An advisory board is a closed room in which a sponsor pays a small panel of named experts for their candid scientific judgment, and almost everything about that room is governed: who is in it, what they were paid, what they were asked, and what they said. A medical affairs lead has eight working days before a Phase 3 cardiometabolic advisory board convenes, and the deliverables stack up on both sides of the meeting. Before it, she needs KOL dossiers for nine experts, a conflict-of-interest summary for each, and an agenda mapped to the medical strategy questions the brand team actually needs answered. After it, she needs a post-meeting report that organizes hours of expert input into something the strategy team can act on, fair-market-value documentation that survives a compliance review, and an attendee record that feeds the Sunshine Act transparency report without an error. An enterprise large language model can compress the drafting hours dramatically. It can also, in a single careless paste, expose the identity and unpublished opinions of a named investigator on a competitor's pivotal trial. This lesson walks the full arc, before and after, and treats the confidentiality of named experts as the constraint that shapes every other decision.

Why the Advisory Board Is a Governed Room, Not a Meeting

The reason advisory boards carry so much process is that they sit at the intersection of three regimes that each watch them closely. The first is transfer-of-value transparency: in the United States, the Physician Payments Sunshine Act, implemented as the Open Payments program under Section 6002 of the Affordable Care Act, requires applicable manufacturers to report payments and transfers of value to covered recipients, and an advisory board honorarium is a reportable transfer of value attached to a named physician. The second is anti-kickback and false-claims exposure: a payment to a prescriber must be for a legitimate, documented service at fair market value, because a payment that looks like it bought prescribing rather than advice is the fact pattern that anti-kickback enforcement is built around. The third is scientific-exchange compliance: the meeting must be a genuine solicitation of expert input on a real scientific question, not a disguised promotional event, which is why the agenda, the questions, and the selection rationale all matter.

Each of those regimes generates a document, and each document is the artifact an AI can help draft. Transparency generates the attendee record and the transfer-of-value log. Fair-market-value defensibility generates the FMV rationale and the honorarium calculation tied to a rate methodology. Scientific legitimacy generates the needs assessment, the selection criteria, the agenda, and the post-meeting report that proves real input was sought and captured. The medical affairs lead is not writing prose for its own sake; she is assembling an evidence file that a compliance officer, an auditor, and potentially a government investigator could read years later. That framing is what separates competent advisory board support from a liability, and it is the frame the model does not bring on its own.

The model brings fluency and structure. It does not bring the knowledge that a given expert sits on a competitor advisory board, that an honorarium rate must trace to a documented methodology, or that an attendee's National Provider Identifier must be correct for the Open Payments submission to reconcile. Those are facts that live in the sponsor's own systems and the human's own judgment, and the entire workflow is an exercise in supplying the model what it needs while never letting it leak what it must not. The named medical affairs lead owns the evidence file under her signature, exactly as a medical writer owns a Clinical Overview under theirs.

The Confidentiality Constraint Comes First, Before Any Prompt

The single most consequential decision in this workflow is made before a word is drafted: what may be placed into the model, and under what data protection. Advisory board materials are dense with information that must not reach a public model. The expert panel itself is competitively sensitive, because the composition of a sponsor's advisory board reveals strategic intent to anyone who sees it. Individual experts carry personal data and, when they are also patients' physicians, sit close to protected health information in any case example they might raise. Most acutely, advisory boards routinely include named investigators on pivotal trials, including competitor trials, and an investigator's unpublished read on emerging data is exactly the kind of material that, if leaked, damages the expert, the sponsor, and the trial's integrity at once.

The operational floor is therefore an enterprise deployment under a contract that guarantees the sponsor's data is not used to train the vendor's models and is not retained beyond the session, the zero-data-retention pattern layered on a business associate agreement where any protected health information could be in play. Pasting a roster of nine named cardiologists, three of whom advise a competitor, into a consumer chatbot is not a productivity shortcut; it is a confidentiality breach that no downstream verification can undo, because exposure is irreversible the moment it happens. The mechanical reason is the one established across this program: a model reasons only over what enters its context, and a consumer endpoint may log, retain, and train on that context outside the sponsor's control.

