a predefined template, flag what’s missing, and route a summary to the responsible person for sign-off. The process is known. The exceptions are finite. The human checkpoint is preserved.

Process characterization involves pulling data from multiple experiments, identifying which parameters matter most, and spotting patterns across runs. Agents are well-suited to the data aggregation and pattern-flagging portions. Where they need human oversight: interpreting whether a parameter relationship is scientifically meaningful versus statistically coincidental. The agent finds the signal. The process engineers/ SMEs decide what it means.

Analytical method transfer is heavily document-intensive: protocols, raw data, statistical comparisons, summary reports. An agent can manage the document flow, run predefined equivalence calculations, and flag any results that fall outside acceptance criteria. The human role shifts from data wrangling to scientific judgment on the exceptions.

PPQ documentation is high-stakes, inspection-facing work. An agent can compile batch data, track completion status across required lots, check that each data point meets predefined acceptance criteria, and flag gaps for human resolution before the summary is finalized. It always hands the pen to a human before anything gets signed.

Quality

Quality workflows are where agentic AI gets interesting fast, because the data volume problem is acute and the stakes of missing a pattern are compliance risk.

Deviation investigation support is gaining traction. AI systems can scan batch records, environmental monitoring data, and equipment logs to surface early signals of nonconformance. They can draft investigation narratives using patterns from past deviations, suggest root cause hypotheses based on historical data, and flag similar prior cases so teams avoid repeating ineffective CAPAs. FDA inspection data consistently shows that investigation documentation remains one of the most-cited deficiency categories. The bottleneck isn’t analytical skill. It’s the time spent writing, formatting, and chasing data that already exists elsewhere in the record.

Environmental monitoring trend analysis. An agent can aggregate EM data across cleanroom suites, correlate it with equipment maintenance logs and production schedules, and flag emerging contamination signals before they become deviations. Published case studies describe pharmaceutical manufacturers seeing measurable reductions in environmental deviations and contamination-related CAPAs after deploying AI-enabled monitoring.

Batch record review is the high-volume, repetitive work that consumes QA bandwidth disproportionately. Agents can check records against acceptance criteria, flag incomplete entries, and surface exceptions for human review. The goal isn’t replacing the reviewer. It’s replacing the hours spent on data wrangling so the reviewer can focus on judgment calls.

Where Quality stays human: CAPA effectiveness verification, risk assessments that require cross-functional context, and anything that touches the “why” behind a quality decision rather than the “what.” Those still need a human with organizational knowledge.

One non-negotiable for quality teams evaluating any of these workflows: any agent that touches GMP data must produce an auditable trail that meets ALCOA+ data integrity standards. If the vendor can’t demonstrate attributable, legible, contemporaneous, original, and accurate record-keeping for the AI’s actions, the conversation is over. And any model update, whether retraining on new data or an architecture change, is a change control event in a GxP environment. Treat it like you’d treat any other validated system change. Most vendor conversations skip both of these points entirely. Don’t let yours.

Supply Chain

Supply chain is where the industry has the longest track record with AI, but most of it has been predictive analytics and demand forecasting. Agentic AI changes the model.

Demand-supply monitoring. Instead of running a forecast model and handing the output to a planner, an agentic system can monitor inventory levels, track supplier lead times, detect signals of disruption (raw material shortages, logistics delays, regulatory holds), and adjust reorder points or flag exceptions to a human planner in real time. The shift is from “here’s a dashboard” to “here’s what changed overnight and what you need to decide by noon.”

Supplier performance tracking. An agent can aggregate delivery data, quality metrics, and audit findings across your CDMO and raw material suppliers, surface trends that a quarterly business review would miss, and flag deteriorating performance before it becomes a supply event. For organizations managing multiple external manufacturing partners, this is the kind of work that should be continuous but rarely is because nobody has the bandwidth.

Cold chain and distribution compliance. For temperature-sensitive biologics, agents can monitor storage and transit conditions in real time, flag excursions against predefined protocols, and route alerts to the right decision-maker. The structured, rules-based nature of cold chain compliance makes it a strong fit.

Where supply chain stays human: CDMO relationship management, escalation decisions, contract negotiations, and sourcing strategy. Those depend on context, trust, and organizational priorities that no agent has access to.

Regulatory

Regulatory affairs is seeing some of the most active vendor development, and also some of the biggest gaps between what’s promised and what’s validated for GxP use.

Submission document preparation. Agents can already pull CMC data from multiple sources, check Module 3 content against ICH requirements, flag inconsistencies between sections, and generate first-draft narratives for human review. One published example from Amgen describes using an LLM-based tool to generate draft Quality Overall Summary (Module 2.3) content from Module 3 data, reducing initial draft time from roughly two weeks to under an hour. The human still owns the final review and sign-off.

Regulatory intelligence monitoring. An agent can track guidance updates, health authority Q&A documents, and inspection trends across FDA, EMA, and other agencies, and surface what’s relevant to your active submissions or marketed products. This is structured, high-volume scanning work that most regulatory teams do inconsistently because of bandwidth constraints.

eCTD gap analysis and pre-submission readiness checks. Before a submission is filed, an agent can evaluate whether each module meets the specific content requirements, flag documentation gaps, and identify cross-references that don’t resolve. This is the kind of systematic check that prevents Complete Response Letters, which remain one of the most expensive outcomes in pharmaceutical development.

Where regulatory stays human: strategy decisions on what to file and when, FDA meeting preparation, interpretation of advisory committee feedback, and any judgment call about benefit-risk framing. The agent organizes and checks. The regulatory strategist decides.

Cross-Functional: The Highest-Value Pattern

Across all of the above, one of the highest-value applications cuts across every function: cross-batch and cross-system data analysis. Spotting trends, flagging drift, surfacing anomalies before they become deviations. This is where agentic AI earns its keep quickly, because it’s doing work humans technically could do but in practice often don’t have bandwidth for consistently.

Here’s an example agentic workflow with human-in-the-loop checkpoints:

How to Evaluate Whether a Workflow Is Safe for This

You don’t need a technical background to make this judgment. You need one simple framework.

Think of it as a traffic light.

Green: Go. The workflow is structured and repeatable. The steps are defined. The inputs are consistent. A human reviews the output before anything moves forward. Start here. This is where the risk is lowest and the value is clearest.

Examples: Tech transfer document gap analysis. Stability data trend reports. Method transfer equivalence summaries. PPQ completion tracking. Batch record review against acceptance criteria. Supplier delivery and quality metric aggregation. eCTD module completeness checks. Environmental monitoring trend analysis.

Yellow: Proceed carefully. The workflow involves some judgment calls. Variable inputs, edge cases the agent may not have seen, or outputs that go directly into a regulatory submission. These workflows can work, but they require careful design of human review points before you deploy, not after.

Examples: Process characterization interpretation (the agent flags correlations, but deciding which are scientifically meaningful requires human judgment). Deviation investigation drafting (investigation narratives can become objective evidence during an inspection, so AI-generated drafts need rigorous human review before they enter the QMS). Cross-batch trend analysis with regulatory implications. Demand forecasting with novel product launches. Module 3 narrative generation for regulatory submission.

Red: Not yet. The workflow requires the kind of contextual judgment that lives outside the data. Organizational knowledge. Relationship dynamics. Regulatory strategy. Risk trade-offs that depend on the full program picture. Agentic AI doesn’t have access to that context. You do.

Examples: Regulatory strategy and filing decisions. CDMO escalation calls. CAPA effectiveness verification requiring cross-functional context. Sourcing strategy and contract negotiations. Risk assessments requiring program-level judgment.

The single most useful question to ask about any proposed agentic workflow: “What happens when the agent encounters something it hasn’t seen before?” If the answer is “it routes to a human,” that’s a sign of a well-designed system. If the answer is unclear, the system isn’t ready for GxP deployment.

What FDA Is Already Saying About This

The FDA has been watching this space closely, and they’ve started drawing lines.

In January 2025, FDA published a draft guidance specifically addressing AI use in drug and biological product development. Earlier this year, FDA and the European Medicines Agency jointly released ten guiding principles for good AI practice in drug development, covering manufacturing explicitly. These are not final rules. But they are a clear signal of where regulatory expectations are heading, and in a GxP environment, signals from regulators are worth reading before the guidance is finalized.

What FDA is asking about when they see AI in a CMC submission:

Think of it the way an experienced auditor thinks about any new process your team introduces. They want to know: what is this doing, who decided it was appropriate for this use, how was it tested, and who is accountable if it produces an error?

The FDA’s framework translates that mindset into four practical questions for any AI tool used in a regulated context. Before we get into them, one sentence that reframes everything: The agent doesn’t sign. You do. That’s not a limitation. It’s the reason your judgment becomes more valuable, not less, as these systems scale.

First: What is the specific context of use? FDA calls this the COU. Is this agent flagging stability data for human review, or generating text for a Module 3 section? Those are different risk levels, and FDA expects validation rigor to match.

Second: How was the model’s risk assessed? Two dimensions: how much the AI output influences the final decision, and what happens to the patient or product if the output is wrong. Know where your pilot sits on that scale before the inspector asks.

