Engagement note

What the First 90 Days Look Like

For biotech-shaped AI-oncology companies — those running their own programs toward an IND under their own name. A parallel variant exists for AI-platform companies.

A fractional engagement should not be a slow ramp into another stakeholder meeting. Ninety days is enough time to retire real decisions if the work is sequenced honestly and the deliverables are concrete. What follows is the operating arc I propose with biotech-shaped AI-oncology companies — four phases, twelve weeks, one specific artifact and one specific decision retired in each.

The phases are not rigid. They are a structure for what the work should produce when the engagement is going well. If the science, the data, or the agency forces a reordering, the phases reorder. The deliverables and the retired decisions do not.

---

Phase 1 — Weeks 1 to 2: Diagnostic

Artifact produced. A one-page Calibration Map of the program. Not a status report. The Map is a structured separation of what is known (cited, source-tiered), what is conditionally true (true if X holds, where X is named), what is asserted but unsupported (named, with the gap), and which of the next eight to twelve decisions in front of the program carry the highest regulatory consequence — ranked.

Both the program-level Map and any partner-portfolio Map carry four small decision fields, written in plain English:

• Risk track — what kind of risk the decision is mostly carrying: endpoint, population, regulatory precedent, manufacturing, biomarker, or operational.
• Reference class — the closest comparable program, submission pattern, or precedent class, with a note on why the comparison is fair and where it breaks.
• Next best learning action — the analysis, conversation, or experiment most likely to change the decision if the current view is wrong. The value-of-information check, made plain: "What would we learn next if we had to spend one unit of scarce time or money?"
• Revisit trigger / date — the milestone, data readout, or calendar point at which the decision is deliberately reopened.

In parallel: read-in to the asset, the prior preclinical and translational data, the current development plan, the AI platform's role in target selection or candidate prioritization (and the validation status of that role), and the current operating cadence with the founders and board.

Decision retired. Scope of the engagement.

By the end of week two, you and I both know what the next 90 days should actually produce — and what they should not. If the diagnostic surfaces a structural problem (the target is too early, the platform validation is not yet load-bearing on the program, the company is fundraising harder than developing), I will say so and we will recalibrate the engagement scope.

What this is not. The Diagnostic is not a due-diligence memo for investors. It is internal calibration work, written for the founders and the operating team.

---

Phase 2 — Weeks 3 to 5: Endpoint and Indication

This phase follows an expansive-then-decisive arc. The three memos at the end of Phase 2 are the output of that arc, not the input to it. The working sessions that produce them deliberately separate what is known (facts, source-tiered), what is possible (destinations the asset could reach, with no constraints applied), what is assumed (beliefs we operate on without definitive evidence), and what narrows the field (regulatory, scientific, operational, resource constraints). The expansive sessions come first; convergence to the memos comes second. The discipline guards against the most common failure mode at this phase — an AI-platform-anchored team locking in a single Target × Indication × Endpoint combination before alternative destinations have been systematically considered.

Weeks 3 to 4: Expansive working sessions

Three short sessions, sequenced over the first two weeks of the phase. Internal calibration material, not external deliverables.

1. Knowledge. What is actually known about the molecule, the disease landscape, the regulatory environment, the competitive field. Facts surfaced; sources tier-labeled; "facts" challenged for whether they are facts or assumptions in disguise. The team explicitly suspends consideration of constraints and ambition during this session — both will get their own.
2. Blue Sky. What destinations are possible for this asset, with no constraints applied. What indications, populations, endpoint types, regulatory pathways, lines of therapy, and combination contexts could this molecule plausibly reach? The point is not to commit to any of them. It is to ensure that the destination ultimately chosen has been chosen against alternatives that were actually considered, not by default. External orientation matters most in this session — KOL conversation, community-oncologist input, patient-advocacy perspective, regulatory-consultant cross-check on adjacent precedent.
3. Assumptions. What the program is currently betting on without definitive evidence, sorted into the safe-to-bet-on category and the foolish-to-bet-on category, each with the evidence base named. AI-platform-derived priors get explicit treatment here: which platform outputs are the program operating on as if they were facts, and what is the actual validation status of each.

Output of the expansive block: a 5- to 10-page working note capturing the surface area. Internal to the founders and operating team; not a deliverable for investors or the agency.

