A medical information specialist opens her queue and finds the same question she has answered eleven times this month, phrased eleven different ways: a physician wants to know whether the company's biologic has data in patients with moderate renal impairment, a population the label does not specifically address. The right answer is not a marketing message and not a clinical opinion; it is a balanced, evidence-based, non-promotional account of what the published and on-file data actually show, delivered the same way to every requester who asks. That account is a Standard Response Document, the SRD, and it is the backbone of a compliant medical information function. An enterprise large language model can convert a pile of publications and clinical data into a first-draft SRD in minutes, which is genuinely useful when the queue is deep and the inquiries keep arriving. It can also, in the same minutes, blur the line between responding to an unsolicited inquiry and promoting an unapproved use, attach a claim to a source that does not support it, and present off-label data as though it were on-label. The non-promotional compliance of the final document is owned by a human, and this lesson is about why that ownership cannot move.

What an SRD Is and Why It Exists

Medical information sits in a specific and carefully drawn legal space. When a healthcare professional, on their own initiative, asks a manufacturer a specific medical question, including a question about an unapproved use, the manufacturer may provide a truthful, balanced, non-misleading, scientific response. This is the unsolicited-request framework, and it is the channel through which off-label scientific information lawfully reaches a clinician, because the clinician asked and the manufacturer did not promote. The Standard Response Document is the pre-prepared, reviewed, and approved answer to a recurring inquiry, so that every requester receives the same vetted, balanced content rather than an improvised answer that drifts toward promotion or away from accuracy.

The SRD therefore carries two jobs at once. It is a scientific document, summarizing the relevant evidence accurately with full attribution, and it is a compliance artifact, structured to stay inside the unsolicited-response framework that keeps it lawful. The structure reflects both jobs: a clear statement of the question, a balanced presentation of the available data with citations, explicit flagging of what is within the approved label and what is outside it, and the standard non-promotional framing that medical information governance requires. A reviewer in the medical-legal-regulatory chain reads the SRD not for whether it sounds good but for whether it is balanced, accurate, attributed, and non-promotional, and whether it correctly distinguishes on-label from off-label content.

This is the frame the model does not bring on its own. To the model, an SRD is a document type with a recognizable shape, and it can reproduce the shape fluently. It does not understand that the shape exists to keep a manufacturer on the lawful side of a line that, if crossed, turns a helpful scientific response into off-label promotion. The named medical information professional owns that line, exactly as a medical writer owns the truth of a Clinical Overview and a medical affairs lead owns the confidentiality of an advisory board. The model accelerates the drafting; the human guarantees the document is non-promotional, balanced, and correctly labeled.

From Published Evidence to a Standardized Answer

The raw material of an SRD is evidence: the relevant sections of the prescribing information, the pivotal and supporting clinical study data, the published literature, and any on-file data the company holds. The drafting task is to convert that evidence into a standardized, readable answer to the specific recurring question, organized so a busy clinician can find what they need and a reviewer can confirm every statement against its source. An LLM is well suited to the synthesis layer of this work: given the loaded evidence, it can produce a clean, well-structured draft that states the question, summarizes the relevant findings, and organizes them in the standard SRD format far faster than starting from a blank page.

The acceleration is real and the structure is consistent, but the same mechanic that makes the model fast makes it dangerous in this context: it generates the most plausible continuation of an evidence summary, not a verified transcription of the evidence. Asked about renal-impairment data, the model can produce a confident statement about a pharmacokinetic finding, a dosing consideration, or a safety signal that reads exactly like a summarized result and may be a generated one. In a Clinical Overview this is a hazard-ratio risk; in an SRD it is the same risk wearing different clothes, a plausible clinical statement that a clinician may act on, sent to every requester who asks the question, with the manufacturer's name on it. The control is the program's control: every factual statement in the SRD traces to a loaded source, and an uncited claim is treated as unverified until a human reconciles it.

There is a standardization discipline specific to the SRD that the model can quietly undermine. The point of a Standard Response Document is that the same vetted content goes to everyone, so a model that subtly rephrases the approved content, introduces a new emphasis, or adds a sentence of interpretation has broken the standardization even if every individual sentence is defensible. When the SRD already exists and the task is updating it with new evidence, the discipline is to integrate the new data without silently rewriting the approved, reviewed content around it, because the reviewed content is reviewed for a reason. The human keeps the approved core stable and treats the model's additions as proposals to be checked, not as improvements to be accepted by default.