Third: What validation evidence exists? This is where most early-stage AI pilots come up short. FDA is asking whether the model has been tested on data that represents the conditions it will operate in. For CMC, that means your actual batch data ranges, your specific document formats, your CDMO’s data outputs. Generic vendor validation packages rarely satisfy this.

Fourth: Who owns the output? Most organizations haven’t answered this before pilots go live. If an agentic system generates a process characterization summary and that summary contains an error that reaches a regulatory submission, a qualified human reviewed and approved it. Period.

One additional signal worth flagging from the January 2026 joint FDA-EMA principles: they explicitly include manufacturing in scope, and they name multidisciplinary expertise as a requirement. That phrase is not accidental. Regulators are signaling that AI governance in manufacturing is not solely an IT or data science function. Operations leaders who understand GxP requirements, process risk, and regulatory accountability are part of what good AI practice looks like.

There's a related risk that current FDA guidance hasn't fully resolved: cybersecurity. When an agentic system has access to your QMS, your batch records, and your CDMO data feeds, the attack surface expands in ways most vendor conversations don't acknowledge. Prompt injection, unauthorized data manipulation, compromised supplier data pipelines. These aren't theoretical in a GxP environment. And in biotech, the stakes go beyond data integrity. A corrupted batch record isn't just a compliance event. Depending on where it sits in your process, it's a product quality event with direct patient safety implications. That's a different risk category entirely, and one I'll be going deeper on in a future issue.

For context on the pace of change: CDER alone reviewed more than 500 drug and biological product submissions with AI components between 2016 and 2023. That's before the current wave of generative AI tools even entered the market. The regulatory bar for AI in a GxP manufacturing context is going to be higher than what most current vendor pilots are built to meet. Understanding that now puts you in a very different position than understanding it during an inspection.

Where Your Team Is Probably Starting (And What’s Actually Ready)

You don’t need to build anything. You don’t need to hire a data scientist. Whether your team has been experimenting for months or hasn’t started yet, here’s what’s practical right now and what actually requires infrastructure versus what you can test tomorrow.

Level 1: AI assistants with your documents. Tools like ChatGPT, Claude, or Microsoft Copilot can already handle significant document-heavy CMC work when given the right structure. Here’s what this looks like on a Monday morning: you upload your tech transfer checklist and 14 source documents. Twenty minutes later you have a gap report that flags three missing validation summaries and an expired supplier qualification. That report would have taken your coordinator atleast two days to compile manually. This isn’t agentic AI yet. You’re still driving every step. But it builds the muscle for what comes next, and it’s available today with no infrastructure investment. (Remember to always check your company’s AI policy first)

Level 2: Workflow automation with AI capabilities. Tools like Claude Cowork, Power Automate, or Veeva Vault’s built-in AI features allow you to start connecting document workflows to AI-assisted review steps. These are closer to agentic. The system follows a defined sequence across steps, still within guardrails your team controls. Appropriate for document management, completeness tracking, and review routing.

Level 3: Purpose-built CMC platforms. Custom-built internal platforms or CMC-specific vendor solutions validated for your environment. These are where purpose-built agentic workflows for process characterization, PPQ documentation, and regulatory submissions will live. Evaluate them with your GxP validation requirements front and center before any pilot.

One honest note on tools: the market is moving fast and vendor claims are running well ahead of validated GxP capability. Before any tool evaluation, agree internally on two things. 1.) What validation evidence you’ll require. 2.) And who owns the accountability if the system produces an error in a regulated document. Those two questions will filter out more vendors than any feature comparison.

What to Do With This Right Now

You don’t need to launch a pilot this quarter. You need to build a point of view before someone asks you for one.

Get curious about what’s already happening. Is your organization piloting anything agentic in operations, quality, or regulatory? If you don’t know, find out. The pilots that matter are often the quiet ones.

Pick one Green-light workflow in your function. (Always check your company AI policy first) If you’re in process dev or MSAT, tech transfer documentation gap analysis is a natural starting point. If you’re in quality, batch record review or EM trend analysis. If you’re in supply chain, supplier performance aggregation. If you’re in regulatory, eCTD completeness checks. Map out the steps, identify the human review checkpoint, and ask yourself whether Level 1 tools could handle the routine portions today. You don’t have to deploy anything. You just need to understand the territory before someone else maps it for you.

Know your non-negotiables before the vendor conversation. Validation evidence. Audit trail. Human accountability at decision points. These three requirements should be in the room before any agentic AI vendor gets a meeting with your team.

Read the January 2025 FDA draft guidance, even just the summary sections. You don’t need to parse every regulatory detail. You need to understand what FDA is asking about when they see AI outputs in a CMC submission: context of use, risk level, validation evidence, and human accountability. A VP who can speak to those four questions is a different conversation partner than one who can’t.

Who Owns AI Governance in Manufacturing? Most organizations haven’t resolved who governs agentic AI deployment in manufacturing. IT, Data Science, and Ops are all logical stakeholders, but GxP validation requirements, process risk, and inspection readiness aren’t evenly distributed across those groups. If you’re in operations or quality, that’s your contribution to the conversation. Better to be in it early than inherit a framework someone else built.

Thank you for being here,

Jizel

He’s not wrong about the pace. But he’s missing something critical about where you work.

Matt Shumer, CEO of HyperWrite, wrote that he now describes what he wants built in plain English, walks away for four hours, and comes back to finished software. No corrections needed. He called it judgment. Taste. The thing everyone said AI would never have.

If you’re in biotech operations, you probably read that and felt two things at once: a jolt of urgency and a wave of “but that’s not how any of this works in my field.”

Both reactions are correct. And the gap between them is exactly where your career strategy lives right now.

The Gap Everyone’s Talking About vs. The Gap That Actually Matters

Shumer’s AI skills gap is behavioral. Tech professionals who adopted AI early are pulling ahead of those who didn’t. His advice: spend an hour a day experimenting. Sign up for the paid tools. Push AI into your actual work. The window is closing.

Fair enough. But that framing assumes something fundamental: that when you want to try something new, you can simply implement at work.

In biotech operations, the AI skills gap isn’t primarily behavioral. It’s structural.

Consider what “just try AI in your workflow” actually means when your workflow touches GMP manufacturing, regulatory submissions, or validated systems. You can’t feed a batch record into ChatGPT and see what happens. You can’t let an AI agent iterate on a tech transfer protocol the way Shumer lets one iterate on app code. Your change control process alone takes longer than most tech companies’ entire product cycles.

This isn’t resistance. It’s reality. And it creates a fundamentally different kind of AI skills gap than the one dominating headlines right now.

Subscribe now

Three Structural Barriers Tech Doesn’t Have

The Regulatory Architecture

In tech, you ship and iterate. In GxP environments, you validate, document, qualify, and then maintain that qualified state through rigorous change control. In January 2025, the FDA published its first draft guidance on AI in drug development, with a 7-step credibility assessment framework. Then just last month (Jan 2026), the FDA and EMA jointly released 10 guiding principles for AI practice across the entire medicines lifecycle, from research through manufacturing and safety monitoring. The regulatory landscape is getting clearer. It’s also getting more layered. The gap isn’t “we should use AI.” It’s “we literally cannot deploy it the same way.”

The Enterprise Adoption Cycle

Multiple enterprise AI studies suggest that a large majority of pilots fail to deliver measurable impact, even after tens of billions in investment. The primary cause wasn’t model capability. It was integration failure. Most organizations need 12-18 months just to evaluate, procure, and approve new digital tools.

To put that timeline in perspective: Anthropic released six major updates to Claude between May 2025 and February 2026. Nine months. The model you start evaluating can easily be multiple generations behind by the time procurement finishes. You’re not behind because you’re slow. You’re behind because the system was designed for a fundamentally different pace of change.

The Accountability Asymmetry

When Shumer’s AI writes code that doesn’t work, he iterates. In a GxP environment, when an AI system generates an incorrect deviation summary or misclassifies a CAPA, it could lead to a potential regulatory finding down the road or any potential patient safety impacts. And every AI tool connected to regulated systems is also a new cybersecurity attack surface, such as prompt injection, data exfiltration, adversarial manipulation.

In tech, that’s a breach notification. In biotech, a compromised system touching batch records, quality data, or manufacturing processes isn’t just a data leak. It’s a patient safety risk. The stakes aren’t symmetrical. And that asymmetry shapes every adoption decision, from which use cases to pilot to who owns the risk when something goes wrong.

Why This Makes Your Judgment More Valuable, Not Less

Here’s where the Shumer narrative breaks down for regulated industries like biotech. He suggests that AI is developing something that feels like judgment, and that nothing done on a screen is safe in the medium term.

That may be true for environments where output quality is the only variable. But here’s what I want you to hold onto:

the structural gap that makes biotech AI adoption harder is the same gap that makes your judgment more strategically valuable.

In biotech operations, good judgment isn’t just about getting the right answer. It’s about knowing what questions can’t be delegated, understanding the regulatory context around a decision, and navigating the organizational dynamics that determine whether a good answer actually gets implemented.