Weeks 4 to 5: Convergence to the three memos

A constraints session narrows the surface area — corporate strategy and capital runway, regulatory landscape and precedent, resource and timeline reality, scientific feasibility. Hidden cognitive biases get named here; perceived constraints are tested against real ones.

Convergence then produces the three memos that are the phase's external deliverable:

1. Endpoint Strategy Memo — what endpoint the program is built to deliver, what the agency has precedent to accept for this mechanism in this disease setting, and (where the science warrants) what reasoned departure from precedent is worth proposing. The memo names the comparable approvals (accelerated approval on early response, tumor-agnostic indication, novel surrogate, external control), surfaces the strongest counterevidence to the proposed endpoint, and states the conditions under which the recommendation would change.
2. Indication and Population Analysis — what patient population the program will actually enroll, pressure-tested against current standard of care, community-oncologist referral pattern, line-of-therapy windows, and biomarker prevalence. Identifies populations the AI platform may have over-counted and populations the platform may have missed.
3. Pre-IND Target Statement — a single page stating what we will ask the agency to accept, on what evidentiary basis, and what the fallback positions are if the agency pushes back.

Decision retired. What we are taking to the agency, and in what population.

Calibrated-departure layer. Where the science calls for a reasoned departure from established precedent, the Endpoint Strategy Memo names it explicitly, locates the agency-acknowledged territory of uncertainty, and proposes the evidentiary package that would make the new path defensible. Fitting the guidance is sometimes the right move; proposing beyond it, in dialogue with the agency, is sometimes the right move. The memo states which is which, and why — and points back to the Blue Sky session where the alternative was first surfaced.

---

Phase 3 — Weeks 6 to 9: Trial Design and FDA Interaction Plan

Artifact produced. A trial design package and a pre-IND interaction plan.

The trial design package includes:

• The Phase I or Phase I/II protocol skeleton — primary objective, eligibility logic (sharpened against the Phase 2 population analysis), dose-escalation framework (per Project Optimus where relevant), biomarker strategy, sample size logic, safety run-in if warranted.
• An operational feasibility check — site identification, expected accrual rate, referral-pattern modeling, competing-trial environment.
• A decision-impact register — what the trial will and will not let the program decide at each interim and final analysis.

The pre-IND interaction plan includes:

• A pre-IND briefing book outline — questions to the agency, supporting data, proposed positions.
• Anticipated reviewer Q&A — separately for the standard-path positions and the reasoned-departure positions. Each anticipated question is paired with the response the briefing book supports.
• A meeting choreography plan — who presents what, in what order, with what fallback if the agency pushes on specific positions.

Decision retired. Trial design submitted for pre-IND (or, if the pre-IND timing is later, the design package locked for board approval and FDA prep).

---

Phase 4 — Weeks 10 to 12: Regulatory Translation and Handoff

The translation package addresses the third failure mode from the Three Ways thesis directly. The AI platform's existing provenance, validation, and drift work is translated into pre-IND-readable artifacts:

• Model Context-of-Use Statement — what role the AI plays in the regulatory decision the agency is being asked to make, at what risk tier per the FDA/EMA joint principles framework.
• Fitness-for-Purpose Documentation — analytical and operational validation of the model for that specific role, in the language a CDER review division will accept.
• Drift Monitoring Plan — how the model's performance will be monitored over the lifecycle of the program, and what triggers retraining or re-validation.

In parallel:

• Strategic Theme one-pager — a single page compressing the calibration trail and the destination into a portable summary. Three sections: WHERE we are taking this asset (the destination, with the patient end in mind), WHY this path and not the alternatives considered (with the calibration trail referenced), and WHAT it will take to win (the next eight to twelve high-stakes decisions, the conditions under which the recommendation would change, and the falsifiers). The artifact that survives the engagement and travels — forwarded to investors, board members, future hires, and any subsequent regulatory consultant.
• Board Documentation Package — the full calibration trail (every load-bearing claim cited, every decision retired, every counterevidence surfaced) compressed into a board-readable format. The Strategic Theme one-pager is the front page; the Board Package is what supports it.
• Next-Stage CMO Recommendation — when the company should hire its full-time CMO, what scope, what profile, what compensation structure, what the recruitment timeline looks like. If the recommendation is that the fractional engagement should continue past 90 days, the recommendation says so and names the conditions under which it should transition.