Literature Attribution Is the Load-Bearing Control

An SRD lives or dies on its attribution. Every claim about what the evidence shows must point to the specific source that shows it, the named publication, the specific study, the section of the prescribing information, so that a clinician can evaluate the evidence and a reviewer can verify it. This is not a stylistic preference; it is the mechanism by which the SRD stays scientific rather than promotional, because a balanced presentation of attributed evidence is a fundamentally different act from an unattributed claim that reads as a manufacturer assertion. Attribution is what lets the document be checked, and a document that cannot be checked cannot be approved.

This is exactly where the model's failure mode is most consequential. A large language model produces citations the way it produces prose, by pattern, and it can attach a real-looking citation to a claim the cited source does not actually support, or invent a publication that does not exist, or cite a real paper for a finding that appears nowhere in it. In a medical information context, a misattributed claim is worse than an unattributed one, because the false attribution lends the claim an authority it has not earned and a clinician may rely on a study that says no such thing. The verification step is non-negotiable and specific: every citation is opened and checked against the claim it supports, not merely confirmed to exist, because a real paper cited for a finding it does not contain passes an existence check and fails a substance check.

The substance check is the harder discipline and the one the model cannot perform for itself. It is not enough that the cited study is real and is about the right drug; the study must actually report the finding the SRD attributes to it, in the population and at the magnitude stated. A model summarizing a renal-impairment publication can shift a finding from one subgroup to another, generalize a narrow result, or soften a limitation the authors emphasized, all while keeping the citation attached. The human reconciles the claim against what the source actually says, treating the citation as a pointer to be followed rather than a credential to be trusted, and rejecting any statement the source does not support even when the citation is genuine.

The On-Label / Off-Label Line the Human Must Draw

The most compliance-critical feature of an SRD is the explicit distinction between what is within the approved labeling and what is outside it. A clinician asking about renal-impairment data deserves to know clearly which statements describe the approved use as reflected in the prescribing information and which describe data outside the label, because that distinction governs how they should weigh the information and governs whether the manufacturer is responding lawfully or promoting unlawfully. The SRD must flag off-label content as off-label, present it in a balanced way that includes limitations, and avoid any framing that encourages the unapproved use rather than informing about it.

This is a judgment the model is structurally unequipped to make reliably. Determining whether a given statement falls within the approved label requires reading the specific approved indication and labeling and reasoning about whether the data in question describes that use or a use beyond it, and a model generating fluent text will often present on-label and off-label findings in the same even register, with no signal that the line was crossed. Worse, a model optimizing for a helpful, complete answer can drift into framing that reads as encouraging the off-label use, because completeness and enthusiasm are patterns it has learned, and the difference between informing about off-label data and promoting it is a legal distinction the model does not represent. The human must read every statement against the label and confirm both that off-label content is flagged and that its framing stays informational.

The stakes of getting this wrong are not editorial. An SRD that presents off-label data without flagging it, or that frames it promotionally, can convert the entire unsolicited-response channel into off-label promotion, exposing the manufacturer to enforcement and undermining the legal basis on which the response was provided at all. This is why the on-label/off-label flagging and the non-promotional framing are owned by the human in the medical information and MLR chain, not delegated to the model. The model can draft the evidence summary; it cannot be trusted to draw the line that keeps the document lawful, and the human who signs the SRD draws that line deliberately, statement by statement.

Balance and the Non-Promotional Register

Beyond the on-label/off-label distinction, the SRD must be balanced and non-promotional in its overall character, which means it presents limitations, contrary findings, and uncertainty alongside favorable data, and does so in a register that informs rather than persuades. A balanced response to a renal-impairment question includes the studies that found a benefit and the ones that did not, the limitations of the available data, and the absence of evidence where evidence is absent, because a one-sided presentation of only the favorable findings is promotional regardless of whether each individual sentence is true. Balance is a property of the whole document, not of its sentences, and it is exactly the property a model optimizing for a clean, confident answer tends to erode.

The model's instinct, learned from a corpus full of persuasive and conclusive writing, is to resolve uncertainty into a clear takeaway, to lead with the favorable result, and to underweight limitations that complicate the narrative. None of these is a lie, and that is precisely why they are dangerous: a draft that omits the negative study, buries the limitation, and states a tentative finding with confident phrasing is more fluent and more satisfying than the balanced truth, and it reads as a better document while being a non-compliant one. The human reviews the SRD for balance as a whole, asking whether a reader would come away with an accurate sense of the evidence including its weaknesses, or with a favorably skewed impression the data does not support.