There is a sharper layer for the dossiers specifically. A KOL dossier compiled from public sources, publications, trial registrations, congress presentations, sits at one risk level. The same dossier annotated with the sponsor's private notes on the expert's prior interactions, their perceived receptivity, or their competitor relationships, is internal intelligence that must stay inside the governed environment and must never be phrased in a way that reads as ranking physicians by commercial value, which is both a compliance and a reputational hazard. The discipline is to separate the public evidence base, which the model can help assemble and summarize, from the private commercial annotations, which a human handles inside the sponsor's own systems with deliberate care about language.

Building the KOL Dossiers Without Letting the Model Invent Credentials

A KOL dossier is a structured profile that lets the medical strategy team understand who is in the room: the expert's clinical and research focus, their relevant publications, their trial involvement, their congress activity, their prior engagement with the sponsor, and the scientific topics on which their input is most valuable. An LLM is genuinely good at the synthesis layer of this work. Given a set of loaded source documents, an expert's publication list, their public trial registrations, their society roles, it can produce a clean narrative profile that a strategy lead can read in a minute instead of assembling from a dozen tabs. The acceleration is real, and the structure it produces is consistent across nine experts in a way that hand-assembly rarely achieves under deadline.

The danger is the same one that haunts every generation task: the model will produce a plausible credential whether or not it is true. Asked to profile a cardiologist, the model can confidently state that the expert was principal investigator on a named trial, holds a particular society fellowship, or authored a pivotal paper, and any of those claims may be a fabrication that wears the costume of a fact. A dossier is read by people who will treat its contents as vetted, and a fabricated trial role or a misattributed publication can send a strategy team into a meeting with a false picture of the expert across the table. The control is to ground every dossier claim in a loaded source and to reject any credential the model asserts without one, treating an uncited credential as wrong until a human verifies it against the primary record.

This is where the discipline from the submission world transfers cleanly. Each dossier is a set of claims, not a paragraph of prose: a focus claim, a series of publication claims, a trial-role claim, an affiliation claim, each with a source that a human checks. The publication list in particular should be reconciled against the actual bibliographic record, because a model that produces a list of an expert's papers is doing exactly the kind of citation generation that fabricates references in other contexts. The dossier that goes to the strategy team should carry only credentials traced to a primary source, and the few minutes that verification costs per expert is trivial against the cost of briefing a brand team on a person the model partly invented.

Conflict-of-Interest Summaries and the Selection Rationale

The conflict-of-interest summary serves two audiences at once. It tells the sponsor's own team what entanglements each expert carries, the competitor relationships, the other advisory boards, the consulting arrangements, the equity positions, so the team can manage the meeting and weigh the input appropriately. And it feeds the compliance record that demonstrates the sponsor selected experts for genuine scientific reasons and managed their conflicts rather than ignoring them. The summary draws on disclosed conflicts from publications, public payment databases, registry disclosures, and the sponsor's own engagement history, and the AI can assemble these disclosed facts into a consistent per-expert summary far faster than a human can.

The hazard is twofold and specific. First, a conflict-of-interest summary is precisely the document where a fabricated or omitted entanglement does the most damage: a model that invents a competitor relationship defames the expert, and a model that drops a real one leaves the sponsor managing a conflict it does not know exists. Every stated conflict must trace to a disclosure source, and the absence of a stated conflict must never be read as confirmation that none exists, because the model does not experience a missing disclosure as a gap. Second, the language matters enormously: a summary that reads as the sponsor weighing how to influence a conflicted expert, rather than how to manage a disclosed conflict, is the kind of phrasing that turns a routine document into evidence in an enforcement matter. The human owns the framing here, deliberately and word by word.

The selection rationale ties the panel together. Compliance expects the sponsor to be able to explain why these specific experts were chosen for this specific scientific question, in terms of expertise and relevance rather than prescribing volume or commercial reach. The model can draft this rationale from the needs assessment and the dossiers, mapping each expert's documented expertise to the agenda topics, and that draft is a real time-saver. But the rationale must be true to the actual selection logic, not a retrofitted justification the model composed to sound compliant, because a rationale that does not match the real reasons the experts were chosen is worse than none. The human supplies the genuine selection logic; the model organizes it into defensible prose.