AI can generate a deviation investigation summary. It cannot accurately determine whether that summary will survive an FDA inspector’s follow-up questions about your rationale. It cannot assess whether your CDMO partner’s corrective action plan is genuinely addressing root cause or just checking a box. It cannot navigate the cross-functional tension between manufacturing timelines and quality holds, nor can it negotiate a highly complicated CDMO contract with various stakeholder timelines and conflicting strategies.

That’s a structural feature of how regulated industries work. And it gives your judgment real strategic weight right now.

This structural gap is a moat, not a wall. It buys us time. It does not make us immune.

The pace of AI capability improvement is unprecedented, and it would be naive to assume that today’s limitations are permanent ones.

Experimenting with AI tools for my own productivity has left me with the same reaction Shumer described: genuine awe. The capability gains are real. And they’re compounding.

Sophisticated agent orchestration, multi-step reasoning systems, and models that can simulate regulatory risk or navigate complex decision trees are not distant possibilities. They’re the next two or three iterations on a curve that is accelerating, not flattening.

The things we say AI “can’t do” in regulated environments today? Some of them will be table stakes in 36 months. Maybe sooner. The difference between biotech and unregulated industries isn’t that we’re protected from this wave. It’s that the structural complexity of our environment gives us a longer runway to prepare. Longer than tech. But not infinite. And not as long as most people in our industry seem to think.

Which makes how we use this time the most consequential career decision most of us will make in the next few years.

The professionals who recognize this have a strategic advantage right now. Because they understand that the skills gap in biotech is really a judgment gap — and they’re using the time the structural moat provides to close it deliberately. The question isn’t whether AI can do the cognitive work. It’s who retains the decision rights over where AI judgment is trusted and where it isn’t.

That’s leadership work. And it’s the work that AI adoption actually accelerates demand for, not replaces.

From the Field

I’ll be honest: I felt the same two-reaction split reading Shumer’s essay. The urgency is real. I see it firsthand as I try to keep up with the AI news and tools for my personal learning. The pace of improvement is genuinely staggering. By late 2025, some orchestrated AI systems could handle tasks that previously demanded hours of expert effort, at least in constrained domains.

But I’ve also seen what happens when other biotech organizations try to import tech-sector urgency without adapting it to regulated-industry reality. The result isn’t transformation. It’s a dozen stalled pilots, a frustrated quality team, and a narrative that “AI doesn’t work here.”

The biotech teams making real progress aren’t the ones panicking about the headlines. They’re the ones quietly building AI fluency in parallel, in the spaces where experimentation is possible, while developing governance frameworks for the spaces where it isn’t yet.

That’s not complacency. That’s strategy. And it’s the kind of strategy that compound interest rewards. As long as you actually start. And I’d encourage everyone, friend and colleagues, reading this to start now.

Your Move

Don’t panic. Reframe the gap. The skills gap in biotech isn’t about whether you’re using AI. It’s about whether you’re developing the judgment to guide how AI gets used in your domain. That’s very different from learning prompts or experimenting with tools. I want to dig into this more in the next month: how to strengthen judgment when more of the routine thinking work becomes automated.

Build AI fluency where you can. Like Shumer, I recommend protecting time everyday to learn and use any AI/LLM of your choice (Claude, Gemini, ChatGPT, Microsoft Copilot, etc.). For personal productivity, literature review, competitive intelligence, meeting prep, draft communications, research, etc. There are countless tutorials and courses available online.

Build the muscle memory and habit of using these tools every day. (My personal favorite is Google’s NotebookLM for learning new things fast. It’s amazing and highly recommend trying it out!)

Think M-shaped. Deep operational expertise is your first pillar. Digital fluency and AI governance is your second. The professionals who integrate both will retain decision authority. I wrote about this briefly in here: LINK

Own the governance conversation. The January 2026 FDA-EMA guiding principles mean this is accelerating. Someone in your organization will define how AI gets deployed in operations. If that person isn’t from operations, the frameworks won’t reflect operational reality. Be the person who brings both AI fluency and regulatory context to that table.

If you lead people, talk to your team about this. They’re reading these headlines too. They need someone to translate the urgency into context that accounts for their actual operating environment. Be that person. It builds trust and it might surface adoption opportunities you might not see alone.

The viral essay is right about one thing: the window for getting ahead of this is real, and it’s closing. But in biotech, “getting ahead” doesn’t look like walking away from your laptop for four hours and hoping the AI figured it out.

It looks like being the person in the room who understands both the technology’s capability and the system’s constraints.

This constraint bought us a little time. Use it wisely. The moat won’t last forever.

Next issue: Why AI makes judgment more valuable, not less. If Shumer’s essay landed differently for you, hit reply and tell me how. I’d love to hear your perspectives.

All content reflects publicly available information and general industry patterns. Views are purely my own.

Thank you for being here,

Jizel

Leave a comment

In March 2025 (11 months ago), Google's ex-CEO, Eric Schmidt, put $430M behind that exact thesis. Last week, his company Hologen moved to raise $150M in Series A funding to scale that exact platform. He’s doubling down.

What just happened, why it matters, and what this means for your career in biotech operations.

Subscribe now

Schmidt Didn’t Fund Discovery AI. He Funded Manufacturing Excellence.

Schmidt didn’t fund another AI drug discovery platform. We’ve got dozens of those already.

He funded something different: “large medicine models” for late-stage trial optimization and manufacturing scale-up. Think ChatGPT, but for the stage where 45% of drugs fail despite passing earlier trials: Phase 3 trials and commercial manufacturing.

The deal structure tells the story:

• $200M upfront to MeiraGTx (gene therapy company)

• $230M committed to fully fund a Parkinson’s gene therapy through commercialization

• Equity stake in MeiraGTx’s manufacturing subsidiary

That last detail matters more than people realize.

This isn’t theoretical. MeiraGTx’s CEO said Hologen’s AI already “significantly de-risked” their Parkinson’s program by analyzing Phase 2 data and identifying disease-modifying changes that weren’t initially obvious.

Phase 3 started mid-2025. Last week's Series A announcement suggests they're confident enough to raise more capital before waiting for Phase 3 results.

Why Operations Just Became Your Competitive Advantage

Four months ago, I wrote about how COVID vaccines proved manufacturing readiness determines market winners. Not just clinical efficacy.

Pfizer produced 3 billion doses by end of 2021. Novavax lagged competitors by 8-12 months despite promising clinical data. Why? Manufacturing problems at multiple sites.

Schmidt’s investment says this pattern isn’t specific to COVID. It’s structural.

When discovery gets 70% faster but manufacturing timelines stay static, operations becomes the competitive advantage. Not science. Operations.

The smart money is now betting on operational excellence at scale, not just molecular innovation.

But here’s what nobody’s talking about: These AI models need manufacturing expertise to actually work.

You can’t de-risk a billion-dollar trial if you can’t consistently produce the drug at scale. You can’t predict patient responses if your manufacturing process introduces variability that confounds the clinical signal.

This is why Schmidt invested in the manufacturing subsidiary. The AI and the manufacturing capabilities are inseparable.

The Career Positioning Move You Need to Make Now

If you’re in CMC, tech transfer, or manufacturing:

Your domain expertise just became more strategically valuable.

But domain expertise alone isn’t enough anymore. I’ve been watching this pattern across the industry. The professionals who survive layoffs and land promotions have something different.

They have what I call the “M-shaped” profile. Two pillars of strategic depth standing on a shared platform.

Here’s the structure:

The left pillar is your deep operational expertise. Supply chain forecasting, CDMO relationships, CMC documentation, tech transfer protocols. This is what got you here. It’s still critical. You already have this.

The platform is your cross-functional fluency. This is what lets your two pillars work together and multiply each other’s value. Can you translate manufacturing constraints for regulatory? Explain supply risks to finance in their language? Coordinate across three functions instead of just reporting status? This connector is what separates good from great.

The right pillar is your strategic complementary skill. This is what you build now. Pick one. Go deep. AI systems fluency. Advanced stakeholder management. Strategic communication. Digital transformation leadership. Make it strategic, not shallow.

Here’s what Schmidt is actually betting on:

Deep CMC process knowledge (left pillar) is still foundational. You need to understand tech transfer protocols, batch variability, regulatory requirements, manufacturing constraints.

But the professionals who can also fluently work WITH AI systems (right pillar) become exponentially more valuable. Not coding. Not data science. The ability to interpret AI-generated insights in the context of manufacturing reality. To know when AI recommendations make sense and when they violate fundamental process constraints. To translate between data scientists and operations teams.

Think about it. A CMC manager who deeply understands analytical method transfer AND can effectively leverage AI for predictive quality monitoring? That person becomes irreplaceable. They’re not just executing tech transfers. They’re using AI to identify potential failures before they happen, then applying their process knowledge to prevent them.

That’s the M-shape. Two pillars of deep expertise standing on cross-functional fluency. Schmidt just validated this structure is worth $580M.

