A competitor presented new Phase 3 data at the oncology congress on Saturday, and by Monday morning the medical-affairs team needs its own congress deck refreshed, with updated speaker notes, a current on-label reference list, and a medical-legal-regulatory annotation package ready to enter Veeva PromoMats by Thursday. The MSL who owns the deck loads the existing slides, the competitor's published abstract, and the product's approved labeling into the enterprise model and asks for the refresh. An hour later a polished updated deck appears, with crisp new speaker notes that contextualize the competitor data and a clean reference list. It also contains one slide where the speaker note now makes a comparative efficacy claim against the competitor's agent that the product's label does not support, phrased so smoothly that it reads like established fact. That single slide is the reason this lesson exists. In promotional and scientific-exchange materials, compliance is not a quality attribute layered on top of the content; compliance is the product, and the MLR reviewers, not the model, own the judgment about what the evidence and the label permit.

Why Compliance Is the Product, Not a Finishing Step

A congress deck and its speaker notes are regulated materials. Whether they are promotional, governed by the FDA's Office of Prescription Drug Promotion expectations and the parallel EU rules, or non-promotional scientific exchange, they live under a strict constraint: every claim must be truthful, non-misleading, fairly balanced, and substantiated by adequate evidence consistent with the approved labeling. A speaker note that states the product is more effective than a competitor's agent is a comparative claim, and a comparative claim requires head-to-head evidence of the kind that rarely exists and that the label rarely supports. The danger of AI here is not that it writes badly; it is that it writes persuasively, and persuasive off-label or unsubstantiated comparative claims are exactly what the entire medical-legal-regulatory apparatus exists to catch.

This reframes the whole task. The deliverable is not a good-looking deck; it is a deck that will clear MLR review, and a deck that clears MLR is one in which every claim is on-label, every claim is substantiated by a cited and adequate source, and the promotional or non-promotional character of the piece is consistent throughout. The model can accelerate the production of such a deck enormously, drafting speaker notes, assembling reference lists, and structuring the annotation, but it cannot determine whether a claim is on-label, because that determination requires reading the specific approved indication against the specific claim with a regulatory judgment the model does not possess. The model produces candidate content; MLR, the medical, legal, and regulatory reviewers, render the verdict, and the workflow must keep that boundary bright.

The reason the boundary is so easily blurred is the same fluency that makes the model useful. When the MSL asks the model to update a slide in light of the competitor's new data, the most natural, helpful-sounding response is one that positions the product favorably against that data, and positioning favorably slides almost imperceptibly into claiming superiority. The model has no internal sense that crossing from our product showed a median survival of X in its pivotal trial into our product is more effective than the competitor crosses a regulatory line, because both are fluent English and the corpus is full of comparative marketing language. The line is invisible to the model and bright to the reviewer, and the workflow's job is to surface every candidate claim for the reviewer rather than letting fluent comparative phrasing pass as settled content.

The Three Deliverables and How Each Goes Wrong

The refresh produces three linked artifacts, and each has a characteristic AI failure mode. The updated speaker notes are the highest-risk piece, because they are where claims live and where the comparative and off-label dangers concentrate. The model, asked to contextualize the competitor's data, will generate speaker notes that interpret and position, and interpretation is exactly where unsubstantiated claims enter. The discipline is that every assertion in a speaker note must be traceable to the approved label or an adequate, cited source, and any note that characterizes the product's performance relative to a competitor must be flagged for MLR as a comparative claim requiring specific substantiation rather than waved through as helpful framing.

The on-label reference list is the second deliverable, and its failure mode is the one this program has named from the first lesson: the fabricated or mis-attributed citation. A model asked to assemble a reference list supporting the deck's claims will, if the supporting paper is not loaded, generate a plausible citation, a real-sounding journal, a plausible author, a plausible year, for a paper that does not exist or that does not say what the slide claims. In a regulated promotional piece this is catastrophic, because the reference list is the substantiation, and a substantiation that points to a nonexistent or non-supporting source is worse than no substantiation, since it implies evidence that is not there. Every reference must be verified to exist, to be the correct citation, and to actually support the specific claim it is attached to, and references must be on-label, meaning they support claims within the approved indication rather than smuggling in evidence for an unapproved use.

