An FDA investigator sits down in the sponsor's regulatory operations room on day two of a Pre-Approval Inspection, opens a laptop, and asks a question that did not appear in any inspection two years ago: "Show me how the Module 2.7.3 efficacy summary was drafted, and show me the audit trail for the AI tool that drafted it." The room goes quiet, because the answer lives in seven different systems, the AI use-log was reconstructed after the fact, and no one in the room can produce the validation record for the Certara CoAuthor deployment on demand. That silence is the gap this lesson closes. As an AI Function Strategist you are not the writer who used the tool; you are the person who has to make the function inspection-ready before the investigator arrives, so that the answer to "show me how the AI drafted this" is a single, calm, rehearsed walk through a controlled record rather than a scramble. This lesson maps exactly what FDA and EMA inspectors are asking about AI in 2026 across the Pre-Approval Inspection, the EMA GCP and GMP inspection, and the FDA Form 483 to Establishment Inspection Report cycle, and it gives you the architecture to be ready before the request lands.

What Changed in the Inspection Room

For most of the last decade an FDA Pre-Approval Inspection focused on data integrity, the manufacturing process, and whether the application's claims were substantiated by the records on site. AI did not feature, because AI was not visibly in the workflow that produced the submission. That has reversed. By 2026 the investigator arriving for a PAI knows that a meaningful fraction of the Module 2 narrative, the CMC stability discussion, and possibly the ICSR narratives were drafted with generative AI, and the investigator's training now includes how to probe that. The shift is not that AI is prohibited; it is that AI use is now a documented system like any other computerized system, and an undocumented system in a regulated workflow is itself the finding. The strategist's job is to ensure that when the question comes, the function answers it as a matter of routine rather than as a crisis, because the difference between those two responses is the difference between a clean Establishment Inspection Report and a Form 483 observation that follows the application into the review.

The three inspection contexts you must architect for are distinct and you cannot treat them as one. The FDA Pre-Approval Inspection is application-specific: it is tied to a pending NDA or BLA and it asks whether the data and the documentation behind this submission are trustworthy. The EMA GCP and GMP inspections, conducted by national competent authority inspectors coordinated through EMA for centralized procedures, ask whether the trial conduct and the manufacturing comply with the standards, and they increasingly ask whether the AI tools touching trial data or batch records are validated and controlled. The Form 483 to EIR cycle is the closeout mechanism for an FDA inspection: the Form 483 lists the investigator's observations, and the Establishment Inspection Report is the agency's full narrative account of the inspection that drives the classification. Each of these has its own AI-relevant questions, and a function that is ready for one is not automatically ready for the others.

The Seven Questions the PAI Investigator Asks About AI

From the pattern of 2025 and 2026 inspections and the FDA-EMA Guiding Principles released on 14 January 2026, a recognizable set of investigator questions has emerged, and you should architect your evidence to answer each one without improvisation. First: which AI tools touched the content in this submission, and what is the inventory? An investigator who finds an AI tool in use that is not on your computerized system inventory has found a governance failure before they have read a single output. Second: what is the validation status of each tool, and can you show the validation package? Third: what is the intended use statement for the tool, and does the actual use match it, because a tool validated for drafting that is being used to make benefit-risk conclusions is operating outside its qualified envelope. Fourth: where is the audit trail for a specific piece of AI-generated content, and does it capture the model, version, prompt, sources loaded, and the human who verified and signed?

The remaining questions go to the heart of how the function manages learning systems and accountability. Fifth: if the tool learns or is updated, what is the change control, and is there a Predetermined Change Control Plan or an equivalent that governs how model updates are assessed before they reach production? Sixth: how do you know the AI did not introduce an error into this specific submission, meaning what is the human verification record and the claim-reconciliation evidence for the content the AI produced? Seventh: who is accountable, named, for the AI-assisted content, because the FDA-EMA principles' accountability principle is explicit that accountability does not transfer to the vendor and a sponsor cannot answer "the AI did it." A strategist who has pre-built a clean answer to each of these seven questions, with the supporting artifact identified and retrievable, has converted the most dangerous moment in the inspection into a demonstration of control. The artifacts that answer them are the AI tool inventory, the validation package, the intended-use statement, the per-document audit trail, the change-control record, the verification log, and the named-accountability map.

