An FDA Advisory Committee meeting is the single most public, most consequential, and least controllable day in a drug's regulatory life. A panel of external experts who do not work for the sponsor, do not work for the agency, and have spent the preceding weeks reading the sponsor's briefing document and the FDA's own review will sit in a room, question the sponsor's presenters in front of a transcript that becomes part of the permanent record, and then vote on questions the agency has framed. The vote is advisory, but a division rarely approves over a strongly negative committee and rarely rejects over a strongly positive one, so the room effectively decides. The sponsor cannot control who is appointed, what the agency asks, or how the discussion turns, which means the only variable the sponsor owns is preparation. This lesson builds the AI-assisted AdComm preparation stack: the sponsor briefing document filed under 21 CFR 14, the presentation deck and speaker notes, the mock voter panel that models the likely committee against the public profiles of its members, the Q&A rehearsal binder of the top fifty anticipated questions and scripted responses, and the post-AdComm minutes interpretation memo. The named failure mode this lesson exists to prevent is the one that sinks unprepared sponsors: voter modeling that under-weights the non-clinical voices in the room, the statisticians, the patient representatives, and the consumer representatives whose questions a clinician-built rehearsal never anticipates.

Why the AdComm Is a Preparation Problem, Not a Persuasion Problem

Sponsors who lose advisory committees rarely lose because their drug was indefensible. They lose because a presenter was surprised by a question, gave a defensive or evasive answer in front of the transcript, and watched a recoverable position become a credibility problem that colored the rest of the discussion and the vote. The committee is not persuaded by polish; it is persuaded by a sponsor that has clearly anticipated every hard question and can answer it directly, with data, without flinching. That is a preparation problem, and preparation at AdComm scale is exactly the kind of large, structured, exhaustive task where AI assistance is transformative, because the work is to anticipate the full space of questions the panel might ask and to have a sourced, rehearsed answer ready for each, across every discipline represented in the room.

The reframing matters because it changes what the AI is for. The model is not there to make the argument more persuasive; the named clinical and regulatory leadership owns the argument and its pitch. The model is there to ensure completeness of anticipation, to generate the question space so thoroughly that nothing in the room is a genuine surprise, and to maintain the consistency between what the briefing document says, what the deck says, what the speaker notes say, and what the scripted Q&A answer says, because an inconsistency across those artifacts is precisely the seam a committee member pulls. Completeness and consistency are the model's gifts. Judgment, calibration, and the decision about how to pitch a contested claim remain the sponsor's, exactly as in the surrogate-defense work.

The Briefing Document Under 21 CFR 14

The sponsor briefing document is the committee's primary preparation material, distributed in advance under the advisory committee framework of 21 CFR Part 14, and it competes for the panel's attention with the FDA's own briefing document, which the sponsor does not write and often does not see until close to the meeting. The sponsor's document has to present the benefit-risk case completely, honestly, and in a way that anticipates the agency's framing, because a committee that perceives a gap between the sponsor's rosy briefing and the FDA's sober one will trust the agency. AI assistance in drafting the briefing document is strongest in the same places it is strong elsewhere: assembling and integrating the evidence base, maintaining consistency with the Module 2.5 and 2.7 summaries the committee may also reference, and ensuring that every claim in the briefing traces to a verifiable source, because a committee member who finds one unsupported claim begins to distrust the document.

The discipline that governs the briefing document is the discipline that governs every AI-assisted regulatory artifact in this program: every factual claim, every efficacy result, every safety number, and every cross-reference must be reconciled to source by a named human before the document goes to the committee, and the briefing document inherits the full 21 CFR Part 11 audit-trail obligations because it is a consequential regulated artifact produced with AI assistance. The briefing document is also where the sponsor must decide how to handle the weaknesses the FDA's document will certainly raise, and that decision is judgment, not retrieval. A briefing document that pre-empts the agency's hardest points and addresses them honestly is far more persuasive to a committee than one that omits them and forces the panel to learn the weakness from the FDA, and the model can help draft that honest treatment but cannot decide how much to concede, which is leadership's call.

Modeling the Voters Against Their Public Record

The mock voter panel is the heart of the rehearsal and the place where the named failure mode lives. The composition of an advisory committee is substantially public before the meeting: the standing members of the committee are listed, the agency announces the meeting and frequently the roster, and each member has a public record, a CV, a publication history, prior AdComm participation, and often a known intellectual position on the therapeutic question. Building the mock voter panel means constructing a persona for each likely voter that reasons the way that voter's public record suggests they reason, and then running the sponsor's positions and answers against the panel to surface the questions each member is most likely to ask and the concerns each is most likely to weight. This is the persona-engineering technique from the reviewer lesson, applied to a known and partly named audience rather than an anonymous reviewing division, which makes it both more powerful and more demanding of accuracy.

