AI in Biopharma Manufacturing: Are We There Yet?
This is the script the AI voices read, so it matches the audio word for word; times are from the render. Researched, scripted and voiced by AI systems under Jack Prior's direction. Sam and Sarah are AI characters; nothing in the episode is Jack speaking, and none of it is a statement of his views or his employer's. Generative AI can be confidently wrong — check the sources. Corrections: jack@jackprior.ai.
00:00 Narrator This is Are We There Yet — a podcast on the evolution of AI in biopharma manufacturing, directed by Jack Prior. A word about how it's made. This episode was researched, scripted and voiced by AI systems. Jack sets the questions and the frames; the AI reads the documents and writes the conversation you're about to hear, between two AI characters: Sam, who plays a manufacturing-science practitioner, and Sarah, who has read the documents. Nothing you hear is Jack speaking, and none of it is a statement of his views or his employer's. Like any generative AI output, it can be wrong — confidently wrong, or missing a nuance — which are exactly the risks this industry is working to mitigate, and exactly what this season is about. Check the sources before you rely on anything. Corrections are welcome at jackprior dot A I. Now, the episode.
00:53 Sam A biologics site, inspected by FDA and by a European authority in alternate years. The soft-sensor vendor ships a new release, and the headline feature is adaptive recalibration: every night the model re-fits itself on the latest batches. The data scientist has the slide ready; prediction error down by a third. The qualified person reads one paragraph of a draft annex and asks a question nobody in the room expected: is this now a dynamic model? Because if it is, the draft says it shouldn't be anywhere near a critical step, whatever the error bars say. Down the corridor, the digital team is piloting an assistant on a language model that drafts deviation investigations. Same paragraph. Same problem. And the plant hasn't changed a single process parameter. It changed a model type.
01:44 Sarah That's the shift tonight, Sam. For twenty years, which kind of model you built was an engineering choice. Under the European draft it is a compliance decision, made before the first line of code.
01:55 Sam And FDA doesn't see it that way at all. Let's read both.
01:58 Sam Sarah, the question.
01:59 Sarah Why is the model you choose now a compliance decision, and where do FDA and the EU disagree? Two sub-questions run through it. Can the model keep learning after it's deployed? And can it be a large language model? Last episode gave us the four tiers and said Annex Twenty-two is a gate you pass through after the tier. Tonight we read the gate.
02:21 Sam Why now, in one breath, because I've said my piece in episode one. The soft sensor, the batch model, the camera: all static, all deterministic, all fine under any version of this rulebook. The two things that arrived in the plant in the last two years are exactly the two things the European draft names and excludes: models that learn in use, and generative models. The rulebook's first binding-track text on AI draws its line precisely where the new technology stands. That's not a coincidence, and it's why this episode matters more to the agent than to anything else on the slide.
02:58 Sam Documents.
02:59 Sarah The anchor is draft Annex Twenty-two to the EU GMP guide, titled Artificial Intelligence. Layer one, law and regulation: an annex to EudraLex Volume Four, so when it is final it is part of the GMP rules, not guidance about them. Maturity: draft, published for consultation on the seventh of July twenty twenty-five, consultation closed the seventh of October. Per the landscape it was prepared by the GMP and GDP Inspectors Working Group at EMA and consulted in parallel with PIC/S, the Pharmaceutical Inspection Co-operation Scheme, which means it reaches inspectorates well beyond the EU. It is six pages. We read all of it.
03:42 Sam Voices.
03:43 Sarah Four. FDA's January twenty twenty-five draft on AI to support regulatory decision-making, regulator guidance, still a draft, for its scope carve-out and its section on lifecycle maintenance. FDA's device guidance on predetermined change control plans for AI-enabled device software, final in December twenty twenty-four and reissued in August twenty twenty-five, because the device side already has an instrument for models that change. EMA's reflection paper on AI in the medicinal product lifecycle, final in September twenty twenty-four, for what it says about generative and incremental learning. And NIST's generative AI profile, AI six hundred dash one, from July twenty twenty-four, standards layer, because it is the document that names the generative risks in plain words.
04:37 Sam Start where the draft starts. Scope.
04:39 Sarah Section one, first paragraph. The annex applies to computerised systems used in manufacturing medicinal products and active substances, quoting, where Artificial Intelligence models are used in critical applications with direct impact on patient safety, product quality or data integrity. And in the same breath it says it provides additional guidance to Annex Eleven for computerised systems in which AI models are embedded. So two things at once: it is about critical applications only, and it is a supplement to the computerised-systems annex, not a free-standing rulebook.
05:17 Sam And what kind of model?
05:19 Sarah Second paragraph: machine learning models, meaning models that obtained their functionality through training with data rather than being explicitly programmed. It adds that a model may consist of several individual models, each automating a specific process step. Then come the three paragraphs that make tonight's episode. Static. Deterministic. And not generative.
05:43 Sam Before we read them, the status. Where is this draft in its life?