The question isn’t “Will AI replace me?” It’s “Am I building the second pillar that makes my first pillar exponentially more valuable?”

Key Takeaways

Schmidt deployed $580M into manufacturing optimization (not Discovery AI) in less than a year ($430M in March 2025, $150M Series A in January 2026). This sustained investment validates that operations is now the strategic bottleneck in drug development.

“Large medicine models” require deep manufacturing expertise to work. AI without ops knowledge fails. The technology and the production capability are inseparable.

The M-shaped profile is what survives: Two pillars on a platform. Left pillar is deep operational expertise. Platform is cross-functional fluency. Right pillar is strategic complementary skill you build now. Not broad shallow skills. Two areas of deep expertise that multiply each other’s value.

Schmidt’s bet proves the integration is what matters. CMC experts who can leverage AI become exponentially more valuable. AI experts without domain knowledge can’t solve the actual problems.

The 6-12 month positioning window is closing faster than expected. Major capital is already moving. Companies are implementing, not just exploring.

FROM THE FIELD

Seeing Schmidt validate this thesis feels surreal.

When I published the AI-CMC Speed Mismatch in November, I wondered if I was overstating the urgency. Maybe CMC becoming the bottleneck was important but not that strategically significant?

Turns out Schmidt had already put $430M on it 8 months earlier. And last week, he doubled down with another $150M. The pattern is accelerating.

This isn’t vindication. It’s a reminder that the people in the trenches often see patterns before the headlines catch up. You’re living the manufacturing bottlenecks. You’re navigating the tech transfer delays. You know where the friction is.

The difference between you and Schmidt? He has $580M to deploy. You have your career to position.

Same insight. Different leverage.

Next issue: Breaking down the M-shaped framework for specific biotech operations roles. What supply chain managers, CMC leads, and program managers should focus on for their right pillar.

Your turn: What are you seeing at your company? Are manufacturing and CMC teams getting AI budgets, or is everything still going to discovery?

Hit reply and tell me. I read every response and I’m collecting data points for the next deep dive on where the smart money is actually flowing.

Missed my original analysis? Read: The AI-CMC Speed Mismatch in the archives.

Thank you for being here,

Jizel

Leave a comment

The professionals thriving in 2026 will be the ones who strengthened what AI can’t replace: systems thinking, judgment under pressure, cross-functional intuition, and the ability to see patterns in complexity.

Those skills aren’t built in a weekend workshop. They’re formed through intentional learning.

I’ve organized this professional development reading list for mid-career biotech operations professionals navigating AI transformation in life sciences. Twelve books organized around the four pillars of what I’m calling the M-shaped biotech leader, a framework for building career resilience I’ll unpack in detail come January.

Full transparency: I haven’t read all of these yet. This is what I’m planning to work through in 2026, and I’m sharing it now because maybe you’re building similar capabilities. If we’re both working on this, we might as well learn together.

For those who prefer audio: I’ve included podcast episodes for each category. Sometimes I need the book. Sometimes I need something for my daily walks or while I’m cooking. Both work.

Subscribe now

SYSTEMS THINKING & JUDGMENT

The AI-proof core

Books

Thinking in Systems — Donella Meadows

Why I’m reading it: I want to get better at seeing bottlenecks before they become crises. When a CDMO says “minor delay,” I want to understand the cascade that’s coming, not just react to it.

Sources of Power — Gary Klein

Why I’m reading it: I need to understand how experts actually make decisions under pressure. I’ve made enough judgment calls with incomplete data to know my process could be better. Klein’s research on naturalistic decision-making keeps coming up in conversations with people whose judgment I trust.

The Fifth Discipline — Peter Senge

Why I’m reading it: Leading cross-functional teams and influencing without authority is most of my job now. This is supposedly the definitive book on organizational learning and systems thinking for teams.

Audio Alternatives

“Thinking in Systems: A Primer” — The Knowledge Project (Shane Parrish)

60 min deep dive into Meadows’ framework. I’m starting here before committing to the full book.

“How Experts Really Think” — Hidden Brain (NPR)

50 min on Klein’s research with real-world examples from firefighters and military commanders. Directly applicable to high-stakes biotech decisions.

“Peter Senge on Learning Organizations” — MIT Sloan Management Review

25 min interview with Senge. Includes explanation of the “beer game” simulation that demonstrates systems thinking in action.

AI FLUENCY FOR NON-TECHNICAL PEOPLE

What I actually need to understand

Books

Prediction Machines — Agrawal, Gans & Goldfarb

Why I’m reading it: This frames AI as “a drop in the cost of prediction,” which makes way more sense to me than the hype. When someone says “AI will optimize our supply chain,” I want to know the right follow-up questions.

A Human’s Guide to Machine Intelligence — Kartik Hosanagar

Why I’m reading it: Best accessible introduction to how algorithms behave and misbehave. No coding required. Just enough to ask better questions when vendors pitch “revolutionary AI solutions.”

Deep Work — Cal Newport

Why I’m reading it: If AI makes shallow work cheap, my competitive edge is the ability to focus on complex problems while everyone else drowns in notifications. This book builds that capacity.

Audio Alternatives

“AI as Prediction Technology” — Freakonomics Radio

50 min interview with Ajay Agrawal (co-author of Prediction Machines). Highly accessible, uses real business examples to explain AI economics.

“AI, Machine Intelligence, and the Future of Work” — a16z Podcast

45 min with Hosanagar discussing practical AI implications for professionals. Cuts through hype with economic frameworks.

“Cal Newport on Deep Work” — Knowledge Project

90 min on focus in a distracted age. Includes specific strategies for knowledge workers. Long but worth it.

BIOTECH OPERATIONAL DEPTH

Sharpen cross-domain intuition

Books

The Goal — Eliyahu Goldratt

Why I’m re-reading it: Theory of Constraints and throughput. Every time I see a tech transfer stall because of one analytical method, I think about bottlenecks differently. Worth a refresh.

Range — David Epstein

Why I’m reading it: Makes the case for breadth over narrow specialization in complex environments. Since I’m explicitly trying to build M-shaped capabilities rather than narrow depth, this feels directly relevant.

Principles of Product Development Flow — Donald Reinertsen

Why I’m reading it (with caution): Supposedly THE book on flow in product development, directly applicable to CMC. Fair warning: everyone who’s recommended this has also said “it’s dense.” I’m prepared to skim sections.

Audio Alternatives

“Theory of Constraints in Manufacturing” — The Lean Blog Podcast

35 min on applying Theory of Constraints to operations. Includes pharmaceutical manufacturing examples.

“Why Generalists Triumph in a Specialized World” — The Jordan Harbinger Show

60 min interview with David Epstein about Range. Discusses why breadth beats depth in complex fields.

“Don Reinertsen on Product Development Flow” — Scrum Master Toolbox Podcast

45 min with Reinertsen explaining his framework. More accessible than the book as an introduction. I’m using this to decide if I want to commit to the full read.

HUMAN SKILLS & INNER DURABILITY

What AI will never touch

Books

The Art of Thinking Clearly — Rolf Dobelli

Why I’m reading it: Short chapters on cognitive biases. I want to make better decisions when my CDMO is telling me one thing, my data is saying another, and leadership wants an answer by EOD.

The Culture Code — Daniel Coyle

Why I’m reading it: Research-backed look at psychological safety and trust—the backbone of cross-functional biotech work. Some CDMO relationships thrive, others implode. I want to understand why and how to build the conditions for success.

Man’s Search for Meaning — Viktor Frankl

Why I’m reading it: For grounding. When I’m questioning whether my career still matters in an AI-transformed industry, I need perspective bigger than quarterly planning. This isn’t biotech-specific, but it’s timeless on resilience and purpose.

Audio Alternatives

“Daniel Coyle on The Culture Code” — WorkLife with Adam Grant

35 min with Coyle on building trust and safety in teams. Includes real examples from high-performing organizations. Practical for anyone leading cross-functional initiatives.

“Cognitive Biases in Decision Making” — Hidden Brain

50 min on how biases affect professional judgment. Features research from behavioral economists. Complements Dobelli’s book.

“Viktor Frankl’s Search for Meaning” — On Being (Krista Tippett)

50 min exploration of Frankl’s work and legacy. Includes audio clips from Frankl himself. Excellent companion to the book.

Leave a comment

BONUS: BIOTECH-SPECIFIC PODCASTS

What I’m adding to my regular rotation

The Long Run — Biotech podcast by Timmerman Report

Interviews with biotech leaders and entrepreneurs. Focus on career journeys, operational challenges, industry evolution. Real stories from people who’ve navigated biotech careers successfully.

a16z Bio — Andreessen Horowitz

Intersection of biology, technology, and business. Episodes on AI in drug discovery, manufacturing innovation, supply chain. Helps me see around corners on emerging trends.

Endpoints News Podcast

Weekly biotech news and analysis. Covers M&A, pipeline updates, industry trends. Keeps me current without doomscrolling LinkedIn.

IF YOU JUST WANT SOMETHING LIGHTER

Four Thousand Weeks — Oliver Burkeman

Grounding and profound about time, limits, and intentional living. Pairs well with end-of-year reflection. Won’t make you better at tech transfer, but might help you remember why you’re doing this.