It is more powerful because the model can ground each persona in a specific person's actual publications and stated positions, producing a question set that mirrors how that expert genuinely thinks rather than a generic skeptic. It is more demanding because the persona is now a model of a real, named individual, and the same hallucination risk that threatens precedents threatens the voter model: a persona built on a misattributed publication or an invented prior position will rehearse the sponsor against a voter who does not exist, leaving the real one unanticipated. Every element of a voter persona, the publications, the stated positions, the prior AdComm votes, must be verified against the actual public record before it shapes the rehearsal, and a persona the team cannot ground in verified public material is a fabrication that wastes rehearsal time on a fictional voter. The voter model is a structured hypothesis about how a real expert is likely to reason, built from their verified public record; it is never a claim to know what that person will actually say.

The Failure Mode: Under-Weighting the Non-Clinical Voters

Advisory committees are not panels of clinicians. They include biostatisticians whose questions are about the analysis, not the disease; patient representatives whose questions are about lived experience, access, and what the trial endpoints meant for actual patients; and consumer representatives whose questions are about safety, marketing, and the public interest. The single most common and most damaging error in sponsor AdComm preparation is building the mock voter panel out of clinical perspectives and treating the non-clinical voters as an afterthought, because a sponsor team is dominated by clinicians and clinical reasoning and naturally rehearses against the questions it would ask. The result is a sponsor that is beautifully prepared for the oncologist's question about subgroup efficacy and visibly unprepared for the statistician's question about the multiplicity-adjusted primary analysis or the patient representative's question about why the trial measured a surrogate instead of how patients actually felt.

An AI-built voter panel is the most effective available defense against this failure mode precisely because the model does not share the team's clinical bias and will, if instructed correctly, build the statistician persona with the same rigor as the oncologist persona and the patient-representative persona with the same care as the regulatory persona. But the model will only do this if the workflow forces it to, which means explicitly enumerating every voter type on the committee, weighting the non-clinical voters at least as heavily as the clinical ones in the rehearsal, and treating an under-represented voter type as a defect in the preparation rather than a reasonable simplification. The statistician's question about whether the alpha was spent in a pre-specified hierarchy, the patient representative's question about whether the surrogate endpoint matters to patients, and the consumer representative's question about how the drug will be promoted are exactly the questions that surprise unprepared sponsors, and the voter model exists to make sure they do not. A mock voter panel that is all clinicians is not a partial rehearsal; it is a rehearsal for the wrong meeting.

The Q&A Rehearsal Binder of Fifty Questions

The Q&A rehearsal binder operationalizes the voter model into the artifact the presenters actually drill against: the top fifty anticipated questions, drawn from the full range of voter perspectives, each paired with a scripted, sourced, rehearsed response. The model is well suited to generating the candidate question set at the scale and breadth a real committee will exhibit, pulling from the clinical, statistical, patient, consumer, and regulatory perspectives, and pressing each question to the hardest reasonable form rather than a softball version, because a binder that rehearses easy questions prepares the sponsor for a meeting that will not happen. The fifty questions are then triaged by likelihood and by danger, and the most dangerous questions, the ones where a weak answer could swing the vote, get the most rehearsal and the most carefully sourced responses.

The scripted responses are where the discipline reasserts itself. Every answer in the binder must be sourced to verified data, because a presenter who gives a confident answer that a committee member then contradicts with the actual record has done more damage than if they had said they would follow up. The model drafts the candidate answers, but every answer is verified against source and owned by the named presenter who will deliver it, because the presenter, not the model, is accountable for what they say in the room and on the transcript. The binder also rehearses the hardest skill in an AdComm, which is answering the question that was asked rather than the question the presenter wished had been asked, and giving a direct, honest answer including an honest concession where the data are genuinely limited, because a committee rewards directness and punishes evasion. The model can generate the questions and draft the answers; the rehearsal, the calibration, and the judgment about when to concede belong to the human presenters and the leadership that coaches them.