Agenda Alignment to the Real Scientific Questions

An advisory board earns its compliance footing by being a real solicitation of expert input on questions the sponsor genuinely needs answered, and the agenda is where that legitimacy is written down. The medical affairs lead starts from a needs assessment: the specific scientific or clinical uncertainties the medical strategy must resolve, framed as questions an expert panel is uniquely positioned to address. The AI can take that needs assessment and a set of background materials and produce a structured agenda, time-boxed sessions, framing context for each discussion topic, and the specific questions to be posed, in a format the panel chair and the strategy team can refine.

The line the model must not cross is the line between scientific exchange and promotion, and it is a line the model does not natively understand. An agenda built to genuinely seek input asks open scientific questions about evidence gaps, unmet need, trial design considerations, and the interpretation of emerging data. An agenda that has drifted toward promotion asks the panel to validate the sponsor's product positioning, to endorse messaging, or to react to a competitor in marketing terms, and a model optimizing for a fluent, on-brand agenda can drift in that direction without signaling that it has. The human reviews every agenda item against the question: is this a genuine request for the expert's scientific judgment, or is it asking them to ratify a commercial position? The former is the purpose of an advisory board; the latter is the fact pattern that recharacterizes the whole meeting as a promotional event.

Agenda alignment also means binding the agenda to the dossiers and the selection rationale, so the thread is visible: these uncertainties drove this panel composition, which maps to these agenda topics, which will be captured in this report structure. When the model produces an agenda that floats free of the documented needs assessment, the alignment breaks and the compliance story weakens, because a reviewer cannot trace the meeting back to a genuine need. The human keeps the thread intact, ensuring each agenda topic answers a documented question and each question maps to an expert chosen for it.

The Post-Meeting Report: Categorizing Expert Input Without Distorting It

After the meeting, the medical affairs lead faces hours of discussion to be turned into a structured report that the medical strategy team can use and that the compliance file can hold. The report typically organizes expert input into categories, by agenda topic, by theme, by the type of input such as evidence-gap identification or trial-design suggestion, and summarizes what the panel said in a way that is faithful to the discussion. This is a task where AI excels at the mechanical layer: given a transcript or detailed notes, it can cluster comments by theme, draft topic summaries, and produce a consistent report structure across all sessions, work that would otherwise consume a day.

The two failure modes here are distortion and attribution. Distortion is the subtle one: a model summarizing a nuanced, divided discussion tends to smooth it into a cleaner consensus than the experts actually reached, because a confident summary is more fluent than an accurate account of disagreement. If three of nine experts dissented sharply on a trial-design point, a report that renders this as general agreement has falsified the very input the sponsor paid to obtain, and the strategy team then acts on a consensus that did not exist. The human reads the report against the actual discussion to confirm that genuine disagreement is preserved and that minority views are represented, because the value of an advisory board is often precisely in the dissent.

Attribution is the confidentiality-critical one. In most contexts, the post-meeting report should capture input thematically without attributing specific statements to named experts, both to protect candor and because attributing a specific unpublished opinion to a named investigator is exactly the leak the whole workflow is built to prevent. A model working from a transcript with names in it will happily attribute, because that is what the source contained, so the human must enforce the de-attribution discipline, ensuring the report reflects what was said without creating a named record of who said what unless the engagement was explicitly structured to permit it. The report is the place where a careless attribution becomes a durable document, and the human owns that boundary.

FMV Documentation and Sunshine-Act-Compliant Attendee Tracking

The honorarium each expert receives must be defensible as fair market value for the service rendered, and the documentation has to show the rate methodology, the time the expert spent, and the calculation. Fair-market-value rates are typically derived from a documented methodology, a benchmarking framework tied to the expert's specialty, credentials, and the nature of the service, and the AI can help draft the FMV rationale and assemble the calculation once the human supplies the rate methodology and the actual hours. The model must not invent a rate: an FMV figure that the model generated to look reasonable, rather than one derived from the sponsor's documented methodology, is the comparability-criterion failure in a new costume, a number that reads as derived but corresponds to nothing. Every honorarium amount must trace to the methodology and the recorded time, computed rather than plausibly asserted.

