On Day 74 of an NDA review cycle, the FDA Office of New Drugs sends a one-paragraph Information Request: the agency asks the sponsor to reconcile the overall survival result stated in the Module 2.5 Clinical Overview against the corresponding analysis in the pivotal study's Clinical Study Report, and to clarify the censoring rule applied at the data cutoff. It is a routine IR, the kind that arrives by the dozens across a review, and the sponsor's regulatory lead has perhaps thirty days to respond. She opens the enterprise AI, points it at the question, and within minutes has a draft that retrieves the relevant CSR efficacy section, summarizes the censoring approach, and assembles a clean response narrative with a proposed Module 1.11 placement. The draft is fast, well-structured, and exactly the kind of acceleration that makes an IR-heavy review survivable. It is also the moment where a sponsor most needs discipline, because an IR response is a formal communication to the agency that becomes part of the application's permanent record, and a response that subtly misstates the censoring rule, or cites a CSR section that does not say what the draft claims, does more damage than no response at all. This lesson is about using AI to draft health authority correspondence and IR responses well: retrieving the right source content, drafting a response that is true to it, sequencing it correctly into Module 1.11, and respecting the time clocks that govern when each kind of IR must be answered.

What an Information Request Is, and Why It Is High-Stakes Correspondence

An Information Request during an FDA review is the agency's mechanism for asking the sponsor to clarify, reconcile, or supplement something in the application without issuing a formal deficiency. During an Office of New Drugs review of an NDA or BLA, IRs arrive throughout the cycle, and the points where the agency communicates are structured around internal milestones: the Day 74 letter near the start of the review communicates the filing review's findings and any planned review issues, and IRs continue through the mid-cycle and late-cycle communications. The sponsor's responses to these IRs are not informal emails; they are formal correspondence that enters the application's record, is read by the review team, and is relied upon in the eventual action. A response that is accurate, responsive, and well-sourced advances the review; a response that is wrong, evasive, or internally inconsistent with the rest of the dossier creates a new problem the sponsor then has to unwind, often under a tighter clock than the original IR allowed.

The reason this is the right place for AI assistance, and the right place for caution, is the same reason: an IR response is a retrieval-and-synthesis task over a large, already-filed body of content. The answer to "reconcile the OS result in the 2.5 against the CSR" lives in documents the sponsor already submitted, and finding the relevant CSR section, the relevant Module 3 content, or the relevant Statistical Analysis Plan passage is exactly what retrieval-augmented AI does well. But the response is a claim about what those documents say, made to a reviewer who will check it against those same documents, so the cost of a retrieval error or a confident misstatement is immediate and visible. The AI that drafts the response is operating on the sponsor's behalf in the most scrutinized correspondence channel the sponsor has, and the named regulatory professional who signs the response owns every word of it.

Retrieving the Right CSR and Module 3 Content

The first job in an AI-assisted IR response is retrieval: pulling the specific source content the IR is asking about out of the filed dossier and into the model's working context. This is where retrieval-augmented generation earns its keep, because the relevant material, a CSR efficacy section, a Module 3.2.S specification, a SAP censoring rule, sits inside hundreds or thousands of pages, and a writer hunting it by hand is slow and prone to grabbing the wrong table. A retrieval layer that indexes the filed submission and pulls the passages most relevant to the IR gives the model the actual source text to ground its response in, which is the difference between a response that quotes the real censoring rule and one that invents a plausible-sounding one.

The discipline in retrieval is that the response can only be as good as what was retrieved, and the model does not know what it failed to retrieve. If the IR concerns a censoring rule that is stated in the SAP but the retrieval pulled only the CSR efficacy narrative, the model will draft a response about censoring from the narrative's summary description rather than the SAP's operative definition, and the two may differ in a way that matters to a statistical reviewer. The same silence-not-a-flag property that governs any context window governs retrieval: a relevant source that was not retrieved is invisible to the model, and the response will read as complete whether or not the operative source was in front of it. The verification step therefore includes confirming that the response was grounded in the authoritative source for the specific question, the SAP for a censoring rule, the CSR table for an efficacy number, the Module 3 specification for a quality question, and not merely in whatever the retrieval happened to surface.