Audio: “Oliver Burkeman on Embracing Your Limits” — Hurry Slowly (45 min)

KEY TAKEAWAYS: Your 2026 Reading Strategy

- Systems thinking separates reactive from strategic biotech leaders — Books like “Thinking in Systems” (Meadows) help you anticipate cascading failures when a CDMO signals “minor delays” (critical for tech transfer and supply chain management)

- AI fluency for non-technical professionals means understanding prediction economics — “Prediction Machines” frames AI as reducing the cost of prediction, giving you the right questions when vendors pitch “revolutionary solutions”

- Operational depth comes from mastering constraints and flow — “The Goal” (Goldratt) and “Principles of Product Development Flow” (Reinertsen) directly apply to CMC timelines and batch scheduling decisions

- Human skills like judgment under pressure remain AI-proof — Books on naturalistic decision-making (Klein’s “Sources of Power”), cognitive biases (Dobelli), and psychological safety (Coyle) build irreplaceable capabilities

- Audio learning works just as well as traditional reading — Every category includes podcast episodes (45-90 minutes) for commute-based professional development (test podcasts before committing to full books)

MY ACTUAL PLAN

I’m not reading all twelve books in January. That would be ridiculous.

Here’s what I’m actually doing:

Q1 2026: One book from Systems Thinking (probably Meadows), one from AI Fluency (probably Prediction Machines). Listening to podcast versions during commute to decide priorities.

Q2 2026: The operations depth books (Goal re-read, Range for sure, Reinertsen if the podcast doesn’t scare me off).

Q3-Q4 2026: The durability books when I need perspective more than tactics.

For audio learners: Start with the podcasts in whichever category feels most relevant. Use your commute this week. See which resonates, then decide if you want the full book.

The goal isn’t comprehensive reading. It’s building the durable capabilities these resources teach over the course of 2026.

Coming in January: The complete M-shaped biotech professional framework, the skill architecture that combines deep operational expertise with AI fluency and systems-level judgment. These twelve books lay the foundation. The January issue shows you how to build the structure.

What book or podcast would you add to this list? Hit reply and tell me what shaped your thinking in 2025.

Thanks for being here,

Jizel

Subscribe now

Now you’re in alignment meetings about alignment meetings. You’re updating trackers nobody reads. You’re explaining, again, why the timeline slipped when the analytical method transfer hit a wall that anyone with CMC experience could have predicted.

Your role has become heavier and thinner at the same time. More responsibility. Less clarity. Higher stakes. Fewer people to share the load.

If you feel that shift, you’re not imagining it. And if you’re a mid-career biotech operations professional (whether you’re in supply chain, CMC, program management, or manufacturing), this shift is reshaping what your role actually means.

This is the starting point for understanding the broader changes reshaping biotech operations, and it is why this series begins here.

What’s Causing the Mid-Career Squeeze in Biotech Operations?

The work that used to sit squarely in mid-career jobs is fragmenting. Some of it is being automated. Some is being absorbed by digital platforms. Some is drifting upward toward senior leadership. The rest is landing back on the same people who already hold the system together.

Underneath that fragmentation are forces most teams feel but rarely name. Rising operational complexity as companies work across more manufacturing sites. Digital systems that increase documentation volume before they reduce it. AI discussions that raise expectations faster than teams can build capability. Restructuring that freezes backfills but not deliverables. These pressures converge in the same place: the middle layer.

Industry reporting from outlets like BioPharma Dive and FierceBiotech highlights widespread layoffs, consolidation, and reduced management layers, which often leave remaining mid-level managers covering responsibilities from unfilled or eliminated roles.

This shows up in familiar ways across teams. A CMC group that lost two coordinators but still runs the same volume of tech transfer work. A supply chain team responsible for more complex CDMO networks with no additional headcount. In one case, a QA organization where review cycles stretched from 5 days to 12 days as documentation volume doubled, before digital quality systems came online. The gap between increased complexity and automation capability is where the squeeze shows up. A program manager juggling three cross-functional gaps created by a hiring freeze.

The specifics differ, but the sensation is the same. The middle is carrying more weight with fewer support beams.

Companies rarely announce this shift directly. You see it in the roles they backfill and the ones they quietly eliminate. You see it in who gets pulled into escalation meetings and who no longer does. You see it when leadership asks for an AI readiness assessment but hasn’t decided who owns the problem.

These aren’t signs of dysfunction. They’re signs of a system reorganizing itself around new capabilities.

Why Your Expertise Alone Isn’t Enough Anymore

Here’s what I keep hearing from CMC leads, QA managers, supply chain directors:

“I know my function. I’ve delivered results for a decade. But suddenly that’s not enough?”

The honest answer is this: your expertise is still necessary. But it’s no longer sufficient on its own.

The companies that are hiring right now aren’t looking for people who can do the job as it existed in 2019. They’re looking for people who can do the job as it’s becoming. Regulatory documentation that integrates with digital submission platforms. Deviation investigations informed by predictive quality signals. Demand planning that accounts for disruptions nobody used to model.

This doesn’t mean you need to become a data scientist. It means the scope of operational competence is expanding. Deep domain knowledge still matters enormously. A supply chain manager who understands CDMO capacity dynamics. A QA lead who can read a process deviation and know instantly what downstream risks it creates. A program manager who can sit in a room with manufacturing, regulatory, and clinical teams and actually translate between them.

That expertise is not replaceable. But it is most valuable when it is paired with curiosity about how the work is evolving.

LinkedIn’s 2024 Workplace Learning Report and skills analyses show a clear shift toward durable skills.

Employers increasingly prioritize problem solving, judgment, and communication over narrow technical credentials. Recent BioPhorum analyses highlight growing demand for roles that combine deep technical expertise with strategic coordination skills, rather than purely specialist positions.

The market is signaling what matters.

Which Skills Will Biotech Operations Professionals Need Most?

Even with all this change, the core of mid-career value in biotech operations has not shifted. The work that stabilizes programs still depends on human judgment, cross-functional translation, context-driven decision making, and the ability to manage relationships across CDMOs, internal labs, and regulatory partners. These capabilities are not close to being fully automated and are expected to remain heavily human-driven. They grow more valuable as the environment becomes harder to navigate.

Even in the middle of this compression, certain capabilities remain the backbone of biotech operations. They have not shifted, and AI is not close to replacing them.

Cross-functional orchestration when priorities collide. Judgment under ambiguity when data is incomplete. Reading relationship dynamics at a CDMO before a gap becomes a deviation. Translating analytical, regulatory, and manufacturing constraints into a decision leadership can act on. These are the skills that stabilize programs. They are also the skills that become more valuable as the environment becomes more complex.

Read my deep-dive into this topic here: 5 Skills Biotech Operations Teams Need That AI Can’t Replace

This is why the squeeze feels so personal. You are being asked to deliver both the work that cannot be automated and the work that used to have a support structure beneath it.

How Are Mid-Career Professionals Successfully Navigating This Shift?

The people navigating this well share a few patterns.

They’re selectively experimenting with AI. Not panicking about it, not ignoring it either. They’re finding one or two places where new tools intersect with problems they already understand. A CMC director exploring how digital platforms could streamline regulatory change controls. A supply chain lead testing forecasting tools that flag risk earlier.

They’re getting specific about problems they solve. Not “program management” but “cross-functional orchestration when three critical workstreams are competing for the same CDMO capacity.” Not “quality assurance” but “deviation trend analysis that actually prevents repeat failures.”

They’re protecting strategic work time even when tactical demands are relentless. Not perfectly. Not every week. But enough to stay visible as someone who thinks beyond their immediate workload.

They’re honest about what’s hard. The professionals who pretend everything is fine sound out of touch. The ones who name the difficulty and describe how they’re adapting come across as clear-eyed and resilient.

FROM THE FIELD

I’ve watched colleagues across the industry navigate this compression for the past eighteen months. Smart people. Experienced people. People whose instincts about tech transfers and vendor management are sharper than anything you’ll find in a playbook.

Some of them feel stuck anyway. Not because they’re doing anything wrong, but because the environment shifted faster than anyone expected.

The ones who seem steadiest aren’t necessarily the ones with the most impressive credentials. They’re the ones who decided to learn something new even when their plates were already full. They tightened their pitch about what they actually do. They stopped waiting for their organizations to figure out what skills matter and started figuring it out themselves.

What Can You Do This Week to Address the Squeeze?

Before Your Next Career Planning Session:

• Audit your last two weeks using the three work categories from Issue #1 (Automatable, Augmented, Irreplaceable). If you’re drowning in Automatable tasks, that’s the squeeze showing up in your calendar.

Building Career Resilience in Biotech: Navigating AI Disruption

• Write down three problems you’ve solved in the past year that required judgment and context, not just execution. The CDMO escalation you navigated, the cross-functional conflict you resolved. That’s your positioning story.

• Pick one AI tool, one use case, one experiment in your function (regulatory submission automation, predictive quality metrics, digital tech transfer platforms). Deep domain expertise plus AI fluency equals the T-shaped profile that survives.