The attendee record feeds the most unforgiving part of the workflow: the Open Payments transparency submission. Each covered recipient who received a transfer of value, the honorarium, sometimes travel and meals, must be reported with accurate identifying information, and the National Provider Identifier and the payment amount must reconcile exactly to the Open Payments submission. This is a reconciliation task, not a generation task, and the distinction is everything. An AI that helps organize the attendee data into the reporting structure is useful; an AI that generates or guesses an NPI, a covered-recipient status, or a payment amount has introduced an error into a federal transparency report, where a mismatch is publicly visible and traceable to the manufacturer. The human verifies every reportable field against the source of truth, treating the attendee record as a set of facts to be reconciled rather than text to be drafted.

There is a timing and completeness dimension that the model will not raise on its own. Transfers of value have reporting deadlines and category rules, certain transfers are reportable and others are excluded, and the determination of what is reportable is a compliance judgment, not a drafting one. The model can structure the record and flag the fields that need values, but it cannot decide reportability, and a writer who lets the model make that call has delegated a regulatory determination to a pattern completer. The medical affairs lead, working with compliance, owns the reportability determination and the accuracy of every field, and uses the model only to assemble and format what the humans have already verified.

What This Means for the Medical Affairs Lead on Monday

The advisory board workflow is a strong case for AI assistance precisely because it is drafting-heavy and structure-heavy on both sides of the meeting, and a lead who uses the model well reclaims hours that go back into strategy rather than formatting. But every one of those hours is reclaimed inside a confidentiality and compliance envelope that the model does not enforce on its own. Before any drafting, the lead decides what may enter the model and under what data protection, and treats the panel roster and the experts' unpublished opinions as material that never reaches a non-governed endpoint. That decision is the one that cannot be corrected after the fact.

From there the discipline is the program's discipline, applied to a new artifact set. Every dossier credential and every disclosed conflict traces to a source; an uncited credential is wrong until verified. The agenda is reviewed item by item against the line between scientific exchange and promotion. The post-meeting report preserves genuine disagreement and de-attributes expert input unless the engagement permits otherwise. The FMV figures are computed from a documented methodology, not generated, and every reportable attendee field is reconciled to the source of truth for the Open Payments submission. The model drafts the dossiers, the agenda, the report, and the FMV rationale; the named medical affairs lead certifies the evidence file, owns the confidentiality of the named experts, and signs what she has verified. The eight days the model saves are worth nothing if one of them produces a leak or a transparency error, and worth a great deal when the envelope holds.

Key Takeaways

• The advisory board is a governed room, and every deliverable is an evidence file, not just prose. Transparency law, fair-market-value defensibility, and scientific-exchange compliance each generate a document that a compliance officer or investigator could read years later. The named medical affairs lead owns that file under her signature, and the model does not bring the compliance frame on its own.
• Confidentiality of named experts is decided before any prompt and cannot be corrected afterward. The panel roster is competitively sensitive and routinely includes named investigators on competitor pivotal trials whose unpublished opinions must never reach a non-governed endpoint. The floor is an enterprise deployment with zero data retention, layered on a business associate agreement where protected health information could be in play.
• Every dossier credential and every disclosed conflict must trace to a source, and an uncited claim is wrong until verified. The model will assert a plausible trial role, fellowship, or publication that may be fabricated, and a conflict-of-interest summary is the document where an invented or omitted entanglement does the most damage. Reconcile publication lists against the bibliographic record and own the framing of conflict language deliberately.
• The agenda must stay on the scientific-exchange side of the line, and the report must preserve genuine disagreement. A model optimizing for a fluent agenda can drift toward asking the panel to ratify a commercial position, which recharacterizes the meeting as promotion. A model summarizing discussion smooths real dissent into false consensus and will attribute statements to named experts unless the human enforces de-attribution.
• FMV is computed, not generated, and every reportable attendee field is reconciled, not drafted. An honorarium must trace to a documented rate methodology and recorded hours, and the National Provider Identifier and payment amount must match the Open Payments submission exactly. Reportability is a compliance determination the human owns; the model assembles and formats only what the humans have already verified.