EMA GCP and GMP Inspections: The European Angle

The European inspection posture differs in emphasis and you must architect for it separately even though the underlying AI controls overlap. A GCP inspection coordinated for an EMA centralized procedure examines trial conduct, and where AI touched the trial, such as AI-assisted central monitoring signal triage under ICH E6(R3) or AI-drafted monitoring visit reports, the inspector asks whether the AI was validated, whether it influenced decisions about data, and whether the sponsor retained oversight. A GMP inspection examines manufacturing and quality systems under EU Annex 11 for computerized systems, and where AI touched batch records, stability narratives, or deviation handling, the inspector applies the Annex 11 validation and audit-trail expectations directly. The European framework leans on EU Annex 11 and GAMP 5 categorization in a way that maps cleanly onto AI tools once you have placed each tool in its GAMP category, which is itself an L4 validation subject, so the strategist who has done that categorization arrives with the European inspector's mental model already satisfied.

The European context also runs on named regulatory windows that the strategist must respect when an inspection intersects with an ongoing centralized review. For an EMA centralized procedure the assessment runs to the Day 120 list of questions and the Day 180 list of outstanding issues, and an inspection finding about an AI tool that contributed to the dossier can become an outstanding issue that the sponsor must resolve before the procedure proceeds. The EMA AI Workplan is the multi-year vehicle through which European expectations on AI are being shaped, and the EMA Innovation Task Force is the early-engagement channel, both of which a strategist should be tracking so that the function's controls anticipate rather than lag the European direction. The practical discipline is that your AI controls must be defensible to a national competent authority inspector applying Annex 11, not only to an FDA investigator applying 21 CFR Part 11, and the union of those two expectations is the bar you build to.

The Form 483 and the EIR Classification

The Form 483 is the list of inspectional observations the FDA investigator issues at the close of an inspection, and it is the single document that most shapes what happens next, so understanding its mechanics is central to inspection-readiness strategy. A Form 483 observation about AI in 2026 typically reads as a computerized-system or data-integrity observation: an AI tool used in a GxP workflow without documented validation, an audit trail that does not capture which model version produced a piece of content, an AI output incorporated into a submission without a documented human verification step, or a learning system updated in production without change control. The FDA recommends that a sponsor respond to a Form 483 within 15 business days of the inspection close-out, because a response received within that window is positioned to be considered before the agency finalizes the Establishment Inspection Report and assigns the classification. A response that arrives later is still considered, but it may not factor into the EIR classification, which is why the 15-business-day discipline is a strategic deadline and not merely a courtesy.

The Establishment Inspection Report is the investigator's full narrative account of the inspection, written after the inspection closes, and it carries the classification that determines the consequence: No Action Indicated, Voluntary Action Indicated, or Official Action Indicated. For a Pre-Approval Inspection, an Official Action Indicated classification can delay or block the approval of the pending application, which is why an AI-related 483 observation is not a quality-department problem to be handled after the fact; it is a submission-jeopardy problem that the function strategist must prevent or, failing that, respond to with a CAPA that is credible within the 15-business-day window. The strategist's contribution to the 483 response is the systemic framing: an investigator who sees one undocumented AI tool will ask whether the whole function lacks AI governance, so the response must demonstrate that the observed gap is being closed inside a governance system that already exists, not invented in reaction to the observation. That is why the risk register, the governance charter, and the validation framework are inspection assets and not just management documents.

Building the Inspection-Readiness Binder for AI

The deliverable that converts strategy into inspection survival is an AI inspection-readiness binder, and as strategist you own its existence even if the quality function maintains it. The binder is not a single document; it is a curated, retrievable set of the seven artifacts that answer the investigator's seven questions, indexed so that any one of them can be produced in minutes during an inspection rather than reconstructed over days. At the top sits the AI tool inventory, naming every AI tool touching regulated content, its GAMP category, its intended-use statement, its validation status, and its accountable owner. Beneath each inventory entry sits the validation package, the change-control history including any Predetermined Change Control Plan, the audit-trail specification, and a representative per-document trace showing the model, version, prompt, sources, and human verifier for a real piece of submitted content. The binder is the physical embodiment of the claim that the function controls its AI, and an inspection-readiness binder that is current is worth more than any policy statement, because the investigator tests records, not intentions.

The binder must be living, not a pre-inspection artifact assembled in panic when the FDA gives notice of a PAI. The discipline that keeps it current is to tie each binder element to the operational system that generates it: the inventory updates when a tool is onboarded through governance, the validation package updates when a tool is re-validated under the change-control trigger, and the per-document traces are sampled continuously from the production audit trail rather than constructed retrospectively. A strategist who has wired the binder to live systems can answer the inspector's request as a query rather than a project, and that single capability, the ability to retrieve rather than reconstruct, is the operational signature of a function that is genuinely inspection-ready rather than inspection-anxious. The retrospective scramble is itself a tell that an investigator reads: if the audit trail had to be assembled after the request, it was not contemporaneous, and contemporaneousness is an ALCOA+ attribute the investigator is specifically trained to test.