There is a second retrieval discipline specific to IR responses: the response must be consistent not only with the source the IR points to but with everything else the sponsor has filed. An IR asking about the OS result is implicitly asking the sponsor to confirm that the 2.5, the Module 2.7.3, the CSR, and the SAP all tell one coherent story about that result, and an AI that drafts a response reconciling the 2.5 to the CSR while ignoring a different statement in the 2.7.3 has solved the IR narrowly and created a new inconsistency the reviewer may catch. The strongest AI-assisted responses retrieve across the relevant modules and surface any internal disagreement to the human before the response goes out, turning the IR into an occasion to confirm dossier-wide consistency rather than to patch one paragraph.

Drafting a Response That Is True to the Source

With the right content retrieved, the drafting job is to produce a response that answers the question precisely, states what the source says without overstating or understating it, and does not introduce any new claim that is not grounded in the filed record. The structure of a good IR response is disciplined: it restates the agency's question, provides the direct answer, supports the answer with specific cross-references to the source location in the dossier, and stops. The AI is genuinely good at this structure, and at maintaining the measured, non-defensive tone that regulatory correspondence requires. The risk is the same risk this program traces through every drafting task: the model writes a confident, fluent answer whether or not the answer is faithful to the source, so a response that says the censoring rule was "administrative censoring at the data cutoff" reads exactly as authoritative whether that is what the SAP actually specifies or a plausible paraphrase that drops a clinically meaningful detail.

The verification posture for an IR response is claim-by-claim reconciliation against the retrieved source, identical to the posture the program teaches for a 2.5.4 efficacy section, but with a sharper edge because the audience is a reviewer who requested this exact clarification and will read it adversarially. Every factual statement in the response, every number, every characterization of a method, every cross-reference, is reconciled to the filed source before the response is signed. A cross-reference in an IR response that points to a CSR section which does not contain what the response says it contains is the same fabricated-cross-reference failure that the program opens with, now committed directly into formal correspondence with the agency, where it is not caught at reference QC but read by the reviewer who asked the question. The drafting acceleration is real and valuable; the reconciliation is non-negotiable, and the named author owns the gap between a fluent response and a faithful one.

Module 1.11 Sequencing and Getting the Response Into the Record Correctly

An IR response is not just a letter; it is a submission, and it has to be placed correctly in the eCTD so the review team finds it where they expect it. Module 1.11 of the US regional eCTD holds responses to agency correspondence and information requests, and an IR response sequence typically carries the response letter in Module 1.11 along with any supporting content placed in its proper module, a revised analysis in Module 5, an updated specification in Module 3, an updated summary in Module 2, cross-referenced from the response. The sequencing job is to assemble the response so that the cover correspondence in Module 1.11 points cleanly to the supporting content wherever it lives, the Form 1571 or the appropriate transmittal carries the correct purpose, and the whole thing validates and lands in the review team's queue as a recognizable, well-formed response to their specific request.

AI assists this sequencing the way it assists any submission-operations task: it can map the response content to the correct eCTD locations, check that cross-references between the Module 1.11 letter and the supporting modules resolve, and confirm the sequence declares the right purpose. The limit is the same as in the form-drafting lesson: the AI reconciles the sequence against the content it is given, and the named regulatory professional owns whether the response is complete, whether it actually answers the agency's question, and whether placing a revised analysis in Module 5 triggers any consistency obligation elsewhere in the dossier. A well-sequenced response that answers the wrong question, or that introduces a revised number in Module 5 without updating the corresponding statement in the Module 2 summary, is a sequencing success and a substantive failure, and only the human can tell the difference.

The IR Time Clocks: Routine, Expedited, and the RTF Conference Window

The single operational fact that shapes every IR response is the clock, and the clocks differ sharply by IR type, so an AI workflow that treats all IRs the same is dangerous. A routine Office of New Drugs IR during a standard review typically carries a response expectation on the order of thirty days, enough time to retrieve, draft, reconcile, and sequence properly, though the exact window is set by the IR and the review team. An expedited discipline-review IR, the kind a statistical or clinical reviewer issues when a specific question is blocking their assessment, can carry a far tighter window, sometimes on the order of five to ten business days, which compresses the entire retrieve-draft-reconcile-sequence cycle and raises the value of AI acceleration precisely when the time pressure makes hand-drafting infeasible. The workflow has to be clock-aware, because a response that is accurate but late can be as damaging as one that is on time but wrong.