The MLR-ready annotation is the third deliverable, the package that maps each claim on each slide to its supporting reference and flags the regulatory considerations, prepared in the structure the review platform expects, here Veeva PromoMats. The model is genuinely useful at producing this structured mapping, linking claims to references and laying out the annotation in the platform's format, and this is real time saved because the annotation is tedious to build by hand. The failure mode is that the model's annotation can assert that a claim is substantiated by a reference when the reference does not in fact substantiate it, producing a tidy annotation that gives MLR false comfort. The annotation accelerates the reviewers' work; it does not replace their verification, and a claim-reference link the model proposes is a claim the reviewer must still check, not a conclusion the reviewer can accept.

Handling the Competitor's Data Without Inheriting Its Claims

The trigger for this refresh is a competitor's new presentation, and the way the model incorporates that competitor data is its own distinct hazard. A competitor's abstract makes its own claims about its own product, framed favorably and substantiated against its own label, and a model asked to update your deck in light of it will tend to absorb the competitor's framing, sometimes restating the competitor's efficacy figures as if they were neutral facts and sometimes constructing the implied comparison the competitor wanted but that your label does not authorize you to make. Discussing a competitor's published data is permissible in scientific exchange when done accurately and fairly, but importing its claims, or building a head-to-head comparison from two separately conducted trials, is exactly the cross-trial comparison that regulators treat as misleading because the trials differ in population, design, and endpoints.

The discipline is to treat the competitor data as a fact to be characterized accurately, not as a benchmark to position against. The model may summarize what the competitor presented, with the figures verified against the competitor's actual abstract and attributed as the competitor's own reported results, and it may not construct a comparative conclusion that the data do not support. Fair balance compounds the obligation: a slide that presents the product's strengths must not omit the limitations or the relevant safety context that a balanced scientific presentation requires, and the model, optimizing for a persuasive narrative, characteristically under-weights the balancing content. The MSL must restore the balance the model trims and must ensure every competitor figure is the competitor's verified, attributed number rather than a value the model reconstructed from memory of the field.

On-Label Is a Regulatory Judgment, Not a Phrasing Choice

The single concept the MSL must hold firmly is that whether a claim is on-label is a regulatory determination about the relationship between the claim and the approved indication, and it is not something the model can settle by phrasing. A product approved for second-line treatment of a specific tumor type in patients with a specific biomarker has a labeled indication with precise boundaries, and a speaker note that discusses the product's use in the first-line setting, or in biomarker-negative patients, or in an adjacent tumor type, is off-label regardless of how carefully it is worded. The model, drafting from the clinical literature, will readily produce content about uses that are studied and published but not approved, because the literature discusses them and the label is just one document among many in the corpus, and that content is off-label even when it is scientifically accurate.

This is why the approved label must be the governing source loaded into the workflow and why the MSL must read every claim against the specific indication rather than against general scientific plausibility. A claim can be true, supported by good data, and still off-label, and the regulated material may not make it. The model collapses the distinction between scientifically true and on-label permitted, because both are simply plausible text to it, and the MSL must hold the distinction open, routing any claim that touches an unapproved population, line of therapy, dose, or endpoint to MLR with the off-label character flagged. The most dangerous version of this is the claim that is true and useful and only subtly off-label, because its truth and usefulness make it tempting to keep, and its off-label character makes it impermissible, and only a reviewer reading the label can adjudicate the tension.

The MLR Reviewers Own the Judgment, and the Workflow Serves Them

The medical-legal-regulatory review is a tri-disciplinary judgment: the medical reviewer assesses scientific accuracy and fair balance, the regulatory reviewer assesses consistency with the approved label and the promotional rules, and the legal reviewer assesses liability and intellectual-property exposure. None of these is a task the model performs, because each is an accountable professional judgment rendered by a named reviewer who can be questioned by a regulator and who signs off on the piece's release. The model's role is to deliver the reviewers a deck and an annotation that make their judgment faster and better informed, with every claim traced, every reference verified, and every comparative or off-label-adjacent claim pre-flagged, so the reviewers spend their time judging the hard cases rather than hunting for the claims.