• Talk to someone who changed roles in the past 6 months. What did they learn about what’s actually valued right now? First-hand pattern recognition beats speculation.

• Protect time for strategic work even when tactical demands are relentless. Not perfectly, not every week, but enough to stay visible as someone who thinks beyond their immediate workload.

Subscribe now

The mid-career squeeze is real. It’s uncomfortable. And it’s also where your future leverage begins.

This is the first in a six-part series on building career resilience in biotech operations. Next up: why organizations are stuck on AI adoption, and what that means for the people inside them.

What’s the squeeze looking like in your corner of the industry? Hit reply and tell me. I read every response.

Until next time,

Jizel

Leave a comment

But nobody’s talking about this: AI compressed discovery by 70%, but tech transfers still take 18-30 months. And a large share face major delays or quality problems.

CMC just became the rate-limiting step in drug development. Not clinical, not regulatory. For anyone managing supply chains, tech transfer, or manufacturing, this changes everything. This isn’t a temporary mismatch. It’s a structural shift that’s reshaping competitive advantage right now.

Why Should Biotech Operations Leaders Care About This Now?

I’ve been tracking this shift across multiple biotech teams over the past year. The companies winning the AI race aren’t the ones with the biggest discovery budgets.

They’re the ones who realized CMC is now their competitive advantage.

This matters because if you’re still optimizing clinical timelines while your tech transfers take two years, you’ve misidentified the bottleneck. And Q1 planning is happening now, which means the teams that get this right will be 12-18 months ahead of everyone else by mid-2026.

The window for proactive response is narrower than you think.

How Big Is the AI-CMC Speed Gap?

The data is unambiguous:

• AI discovery: 12-18 months (down from 3-6 years, a 60-75% reduction)

• Tech transfers: Still 18-30 months

• Quality problems: Many organizations informally estimate that roughly half of tech transfers encounter serious quality or execution problems

• Cost per transfer: Over $5 million

Rentosertib (Insilico’s fully AI-discovered drug) hit Phase 2 proof-of-concept in Nature Medicine after 4 years total. Exscientia’s DSP-1181 completed discovery in under 12 months versus 4.5 years traditionally.

Meanwhile, tech transfers haven’t improved. External transfers add 5.8 months compared to internal. Analytical assay transfer problems are the #1 reason for delays. When your CDMO can’t reproduce testing methods at the receiving site, you can’t validate product quality.

You know this pain point. That moment when your CDMO can’t replicate the analytical results on reference standards and the entire transfer stalls for months.

What COVID Taught Us (That Most Companies Ignored)

The EMA stated it explicitly in their 2022 lessons learned report: CMC was “generally on the critical path to authorization” for COVID vaccines. Not clinical development.

Pfizer produced 3 billion doses by end of 2021. Novavax’s authorization lagged competitors by 8-12 months despite having promising clinical data. Why? Manufacturing struggles at multiple sites.

Manufacturing readiness, not clinical results, determined who won the race.

This matters for your career: The function you work in just became strategically critical in a way it hasn’t been in decades. Supply chain and manufacturing concerns are now the #1 priority for pharmaceutical executives (GlobalData Q1 2024). Up to 80% of late-stage clinical trials experience preventable delays, and manufacturing and supply problems are a major, growing contributor alongside recruitment and protocol issues.

For complex therapies like cell and gene therapy, a disproportionate share of regulatory questions and clinical holds now stem from CMC and quality issues, far more than was typical for small-molecule drugs.

The Acceleration Pathways Actually Working

I’ve been digging into what’s actually generating results. Not vendor marketing claims, but verified implementations with documented ROI. Two specific AI applications are cutting tech transfer timelines and improving manufacturing efficiency right now:

1. Digital CMC Platforms for Tech Transfer Automation

Traditional tech transfer relies on hundreds of PDFs, Excel spreadsheets, and Word documents. Critical information is buried, incomplete, or never formally documented. That tacit knowledge (the “tips and tricks” experienced operators use) rarely transfers effectively.

Digital CMC platforms atomize this knowledge into structured, interconnected data that can be searched, analyzed, and transferred instantaneously.

In published QbDVision case studies, individual sponsors report substantial benefits:

• Up to 75-80% reductions in per-transfer costs (global enterprise drug developer)

• About 80% reduction in labor costs for site-to-site transfers

• 50% reduction in time for gap assessments

• In some cases, 3× faster tech transfers with 6-12 months removed from timelines (mid-sized US developer)

• Multi-million dollar savings per transfer in certain implementations

A senior director at Bayer noted that after completing numerous tech transfers with QbDVision, the platform became essential to their process.

2. Smart Manufacturing with Predictive Maintenance

Remember that equipment failure that shut down your fill-finish line for three days last quarter? Predictive maintenance using IoT sensors and machine learning can forecast failures before they happen.

Quantified savings in industry case studies:

• Around $200,000 saved per avoided downtime incident

• Unplanned manufacturing downtime can cost $260,000/hour on average

• 30% reduction in unplanned stops

Some FactoryTalk case studies describe pharmaceutical customers eliminating certain contamination-related batch losses and saving hundreds of thousands of dollars annually through predictive maintenance and improved monitoring.

The pattern I’m seeing: Companies that deployed these systems 12-18 months ago are now seeing the compounding benefits. Those starting now will hit their stride in late 2026.

What FDA’s January 2025 Guidance Means for You

If you’re deploying AI in manufacturing operations, you need to understand FDA’s new framework. They released comprehensive guidance on AI in drug development in January, establishing a 7-step risk-based credibility assessment process.

This isn’t optional. FDA has indicated it has already reviewed hundreds of drug and biologic submissions with AI or advanced modeling components since 2016.

Key requirements you need to know:

• Model risk assessment: Combine model influence with decision consequence

• Credibility assessment plans and reports: Required for regulatory submissions

• ALCOA+ data integrity: Training data must be attributable, legible, contemporaneous, original, accurate, complete, consistent, enduring, and available

• Change control for model updates: Retraining and architecture changes require proper change management

• Early FDA engagement: For high-risk applications, schedule Pre-IND or Type B meetings

If you’re deploying AI without a validation strategy, you’re creating compliance risk that will cost significantly more to remediate than to do correctly from the start.

FDA wants to support innovation. They’ve established specific pathways for early engagement. Use them.

What Does This Mean for Your CMC Career?

I’ve been watching this play out from both sides. Operational execution and strategic planning. The teams getting traction aren’t the ones with the biggest AI budgets or the most sophisticated tools.

They’re the ones who recognized that the bottleneck shifted. They stopped optimizing last decade’s problems and started addressing this decade’s reality.

What strikes me is how many organizations still have their best people focused on clinical optimization while tech transfers languish. It’s like optimizing checkout lines while the warehouse is on fire.

From what I’m seeing in current searches: CMC and manufacturing experts for late-stage biotechs are often treated as exceptions to hiring freezes right now. Not signs of broad recovery. Just exceptions. The market is brutal, but specialists who combine deep process knowledge with digital fluency are still in demand.

Skills becoming more valuable: Regulatory affairs and CMC documentation expertise, digital fluency (Veeva, IQVIA, cloud LIMS, AI platforms), data analysis and AI literacy, cross-functional leadership for AI implementation, and process troubleshooting that AI can’t replicate.

Skills getting commoditized: Routine data compilation and document assembly, manual gap assessments, basic project coordination without strategic value, and entry-level data entry.

From what I’m seeing, the window to get ahead of this is maybe 6-12 months. After that, you’re catching up instead of leading.

Key Takeaways: The AI-CMC Speed Mismatch

• AI compressed drug discovery timelines by 60-75% (from 3-6 years to 12-18 months), but tech transfer timelines remain static at 18-30 months with many organizations informally estimating that roughly half encounter serious quality or execution problems

• CMC became the rate-limiting step in drug development. COVID vaccines proved manufacturing readiness, not clinical results, determines market entry (Pfizer succeeded, Novavax lagged 8-12 months)

• Digital CMC platforms show up to 75-80% cost reduction and 6-12 month timeline cuts in published case studies, while predictive maintenance can save around $200K per avoided downtime incident with 30% reduction in unplanned stops

• FDA’s January 2025 AI guidance requires 7-step risk-based credibility assessment, ALCOA+ data integrity for training data, and early engagement for high-risk applications before regulatory submission

• Operations professionals who act in the next 6-12 months will establish compounding advantages while competitors continue optimizing outdated bottlenecks. CMC expertise combined with digital fluency is now the strategic differentiator

Your 6-Month Priority Framework

Based on what I’m seeing work across multiple organizations, this is the actionable roadmap:

Months 1-2: Assess and educate

• Inventory current AI/digital systems and document biggest pain points

• Apply FDA’s 7-step framework to existing or planned AI systems

• Train leadership team on January 2025 guidance

• Research and shortlist vendors for priority applications

Months 3-4: Launch pilot program

• Select ONE high-impact pilot (digital CMC platform for single tech transfer, OR predictive maintenance on highest-downtime equipment)

• Define clear success metrics: time savings, cost reduction, quality improvements

• Document rigorously for validation and FDA credibility assessment

Months 5-6: Scale and systematize

• Develop phased rollout plan based on pilot results

• Formalize AI governance committee with cross-functional representation

• Create SOPs for AI system validation and change control

• Engage FDA early for high-risk applications

Common mistakes to avoid: Deploying AI without validation strategy. Underestimating data quality requirements. Ignoring change management. Over-relying on vendor claims. Failing to engage FDA early. Pursuing AI for AI’s sake without clear business case.