The PCCP as an Inspection Instrument for Learning AI

The hardest AI question an investigator can ask is about a tool that changes, because a learning or frequently updated AI system challenges the traditional validate-once model that regulated computerized systems were built around. The FDA's Predetermined Change Control Plan, finalized for AI-enabled device software functions and now adopted by industry as the working pattern for any learning AI in regulated workflows, is the instrument that answers this question, and the strategist should treat it as an inspection artifact and not only a device-regulatory concept. A PCCP describes in advance the modifications a system is expected to undergo, the protocol by which those modifications will be assessed, and the impact assessment that determines whether a given change stays inside the qualified envelope or triggers re-validation. When an investigator asks how you know that a model update last quarter did not degrade the quality of the Module 2.7.3 narratives, the PCCP is the document that shows you anticipated the update, bounded it, and assessed it before it reached production.

For AI tools that are not formally SaMD, which describes most generative AI used in submission drafting, the strategist adopts the PCCP framework conceptually rather than as a regulatory filing, building an internal change-control plan that names the acceptance criteria for tool performance, the drift signals that trigger review, and the assessment that gates a model version into production. This matters because vendors update their models, and a sponsor who lets a vendor model update flow silently into a GxP workflow has lost change control over a validated system, which is a direct Annex 11 and Part 11 problem. The strategist's defensible position is that every model version reaching production passed through a documented assessment against pre-stated acceptance criteria, and that the log of those assessments is part of the inspection-readiness binder. An investigator who sees that pattern sees a function that has solved the learning-system problem rather than one that has ignored it, and that perception shapes the EIR classification as much as any single output does.

Rehearsing the Walkthrough Before the Investigator Arrives

Inspection readiness is a performance as much as a record set, and the strategist who has not rehearsed the AI walkthrough has not finished the job. The rehearsal is a mock inspection in which a senior person plays the investigator and asks the seven questions, and the function demonstrates that it can retrieve the inventory, open a validation package, produce a per-document audit trail, and name the accountable human, all within the time an actual investigator would tolerate. The rehearsal surfaces the gaps that the binder hides on paper: the inventory entry that points to a validation package no one can find, the audit trail that captures the prompt but not the model version, the tool whose intended-use statement was written for drafting but whose actual use has drifted into conclusion-making. Finding those gaps in a rehearsal is cheap; finding them when the investigator finds them is a Form 483 observation and a submission delay.

The rehearsal also trains the people who will be in the room, because the wrong answer delivered confidently is more damaging than a brief pause to retrieve a record. The person who says "the AI handles that" has just told the investigator that the function does not understand its own accountability, and a single answer like that can reframe an entire inspection from routine to adversarial. The rehearsed answer is calm, specific, and grounded in the artifact: "That section was drafted with our validated tool, here is the intended-use statement, here is the audit trail for this document, and here is the named author who verified every claim against the source." The strategist who has built the binder, wired it to live systems, adopted the PCCP pattern for learning tools, and rehearsed the walkthrough has done the entire job, which is to make the most feared question in the 2026 inspection room into the most controlled moment in it.

Key Takeaways

• AI use in a regulated workflow is now a documented computerized system, and an undocumented AI tool is itself the inspection finding. The FDA Pre-Approval Inspection, the EMA GCP and GMP inspections, and the Form 483 to EIR cycle each probe AI differently, and a function ready for one is not automatically ready for the others. Architect the evidence for all three before the investigator arrives.
• Pre-build clean answers to the seven PAI questions: tool inventory, validation status, intended-use match, per-document audit trail, change control, human verification, and named accountability. Accountability never transfers to the vendor under the FDA-EMA principles, so "the AI did it" is not an answer a sponsor can give.
• The Form 483 response window is a strategic deadline: the FDA recommends responding within 15 business days of inspection close-out so the response can be considered before the Establishment Inspection Report is finalized and the classification (NAI, VAI, or OAI) is assigned. For a PAI, an OAI classification can delay or block the pending application's approval.
• The inspection-readiness binder must be living and wired to production systems, so the answer to "show me the AI audit trail" is a query, not a reconstruction project. Reconstructing an audit trail after the request fails the ALCOA+ contemporaneousness attribute that investigators are specifically trained to test.
• Adopt the Predetermined Change Control Plan framework as the inspection instrument for learning and frequently updated AI, even for non-SaMD tools. It proves that every model version reaching production passed a documented assessment against pre-stated acceptance criteria, solving the learning-system problem that defeats validate-once thinking, then rehearse the walkthrough so the answer is calm, specific, and artifact-grounded.