A note on the recommendation's structure. Where the recommendation favors continuing the fractional engagement past 90 days, the criteria for that recommendation are stated explicitly — so the board can adjudicate the recommendation against the criteria, not against the recommender. This is a deliberate constraint. In the cases where the recommendation favors continuing my own engagement, the constraint protects both the board's decision quality and the integrity of the calibration discipline. The same constraint applies when the recommendation favors transitioning to a full-time CMO: the criteria are named, not assumed.

Decision retired. Post-90-day operating model.

---

What the 90 days do not deliver

• A full-time clinical-regulatory culture (that is a multi-year hire, not a 90-day deliverable).
• A locked Phase I protocol (the trial design package supports pre-IND; the protocol is sharpened post-pre-IND).
• A guarantee of FDA acceptance (no honest operator can deliver one).
• A substitute for the company's eventual internal regulatory and clinical-operations teams.

What it delivers is calibration on the next eight to twelve decisions that determine whether the molecule reaches Phase II — and a working system for making those decisions auditable, defensible, and revisable as the science and the agency evolve.

---

Working logistics: Madrid and Central European Time

I am based in Madrid, Spain — Central European Time (UTC+1 winter, UTC+2 summer). For US-based clients this is a real consideration, designed into the engagement from the start rather than discovered mid-flight.

• Working sessions for East Coast clients. Late-Madrid-morning maps to mid-East-Coast-morning. The overlap window is generous — roughly 14:00 to 18:00 Madrid / 08:00 to 12:00 EST — and most weekly working sessions, pre-IND prep blocks, and decision-retirement meetings fall there.
• Working sessions for West Coast clients. The overlap window is narrower — roughly 17:00 to 19:00 Madrid / 08:00 to 10:00 PST. Workable, with deliberate scheduling and a willingness on the Madrid side to take occasional evening sessions for high-stakes meetings.
• Asynchronous work fills the gaps. The calibration discipline — citation-required outputs, externalized reasoning, structured refusal — produces written artifacts that travel well across time zones. Decisions are captured in writing, not in meetings.
• In-person sessions when justified. Pre-IND meetings, FDA interactions, board meetings, engagement-defining moments — travel is built into the engagement scope. Three to five trips per year of moderate length is typical for a fractional CMO engagement; co-founder structures imply more.

The time-zone gap is a real constraint, not a hidden one. The engagement is designed around it rather than against it.

---

A note on this schema

The 90-day arc described above is a designed schema, not a tested-and-validated process. It draws on three decades of biopharma drug development experience — across Pfizer, Millennium, and Takeda; across seven anticancer approvals at FDA, EMA, PMDA, and NMPA; across pre-IND meetings, advisory committees, and the long quiet work of explaining to families why a trial closed early — and on collaborative work with AI to articulate the calibration discipline that has been implicit in that experience for years.

What this means in practice: specific engagements will reorder phases, adjust artifacts, and surface gaps the schema does not yet handle. When that happens, the schema is updated — visibly, in the document's version history — rather than papering over the gap. The same calibration discipline applied to client work is applied to the schema itself: confidence about what is known, refusal where the evidence does not yet support a claim, and an explicit feedback loop for what each engagement teaches the next.

The schema is honest about being a starting point. The discipline behind it is not.

---

How this differs from a typical "fractional CMO" engagement

Most fractional CMO engagements default to weekly check-ins, monthly board reports, and "advise as needed" scope. That structure produces hours, not decisions. The 90-day arc above is built to retire decisions, not to bill hours.

Three features distinguish this engagement:

1. Every claim of consequence is cited. Internal memos carry source-tier labels per the Citation-Required Output Contract. When the evidence does not support a recommendation, no recommendation is made — refusal is part of the output.
2. The calibrated-departure disposition is explicit. Where the science warrants proposing beyond the published guidance, the proposal is named, the evidentiary package is built, and the agency-acknowledged territory of uncertainty is identified.
3. The expansive-then-decisive arc is structural. Phase 2 is built so that the decisive output — the endpoint, the population, the pre-IND position — is reached only after an expansive surface of alternatives has been generated and considered. Most fractional engagements skip the expansive moment because it appears slow. It is not slow. It is what prevents the calibration trail from becoming a defense of the destination the team had already chosen on day one.

← Back to Writing  ·  Conversation about an engagement →