The teams that execute this framework in Q1 2026 will be 12-18 months ahead by 2027.

Look, I get the exhaustion. Another trend to track. Another skill gap to worry about. Another thing that might make your expertise obsolete.

But this one’s different. This isn’t about learning to code or becoming an AI expert. It’s about recognizing that the part of drug development you know intimately (CMC, tech transfer, manufacturing) just became the critical path. The bottleneck shifted to your domain.

That’s not a threat. It’s an opportunity.

The next 6-12 months matter because the companies figuring this out now will set the pace for everyone else. And the professionals positioned at the intersection of deep CMC knowledge and practical AI fluency? They’re not just surviving this transition. They’re defining it.

Next issue: I’ve been digging into McKinsey’s latest AI research, and one stat jumped out: only 6% of companies are seeing real impact from AI. In biotech ops, it’s probably closer to 2%. The gap between the 6% and everyone else? I’ll walk you through what that actually means.

For now: What’s your longest tech transfer in the queue right now? How many months has it been running? Hit reply and tell me. I’m collecting patterns for a deeper dive on what’s actually causing the delays.

The answers are rarely what people think.

—Jizel

The truth is: nobody’s asking the right question yet.

AI is already handling parts of biotech operations: automating regulatory documentation, flagging deviations in batch records, tracking tech transfer milestones. That’s not future talk; it’s happening now.

But here’s what I’m seeing across CMC and regulatory operations teams: the ones succeeding with AI aren’t the ones spending the most on tools. They’re the ones who developed these five capabilities first.

This matters because leadership is asking for AI roadmaps right now. And if your answer starts with “which vendor” instead of “what critical gaps do we actually have,” you’re building on sand. According to LinkedIn’s 2024 analysis, 76% of biotech operations job postings now prioritize “durable skills” such as problem-solving, judgment, and communication over technical credentials alone.

These Aren’t Soft Skills: They’re Your Competitive Advantage

These aren’t “soft skills”. They are the hard competitive advantages in our field. AI can process batch records and track milestones. These five capabilities are still uniquely human and increasingly valuable.

Skill 1: Problem Framing: Defining the Bottleneck Before Buying the Tool

The first question in most AI conversations is “which tool should we use?” That’s backwards.

Here’s what actually works: identifying the real constraint first.

When CMC teams talk about AI for tech transfer, they usually want to talk about automating batch record reviews. But when asked “What decision in your tech transfer process actually takes too long, and why?” the answer is almost never just batch record reviews.

It’s analytical testing timelines. It’s stakeholder alignment and approval cycles. It’s defining what “ready to transfer” actually means across manufacturing, regulatory, and the CDMO.

Some teams spend months evaluating AI tools, only to discover the real bottleneck was stakeholder alignment, not documentation. No AI fixes that.

When you frame the problem correctly, the solution becomes obvious, and it’s often not the tool you thought you needed.

Skill 2: Critical Judgment: Reading What AI Misses in Context

AI is brilliant at pattern recognition. It’s terrible at understanding the full context (caveat: though modern AI systems are improving when context is explicitly provided). They still lack the tacit knowledge, institutional memory, and real-time situational awareness that seasoned biotech professionals bring to complex decision-making.

Here’s a scenario: AI flags a deviation pattern in batch records and recommends a process change. The data is compelling. The recommendation is logical.

But you know three things the AI doesn’t: your CDMO is already at capacity, quality is backlogged on reviews, and this change requires validation work that delays your IND timeline by 3 months.

AI analyzes data. You synthesize context such as regulatory risk, organizational capacity, relationship capital, and timeline constraints. That’s why leadership needs you in the room when decisions get made, not just AI-generated reports.

Your pattern-matching happens across dimensions AI can’t see. That’s not a weakness to fix with more training. That’s your strategic value.

Skill 3: Structured Communication: Translating Complexity Into Decisions

An IND submission has 47 open action items. Three functions are involved. There’s an external CDMO. Everything is “yellow” status.

AI can generate a comprehensive status report with every detail tracked and color-coded.

But what your VP actually needs is simple: the 3 decisions that matter today. Do we delay the validation run? Do we escalate the CDMO resource gap? Do we modify the protocol?

And more importantly, here are the trade-offs for each option, the timeline implications, and my recommendation.

As AI handles routine status tracking, the premium on “can you make this actionable for leadership” goes exponential. Anyone can pull a report. Not everyone can distill complexity into decisions with clear options and implications.

That synthesis is a premium skill.

Subscribe now

Skill 4: Cross-Functional Orchestration: Aligning Humans When Stakes Are High

MSAT or Tech transfer team says “we’re ready to transfer.” Your CDMO says “we need 2 more weeks for resource allocation.” Regulatory says “we need this for the IND filing next month.”

AI can flag the conflict on a dashboard. It can’t resolve it.

What resolves it? You understanding the political landscape: who needs what, when, and why. You orchestrate the conversation that gets everyone aligned on what’s actually possible.

This isn’t just project management. It’s reading the room across organizational boundaries. It’s knowing when someone’s “yes” is actually “yes, but...” It’s building trust with your CDMO contact so they’ll tell you about the resource constraint before it becomes a crisis.

A biotech startup struggling with CDMO alignment implemented a “definition of ready” document that all three teams (manufacturing, CDMO, regulatory, and quality) agreed on upfront. Result: tech transfer kickoffs that previously took 3 weeks of back-and-forth now start in 5 days.

AI can flag conflicts. You resolve them. That’s the difference between a dashboard and a functioning program.

Skill 5: Agency Over Tools: Directing AI Toward Business Outcomes

Here’s where most AI conversations go wrong: “Can AI help with our upcoming IND submission?”

That’s a reactive question. It positions you as waiting for AI to tell you what’s possible.

Here’s the directive approach: “I need to evaluate 3 regulatory filing scenarios: accelerated IND with interim stability data, standard timeline with full 6-month stability, or rolling submission with manufacturing sections first. AI, model the regulatory risk, review timelines, and resource requirements for each option and show me the trade-offs.”

See the difference? You’re not asking what AI can do. You’re directing it like a strategic analyst toward specific business questions you need answered.

The CMC program managers who thrive aren’t being replaced by AI. They’re using it as a tool under their direction and getting promoted for delivering better strategic insights faster.

You’re the pilot. AI is your instrument panel. Don’t confuse the two.

Share

The Bottom Line

• Problem framing matters more than tool selection—ask “What decision in our tech transfer process takes too long?” not “Which AI tool should we use?” (Forces concrete problem definition instead of solution-seeking)

• Critical judgment separates good from great—AI flags deviation patterns; you assess whether the fix conflicts with CDMO capacity, regulatory timelines, or validation requirements (Context AI can’t see is your strategic value)

• Structured communication becomes exponentially valuable—distill 47 tech transfer action items into the 3 decisions leadership needs to make today (Anyone can pull a report; synthesis is the premium skill)

• Cross-functional orchestration is irreplaceable—when manufacturing, CDMO, and regulatory need alignment, AI can’t read the room or negotiate trade-offs (Political landscape navigation beats dashboards)

• Agency over tools means directing AI—use it as your strategic analyst, not hoping it replaces your judgment (You’re the pilot, AI is the instrument panel)

Subscribe now

Next issue: AI compressed drug discovery timelines by 70%. Tech transfers still take 18-30 months. Here’s why CMC is now the bottleneck and the three places to deploy AI there instead.

Until then, which of these 5 skills shows up most in your CMC or regulatory operations work? Hit reply and tell me. I read every response.

—Jizel

You’re scrolling LinkedIn during your morning coffee. Another headline: “Major Biotech Announces 516 Layoffs in New Jersey.” Swipe. Next post: “Johnson & Johnson Acquires Intra-Cellular Therapies for $14.6 Billion.” Swipe again. Someone’s celebrating: “Thrilled to announce AstraZeneca’s $1 billion acquisition of in-vivo CAR-T pioneer EsoBiotec!”

The whiplash is real.

One post screams crisis. The next screams opportunity. Both are happening at the same time, and both are reshaping biotech operations (program management, supply chain, clinical operations, CMC, regulatory affairs, digital transformation) faster than most professionals realize.

Here’s the uncomfortable truth: the biotech industry isn’t contracting or expanding. It’s transforming. According to the World Economic Forum’s 2025 Future of Jobs Report, 39% of the skills you rely on today will be outdated or fundamentally transformed by 2030.