The Post-AdComm Minutes Interpretation Memo

The meeting ends, the vote is recorded, the transcript is published, and the sponsor's job shifts immediately to interpretation, because the vote tally is the least informative part of the outcome. What matters for the division's eventual decision is the substance of the discussion: which concerns the committee weighted, which questions the panel kept returning to, where the agency's own questions revealed its thinking, and what the dissenting and concurring voters actually said. The post-AdComm minutes interpretation memo synthesizes the transcript and the meeting into an assessment for the sponsor's executive team of what the committee's discussion implies for the likely regulatory outcome and what the sponsor should do before the decision, and AI assistance in producing it is genuinely valuable because the transcript is long, dense, and time-pressured to analyze.

The model can summarize the transcript, cluster the recurring themes, and surface the questions the committee weighted most heavily, which accelerates the analysis enormously. What it cannot do is interpret the regulatory significance of those themes, because that requires knowledge of the division's posture, the agency's framing of its questions, and the soft signals from the meeting that an experienced regulatory lead reads and a model cannot. A memo that treats a narrow positive vote as a clean win, when the discussion revealed a safety concern the committee voted to set aside but the division will not, misreads the outcome in the most expensive possible way. The model produces the synthesis; the named regulatory leadership produces the interpretation, and the audit trail records both the AI-assisted transcript analysis and the human judgment built on top of it, so that the basis for the sponsor's post-meeting strategy is fully reconstructable. The transcript is the data; the regulatory read is the conclusion, and the conclusion is the human's.

The Audit Trail Across the Whole Preparation Stack

The entire AdComm preparation stack is produced with AI assistance and feeds a public, high-consequence regulatory event, so every artifact in it inherits the ALCOA+ and 21 CFR Part 11 obligations the program applies throughout. The briefing document, the deck, the speaker notes, the voter personas, the Q&A binder, and the interpretation memo each carry the record of the AI runs that produced them, the sources loaded and verified, and the named human who reviewed and signed each consequential claim. This matters especially for the briefing document, which becomes part of the public record under 21 CFR Part 14, and for the voter personas, which model real named individuals and must be grounded in verified public material rather than fabricated positions that could be both inaccurate and reputationally inappropriate.

The deeper point is that the AdComm preparation stack is a system, and its defensibility comes from the consistency the audit trail enforces across the system as much as from any single artifact. The briefing document, the deck, the speaker notes, and the Q&A answers must tell one story sourced to one verified evidence base, and the audit trail is what lets the team prove, and the sponsor later demonstrate, that they did. A committee member who finds the deck claiming a result the briefing document does not support has found a crack, and the discipline that prevents that crack is the same source-reconciliation and run-capture discipline that governs every other artifact in this program, applied across an integrated set of documents that must agree under public scrutiny. The model builds the stack and keeps it consistent; the named humans verify every claim, calibrate every contested position, interpret the outcome, and own the day.

Key Takeaways

• The AdComm is a preparation problem, not a persuasion problem, and that is exactly the kind of exhaustive, structured task where AI assistance is transformative. Sponsors lose not because the drug was indefensible but because a presenter was surprised; the model's job is completeness of anticipation and consistency across the briefing document, deck, speaker notes, and Q&A binder, while leadership owns the argument and its pitch.
• The single most damaging error is voter modeling that under-weights the non-clinical voters. A mock panel built only of clinicians leaves the sponsor unprepared for the statistician's multiplicity question, the patient representative's question about whether the surrogate matters to patients, and the consumer representative's safety and promotion questions; that is a rehearsal for the wrong meeting. An AI panel is the best defense only if the workflow forces equal rigor across every voter type.
• Voter personas model real named individuals and must be grounded in their verified public record. The same hallucination risk that threatens precedents threatens the voter model: a persona built on a misattributed publication rehearses the sponsor against a voter who does not exist. The voter model is a structured hypothesis about how a real expert is likely to reason, never a claim to know what they will say.
• The Q&A binder rehearses the top fifty questions in their hardest reasonable form, each with a sourced, presenter-owned answer. A binder of softball questions prepares the sponsor for a meeting that will not happen; every answer must be verified against source because a presenter contradicted by the record on the transcript does more damage than one who said they would follow up, and the rehearsed skill is answering the question asked, directly and with honest concession where the data are limited.
• The post-AdComm memo separates synthesis from interpretation, and the audit trail enforces consistency across the whole stack. The model summarizes and clusters the transcript; the named regulatory leadership reads the regulatory significance, because a narrow positive vote can hide a concern the division will not set aside. Every artifact carries its ALCOA+ and Part 11 record, and the briefing document, deck, notes, and answers must tell one story sourced to one verified evidence base under public scrutiny.