The most valuable thing the MSL can do with the model, then, is not to make the deck look finished but to make it maximally reviewable. A deck where the model has surfaced every candidate claim, attached its proposed substantiation, and explicitly flagged the three comparative claims and the one first-line-setting mention as requiring MLR adjudication is a far better input to review than a polished deck that has smoothed those same issues into confident prose. The smoothing is the danger; the surfacing is the value. A reviewer can quickly approve a flagged comparative claim that turns out to be substantiated, or quickly kill one that is not, but a reviewer cannot act on a comparative claim that has been phrased to look like settled fact and buried in an otherwise clean speaker note. The workflow's purpose is to raise the claims to the surface, not to hide them in fluency.

This is the same accountability structure that runs through every lesson in this program, applied to medical affairs. The named author owns the gap between a claim that looks compliant and one that is compliant, and the MLR reviewers own the determination of which is which. The MSL who treats the model's polished deck as MLR-ready has misunderstood the task; the MSL who treats the model's output as a structured, pre-flagged input that makes the reviewers faster has understood it exactly. Compliance is the product, the reviewers are the judges, and the model is the accelerant that serves them, never the authority that replaces them.

Documenting the Refresh So the Annotation Is Defensible

A congress deck cleared through MLR and pushed into Veeva PromoMats carries an audit trail of its own, and the AI-assisted refresh inherits the obligation to make the AI involvement reconstructable within it. The defensible record captures the model and version, the loaded sources, principally the approved label, the existing deck, and the competitor's published data, the prompt, the timestamp, and the final reviewer-approved deck and annotation, linked to the PromoMats record and its review history. If a question ever arises about how a particular comparative claim entered a slide, the record must show whether it was a model-generated candidate that MLR adjudicated and approved, or an unreviewed model assertion that slipped through, because only one of those is defensible.

The point that ties this lesson to the rest of the medical-affairs chapter is that the MLR review itself is the irreplaceable human contribution, and the documentation must preserve that the review happened on the substance and not merely on the formatting. A claim-reference annotation that the model generated and that MLR rubber-stamped without verifying each link is the medical-affairs version of automated negligence, the same failure named in the literature-surveillance and signal-detection lessons, now operating on the promotional record that an advertising regulator can request. The model produces the reviewable deck; the reviewers verify each claim, each reference, and each regulatory consideration; and the record shows, claim by claim, that the judgment was theirs. That is what makes an AI-accelerated congress refresh defensible rather than merely fast.

There is a version-control dimension that the PromoMats record makes concrete and that the AI refresh makes easy to mishandle. A congress deck lives through cycles of review, approval, use at a meeting, and re-approval when the data or the label change, and each approved version is a controlled record tied to the specific claims and references that cleared review on that date. When the model refreshes the deck, it produces a new candidate version, and the discipline is that the refreshed deck re-enters review as a new version rather than silently inheriting the prior version's approval, because the changed slides carry new claims that the prior approval never covered. The failure to guard against is the speed of the refresh outrunning the review cycle, where a deck updated in an hour is used at a meeting before MLR has cleared the new comparative claim, which converts a fast refresh into an unapproved promotional use. The record must show that the version used was the version approved, with the AI-assisted changes captured in the review history.

Key Takeaways

• In congress and promotional materials, compliance is the product, not a finishing step. Every claim must be truthful, non-misleading, fairly balanced, substantiated, and consistent with the approved label, and the model writes persuasively, which makes unsubstantiated comparative and off-label claims read like settled fact. The MLR reviewers own the judgment the model cannot make.
• The three deliverables fail in three ways. The speaker notes concentrate the comparative and off-label claim risk, the reference list carries the fabricated or non-supporting citation risk, and the MLR annotation can assert substantiation that does not hold; each requires verification, not acceptance.
• On-label is a regulatory judgment about the claim and the approved indication, not a phrasing choice. A claim can be scientifically true, well supported, and still off-label if it touches an unapproved population, line, dose, or endpoint; the model collapses true and permitted, and the MSL must hold them apart and route the difference to MLR.
• Make the deck maximally reviewable, not maximally finished. A deck with every candidate claim surfaced, every reference attached, and every comparative or off-label-adjacent claim pre-flagged is a better input to review than a polished deck that smoothed those issues into confident prose; the smoothing is the danger and the surfacing is the value.
• Document the refresh so the annotation is defensible. Capture the model, the loaded sources including the approved label, the prompt, and the final reviewer-approved deck linked to the PromoMats record; an MLR annotation rubber-stamped without verifying each claim-reference link is automated negligence on a promotional record an advertising regulator can request.