The non-promotional register is a related but distinct control. Even a balanced, accurate SRD can adopt phrasing that reads as marketing, superlatives, comparative claims, framing that positions the product, and that phrasing can recharacterize the document regardless of its factual content. The model, trained partly on promotional and persuasive prose, can introduce this register without intending to, and the human strips it out, holding the document to the flat, informational tone that medical information requires. The discipline is to read the SRD twice: once for whether each claim is true and attributed, and once for whether the document as a whole is balanced and non-promotional, because the second reading catches what the first does not.

The MLR Review and the Audit Trail

An SRD does not become usable because a specialist drafted it; it becomes usable when it has passed medical-legal-regulatory review and been approved into the medical information content library, typically managed in a governed system such as a Veeva MedComms or PromoMats environment. The MLR reviewers, medical, legal, and regulatory, each read the SRD for their dimension: medical for scientific accuracy and balance, legal for liability and the unsolicited-response framework, regulatory for label alignment and promotional risk. An AI-drafted SRD enters this review like any other, and the specialist who submits it is attesting that it is ready for that scrutiny, which means the verification work, the source reconciliation, the on-label/off-label flagging, the balance check, has to happen before submission, not be discovered in review.

The audit trail matters here for the same reason it matters in a submission. Because the SRD was AI-assisted and because output varies run to run, a defensible record captures how it was produced: the sources loaded, the model and version, the prompt, and the human verification steps that reconciled each claim, flagged the off-label content, and confirmed the balance and register. When a regulator or an internal auditor asks how the medical information function ensures its responses are accurate and non-promotional, the answer cannot be that a specialist trusted the tool; it must be a documented workflow in which a named human verified the document against its sources and against the compliance framework before approval. The AI use log is the evidence that the human did the work the model cannot do.

What This Means for the MI Specialist on Monday

The SRD workflow is a strong fit for AI assistance because it is repetitive, evidence-based, and structure-heavy, and a specialist facing a deep queue of recurring inquiries reclaims real time by letting the model produce the first-draft synthesis. But the time is reclaimed only inside a verification discipline that the model does not supply, and the discipline follows directly from what the document is for. The SRD exists to deliver balanced, attributed, correctly labeled, non-promotional scientific information through the narrow lawful channel of the unsolicited response, and every one of those properties is a human responsibility that the model can draft toward but cannot guarantee.

So the specialist loads the real evidence before drafting, because the model reasons only over what it is given. She reconciles every claim against its source and opens every citation to confirm the source supports the claim, not merely that it exists. She reads every statement against the label, flags off-label content as off-label, and confirms its framing informs rather than encourages. She reads the document as a whole for balance and for a non-promotional register, separately from checking each sentence. She keeps the approved standardized content stable when updating, and she captures the run so the audit trail shows the work. Then she submits to MLR, attesting that the document is ready. The model drafted the SRD in minutes; the named specialist made it lawful, and the non-promotional compliance of the final document never left her hands.

Key Takeaways

• An SRD is both a scientific document and a compliance artifact, lawful only inside the unsolicited-response framework. It exists so every requester receives the same vetted, balanced, non-promotional answer, including for off-label questions a clinician raises on their own initiative. The model reproduces the shape of an SRD but does not understand the legal line the shape exists to hold, and the named specialist owns that line.
• Every factual statement must trace to a loaded source, and standardized content must stay stable. The model generates plausible clinical statements that read like summarized results and may be invented, and when updating an SRD it can silently rewrite approved, reviewed content. The human treats uncited claims as unverified and the model's additions as proposals to be checked, not improvements to accept by default.
• Literature attribution requires a substance check, not an existence check. The model can attach a real citation to a claim the source does not support, cite a real paper for a finding it does not contain, or invent a publication. Every citation is opened and checked against the claim it supports, because a misattributed claim is worse than an unattributed one and a clinician may rely on a study that says no such thing.
• The on-label / off-label line is a human judgment the model cannot reliably make. The SRD must flag off-label content as off-label and present it informationally, never promotionally, because a model presents both registers evenly and can drift into framing that encourages the unapproved use. Getting this wrong can convert the entire unsolicited-response channel into off-label promotion.
• Balance and a non-promotional register are properties of the whole document, checked in a separate reading. A model resolves uncertainty into a clean takeaway, leads with favorable data, and underweights limitations, producing a more fluent but non-compliant draft. The specialist reads the SRD once for truth and attribution and again for balance and register, captures the run, and submits to MLR attesting it is ready.