05:46 Sarah Draft for consultation, July to October twenty twenty-five, and the landscape records roughly thirteen hundred comments received. EMA held a two-day workshop on the thirtieth of June and first of July twenty twenty-six that looked at exactly the excluded categories, adaptive and generative models, and the draft text was not amended afterwards. As I understand it, the final text is expected to go to the Commission around the end of twenty twenty-six, with an effective date around twenty twenty-seven. Treat those last two as estimates. What is not an estimate is that the text you can read today is the text an inspector has read.
06:28 Sam Which is the house rule. It's a draft; it's also the current thinking of the people who will walk your floor. Now the three paragraphs.
06:36 Sarah First, static. The document applies to static models, quoting, models that do not adapt their performance during use by incorporating new data. The next sentence: the use of dynamic models which continuously and automatically learn and adapt performance during use is not covered by this document, and, quoting the end of the sentence, should not be used in critical GMP applications.
07:01 Sam Read that structure again, because it's the whole point. Not covered, and should not be used. Those are two different statements.
07:08 Sarah They are. Not covered would mean the annex is silent and you reason from Annex Eleven. Should not be used is a position. In GMP drafting, should is the operative word throughout; the annex uses should for everything it expects. So a dynamic model in a critical application is not a gap in the rules. It is something the draft tells you not to do. And the glossary fixes the meaning of static: quoting, Frozen model: a model where all parameters have been finally set, not allowing further adaption to new data.
07:41 Sam So my nightly-recalibrating soft sensor from the cold open.
07:44 Sarah Adapts its performance during use by incorporating new data. Dynamic, by the draft's definition, and if the step is critical, should not be used. The accuracy improvement is not a factor the text recognises.
07:57 Sam Second paragraph. Deterministic.
08:00 Sarah The document applies to models with a deterministic output, quoting, which, when given identical inputs, provide identical outputs. Models with a probabilistic output, which given identical inputs might not provide identical outputs, are not covered and, again, should not be used in critical GMP applications. Note the test is behavioural: same input, same output. A classifier that emits a probability score is still deterministic in this sense if the score is the same every time you show it the same image. What the draft excludes is a model that samples, so that two runs on one input can disagree.
08:42 Sam Which is what a language model does by default. So here's the practitioner's dodge: set the sampling temperature to zero, pin the weights, and now identical inputs give identical outputs. Am I through the gate?
08:54 Sarah No, and the draft closes that door deliberately. The third paragraph opens with the words Following the above, and then, quoting, the document does not apply to Generative AI and Large Language Models, and such models should not be used in critical GMP applications. It excludes them by name, as a category, not by running them through the static and deterministic tests. A pinned, zero-temperature language model may pass both tests and it is still named out.
09:24 Sam So the gate has three bars. Two are properties you can test for, and one is a name. That's going to matter when we get to the agent.
09:31 Sarah And the same paragraph says what happens outside critical use. If generative models are used in non-critical GMP applications, defined as those which do not have direct impact on patient safety, product quality or data integrity, then, quoting, personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use. It names that a human-in-the-loop, and adds that the principles in the document may be considered where applicable.
10:01 Sam Always. Not risk-based, not where appropriate. Always a person responsible for the output.
10:05 Sarah Always, for generative models in non-critical use. And there's a second human-in-the-loop clause later, three point three, written for models in scope. Where a model gives input to a decision made by a human operator, and the effort to test the model has been diminished because of that, the intended use should include the operator's responsibility, and the operator's training and consistent performance should be monitored, quoting, like any other manual process. Clause ten point five adds that records should be kept of that review, and depending on criticality and how much the model was tested, that may mean review of every output, according to a procedure.
10:46 Sam Let me park something for episode seven, because it starts right here. Three point three assumes the operator catches what the model gets wrong. If what the operator is doing is checking the output against defined bounds, that's a rule, and the system should run it; the human is adding a signature, not control. If the operator is meant to stop something unexpected, that takes either process expertise, this looks physically wrong for this bioreactor, or model expertise, the model is misbehaving. The draft says operator, and adequate qualification, and stops. Which competence, evaluating which inputs? Parked. Episode seven.
11:27 Sarah Parked. And for tonight, note what the two clauses together do to a generative system: in critical use it is out; in non-critical use a qualified person always owns the output, and that person is a manual process with its own performance monitoring. The draft never gives a generative model a route into a critical step, with or without a human.
11:49 Sam Where we are. Annex Twenty-two is a draft annex to EU GMP, consulted with PIC/S, and its whole scope is static, deterministic machine learning in critical applications. Dynamic models, probabilistic models and generative models are told not to be used in critical applications, and generative models in non-critical use always get a qualified human owning the output. Thirteen hundred comments and a workshop, and no revised text yet — though EMA says it's revising. Now the other side of the Atlantic, which never says any of this.