A distinct and frequently confused clock governs the refuse-to-file situation. When FDA issues a refuse-to-file letter, the sponsor has thirty days from the RTF notification to request an informal conference under FDA SOPP 8404, and that conference request and the supporting gap-closure reasoning are a different species of correspondence from a routine IR response: the sponsor is contesting or addressing a filing decision, not answering a clarification within an ongoing review. The thirty-day RTF informal-conference window is commonly conflated with the routine IR response window and with the IND safety-report clocks from the prior lesson, and keeping them distinct is part of the clock-awareness this workflow demands. AI can help track which clock applies to which piece of correspondence, draft the appropriate response type for each, and flag when a draft response is being prepared against the wrong clock, but the regulatory professional owns the determination of which clock governs and the decision of what posture the response should take, especially in the RTF case where the wording carries strategic weight.

The Response That Creates a New Problem, Traced Forward

Return to the Day 74 IR about the OS result and the censoring rule, and trace what happens when the AI-drafted response goes out without full reconciliation. Suppose the retrieval pulled the CSR efficacy narrative but not the SAP, and the model drafted a response describing the censoring as a clean administrative censoring at cutoff, when the SAP actually specified a more nuanced rule that censored certain subjects at their last adequate assessment. The response is fluent, well-sequenced into Module 1.11, and submitted on time. The statistical reviewer who asked the question reads the response, checks it against the SAP they already have, and finds that the sponsor's own response misdescribes the sponsor's own censoring rule. Now the IR has not been closed; it has been reopened, with a new and worse question attached: why does the sponsor's formal correspondence not match the sponsor's filed SAP, and what else in the application has the same problem.

The cost compounds the way every uncaught AI error in this program compounds. The reviewer's trust in the sponsor's responses drops, so subsequent IR responses are read more skeptically and the review slows. A follow-up IR now consumes a clock that may be tighter than the first. And the sponsor's team spends time it does not have reconstructing what the censoring rule actually was and drafting a correction, under exactly the time pressure the original AI acceleration was supposed to relieve. The lesson is not that AI should be kept out of IR responses; AI-assisted retrieval and drafting are what make an IR-heavy review survivable, and a team that hand-drafts every response will fall behind. The lesson is that the acceleration only pays if the reconciliation discipline is applied: retrieve from the authoritative source, ground every claim in it, reconcile every cross-reference against the filed record, sequence it correctly into Module 1.11, answer against the correct clock, and have the named regulatory professional verify and sign before the response enters the permanent record. The fast draft is the start of the response, not the end of it.

Key Takeaways

• An Information Request is high-stakes formal correspondence that enters the application's permanent record and is read adversarially by the reviewer who asked it. During an Office of New Drugs review, IRs cluster around milestones like the Day 74 communication, and a wrong or inconsistent response creates a new problem the sponsor must unwind under a tighter clock than the original.
• Retrieval is the first job, and the response is only as good as what was retrieved. A relevant source that was not retrieved is invisible to the model, so the verification step confirms the response was grounded in the authoritative source for the specific question, the SAP for a censoring rule, the CSR table for an efficacy number, not merely in whatever the retrieval surfaced.
• The response must be true to the source and consistent across the whole dossier. Every number, method characterization, and cross-reference is reconciled to the filed record, because a fabricated cross-reference in an IR response is the program's opening failure committed directly into agency correspondence, where the reviewer catches it rather than reference QC.
• Module 1.11 sequencing places the response where the review team expects it, with supporting content in its proper module. AI maps content to eCTD locations and checks that cross-references resolve, but the named professional owns whether the response answers the agency's actual question and whether a revised number triggers a consistency obligation elsewhere in the dossier.
• The clocks differ by IR type and must not be conflated: a routine OND IR runs on the order of thirty days, an expedited discipline-review IR can be five to ten business days, and the RTF informal-conference request is thirty days from the RTF notification under SOPP 8404. AI can track and flag the applicable clock, but the regulatory professional owns which clock governs and the posture the response takes.