Whether you’re managing CDMO relationships, coordinating tech transfers, running clinical trials, or navigating regulatory submissions, the question isn’t whether AI will change your role. The question is whether you’ll position yourself to navigate that change or get swept aside by it.

🎯 Key Takeaways

• The biotech paradox is accelerating: 2024 saw 192 layoff rounds (15,134+ jobs lost) while M&A deals surged. Biotech is cutting operational overhead while investing billions in novel modalities and digital transformation

• 69% of pharma and biotech firms actively use AI (NVIDIA 2025 Healthcare Report), yet companies are “hungry for talent at the intersection of biology and technology.” Professionals who can bridge domain expertise with AI capabilities are in demand

• Career resilience = T-shaped skills: Deep domain expertise (your vertical bar) + adjacent AI capabilities (your horizontal bar). You don’t need to become a data scientist. You need to become the translator between AI tools and biotech operational reality

• Work is splitting into three categories: Automatable (AI handles it), Augmented (AI assists, you decide), and Irreplaceable (human judgment AI cannot replicate). Your career value lives in the last two

What Skills Will Keep Biotech Professionals Valuable as AI Advances?

Every biotech operations role is splitting into three distinct work categories, and only two of them will matter for your career in 2030.

Whether you’re in supply chain, clinical operations, regulatory affairs, CMC development, quality systems, biomanufacturing, or program management, this transformation is happening now.

Automated: Tasks AI Can Handle Today

Automatable work is everything AI can handle without human judgment. This includes documentation, data entry, reporting dashboards, routine tracking, meeting summaries, status updates, and compliance checklists.

Real example: One biotech reduced tech transfer documentation review from 6 weeks to 11 days using AI-powered document analysis. The AI didn’t make decisions. It flagged discrepancies, identified missing sections, and pre-populated standard language according to regulatory templates.

Your move: Identify which of your tasks fall here and get comfortable delegating them to AI tools. This frees your capacity for higher-value work where human expertise matters.

Augmented: AI Assists, You Decide

This is where career value multiplies. AI provides intelligence and analysis, but you apply operational judgment, relationship context, and strategic thinking.

Forecasting, risk assessment, modeling, data analysis, scenario planning. These tasks need both AI speed and human wisdom.

Function-specific examples:

• Supply Chain: AI predicts CDMO capacity constraints and delivery risks; you layer in relationship context, partnership history, and external manufacturing dynamics

• Clinical Ops: AI analyzes site performance metrics and patient demographics; you factor in investigator relationships, community engagement realities, and local regulatory nuances

• Regulatory: AI tracks global intelligence across 40+ agencies; you craft agency strategy based on years of reading between the lines in feedback letters

• Program Management: AI models scenarios and highlights risks; you navigate cross-functional politics, stakeholder dynamics, and competing organizational priorities

Real example: When one clinical operations team implemented AI-powered site selection, patient recruitment timelines dropped from 18 months to 12 months. The AI analyzed historical trial data across 200+ sites. But the human ops manager made the final call by factoring in recent investigator turnover, regulatory inspection history, and community relationships the AI couldn’t see. Result: 6-month faster enrollment without compromising site quality.

Another example: A regulatory affairs team used AI to monitor regulatory changes across global markets. The AI flagged 15 potentially relevant guidances per week. The human regulatory lead identified which 3 actually impacted their IND timeline, saving the team from reading 12 irrelevant documents and focusing resources where they mattered. Result: 40% reduction in regulatory intelligence review time while maintaining compliance.

Your move: Test AI tools in this category. You stay in control, make the strategic calls, and work dramatically faster.

Irreplaceable: Human Expertise AI Cannot Replicate

These are the capabilities that will define career success through 2030 and beyond.

Relationship management when trust is broken. Crisis navigation when the unexpected happens. Strategic trade-offs when there’s no “right” answer. Cross-functional coordination when priorities conflict. Reading the room during tough conversations. Building coalitions across departments. Negotiating under uncertainty when relationships matter more than contracts.

Career Resilience: The ability to maintain and grow your professional value as technology, organizational structures, and industry dynamics shift. In biotech operations, this means positioning yourself at the intersection of deep domain expertise and AI-augmented decision-making.

Your move: This is where you focus your time and build your long-term career value. These are the skills that get you promoted, recruited by competitors, and consulted when crises hit.

The pattern is clear: The biotech professionals getting promoted aren’t doing the most work. They’re positioning themselves at the intersection of AI capability and irreplaceable human judgment.

How Do You Build a Career That’s Both Deep and Adaptable?

Think of your career as a “T”:

The vertical bar = your deep domain expertise

(Research and development, supply chain, clinical ops, regulatory, CMC, quality systems, biomanufacturing, program management, medical affairs)

The horizontal bar = adjacent capabilities that multiply your value

(AI fluency, data analysis, cross-functional leadership, business acumen, digital tool proficiency, change management)

According to the WEF report, 59% of the global workforce will need reskilling or upskilling by 2030. Skills gaps are the #1 barrier cited by 63% of employers.

You don’t need to become an AI expert or learn to code. You need to become the T-shaped professional in your function who knows:

• Where AI adds value in your specific work (and where it creates more problems than it solves)

• Where human expertise remains irreplaceable and why

• How to translate between AI capabilities and biotech’s complex regulatory, quality, and partnership constraints

What’s Your 3-Week Action Plan to Start Building Career Resilience?

Week 1: Audit Your Role

Time investment: 2 hours total

• List all your regular responsibilities (meetings, deliverables, decisions, communications, recurring tasks)

• Categorize each task: Automatable | Augmentable | Irreplaceable

• Analyze your time allocation: Where am I spending most of my energy? Am I stuck in automatable tasks that drain capacity? Am I building irreplaceable skills that increase my value?

• Research ONE AI tool relevant to your function (don’t implement yet, just explore what’s available)

Week 2: Test AI Fluency

Time investment: 30 minutes daily for 5 days

Pick 2 AI tools and actually use them for one week with real (non-confidential) work scenarios.

Universal Starter Tools (Everyone Should Test):

• ChatGPT (OpenAI) or Claude (Anthropic) or Microsoft (Copilot) – Meeting summaries, document drafting, email composition, brainstorming

  • Free tiers available. Start here to learn AI interaction patterns.

• Perplexity AI – Research synthesis, competitive intelligence, literature scanning, regulatory guidance summaries

  • Better than Google for biotech-specific questions. Cites sources automatically.

Function-Specific Examples:

• Supply Chain: Anaplan, Coupa AI

• Clinical Ops: Deep 6 AI, Florence Healthcare, Medidata AI

• Regulatory: Veeva Vault RIM AI, Compliance.ai

• Program Management: Asana, Notion AI

• CMC/Quality: MasterControl, Mareana

• Research and Development: AlphaFold, Benchling AI

Your action this week: Pick ONE universal tool and ONE function-specific tool. Test and learn both for 5 days with sanitized work examples. Document what works, what’s clunky, and where AI misses critical context.

🚨 Critical Security Reminder

Follow your company’s AI usage policies strictly. Never upload proprietary information, trade secrets, confidential data, patient information, or partner details. Use “sanitized” examples only. Many biotechs have approved AI platforms. Check with IT or compliance before testing any external tool.

Week 3: Position Strategically

Time investment: 90 minutes total

• Update your LinkedIn profile: Add “AI Integration” or “Digital Transformation” alongside your core expertise

  • Example: “CMC Program Leader | AI-Augmented Operations | Strategic Partnership Management”

• Volunteer for a pilot project: Approach your manager with specific value

  • “I’ve been testing [specific tool] for [specific use case]. Want me to run a 2-week pilot and report findings to the team?”

• Initiate one strategic conversation: Schedule 15 minutes with your director

  • “I’ve identified where AI could help our team with [specific bottleneck]. What concerns should I address first before proposing a pilot?”

The goal: Become the T-shaped professional in your function who bridges AI capability with irreplaceable human judgment.

Worth Your Time

📖 World Economic Forum Future of Jobs Report 2025 – Skills transformation data across all biotech functions and what employers actually need

🎧 NVIDIA 2025 Healthcare AI Report – 600+ pharma/biotech professionals on AI adoption realities (cuts through the hype)

What resonated?

The whiplash isn’t going away. The M&A activity will continue. The layoffs will continue. The AI transformation will accelerate.

The question is whether we’re positioning ourselves to navigate it or standing still hoping it passes.

Coming in two weeks: The five irreplaceable skills that AI can’t replicate (and how to build them fast). Plus: the exact 30-day visibility plan to position yourself as the AI-ready professional on your team, without becoming a data scientist. 🔍 Deep Dive Edition

Where’s your T-shape right now? Deep expertise but narrow AI skills? Or are you building your horizontal bar? Hit reply and tell me your function and what AI tools you’ve tested. I genuinely want to hear what’s working (or not).

Know someone navigating AI disruption in biotech?

Forward this. They can subscribe here: https://substack.com/@resilientfutures

Until next time,

Jizel

Disclaimer: All AI tool recommendations are for educational purposes only. Follow your company’s AI policies and never upload proprietary information to third-party platforms without authorization.