12:23 Sarah FDA's January twenty twenty-five draft states its position on model type in its introduction, in one sentence: quoting, This guidance does not endorse the use of any specific AI approach or technique. That is the whole of it. Model type never appears as a gate anywhere in the document. The framework, as we read in episode two, grades the decision, and step four then asks you to describe whatever model you chose.
12:51 Sam And on models that keep learning?
12:53 Sarah Section four B, lifecycle maintenance, which it defines as managing changes to AI models whether incidentally or deliberately, so the model stays fit for use over the product lifecycle. It says lifecycle maintenance is important for AI in the manufacturing phase specifically. Then the sentence that sets up the divergence: AI-based models may be highly sensitive to changes in inputs because they are data-driven and can be, quoting, self-evolving, that is, capable of autonomously adapting without any human intervention. Performance metrics should be monitored on an ongoing basis, and, quoting again, sponsors should anticipate inherent, model-directed changes and the need to identify and evaluate those changes.
13:41 Sam So the same thing Europe says should not be used, FDA says plan for. Not permit; plan for. That's a different verb and a different posture.
13:49 Sarah And it says where the plan lives. Changes to the model, or changes in manufacturing that may affect the model, go through the manufacturer's change management system within the pharmaceutical quality system; it gives newly available manufacturing data and model-directed changes as examples. If a change affects performance, it should be reported to the Agency in accordance with regulatory requirements. Detailed lifecycle plans, with performance metrics, a risk-based monitoring frequency and triggers for retesting, sit in the site's quality system, with a summary in the marketing application for product or process-specific models. And it points to ICH Q12: sponsors may propose model-related elements as established conditions, with a plan to manage changes to them, so that some changes would not require submission before they are made.
14:45 Sam Which is the change episode's territory, so I'll leave the instruments there. But one more thing in FDA's text that speaks to model type.
14:52 Sarah Step seven. If credibility is not sufficiently established for the model risk, one of the five outcomes FDA lists is that the sponsor may change the modelling approach. So in FDA's framework model type is a lever you can pull after the evidence comes in. In Annex Twenty-two it is a gate you pass before the evidence is gathered. That is the divergence in one sentence: for FDA the model type is an answer to a credibility problem; for the European draft it is a precondition.
15:22 Sam And FDA has already built the instrument for a model that changes. On the device side.
15:27 Sarah The predetermined change control plan, or PCCP, in the guidance for AI-enabled device software functions, originally issued the fourth of December twenty twenty-four and reissued the eighteenth of August twenty twenty-five. Its scope statement is the interesting part for tonight. It applies to device software the manufacturer intends to modify over time, and that includes modifications implemented automatically by software, which it says is also known as continuous learning, modifications implemented manually with human review and decision, or both.
16:02 Sam So a continuously learning model is not only contemplated, it has a submission pathway.
16:07 Sarah Three components: a Description of Modifications, meaning the specifications for the characteristics and performance of the planned changes; a Modification Protocol, meaning the verification and validation activities with pre-defined acceptance criteria; and an Impact Assessment of the benefits and risks. An authorised plan becomes a technological characteristic of the authorised device, and modifications made within it do not trigger a new marketing submission. Two limits for our purposes. It is a device instrument; the guidance says it applies to the device constituent of a combination product and not to the drug or biologic constituent. And the plan is authorised with the device, so there is a regulator on the other side of the table before the first automatic update.
16:56 Sam So the transatlantic picture for a model that learns in use: Europe, should not be used in critical GMP; FDA drugs, anticipate it, monitor it, put the plan in the quality system and report performance-affecting changes; FDA devices, pre-authorise the envelope. Three postures for one model type. A global site lives under all three.
17:16 Sam Restate. Europe gates by model type before evidence: static, deterministic, not generative. FDA doesn't endorse or exclude any technique, tells you to anticipate self-evolving models under lifecycle maintenance, and lets you change the modelling approach if credibility falls short. The device world has a pre-authorised plan for continuous learning that doesn't reach drug CMC. Now the second sub-question. Can it be a language model? FDA first, because FDA's answer is silence.
17:49 Sarah The words generative and language model do not appear in FDA's draft. The only handle is the scope carve-out, lines forty-seven to fifty-one. The guidance does not address AI used, quoting, for operational efficiencies, and the examples in brackets are internal workflows, resource allocation, and drafting or writing a regulatory submission. But the carve-out has a condition attached: those uses are out of scope only where they do not impact patient safety, drug quality, or the reliability of results from a nonclinical or clinical study. And FDA adds that sponsors uncertain whether a use is in scope should engage early.
18:32 Sam So drafting is named as an example of what's out. And then the condition takes it back the moment the draft matters.
18:40 Sarah That's the structure. A language model drafting an internal memo is out. A language model whose drafted investigation shapes a batch disposition impacts drug quality, and is in. Once it is in, there is no generative-specific chapter to turn to; the seven steps apply as to any model. Step four does ask two things a foundation model forces on you: specify whether a pre-trained model was used, and if so, which dataset it was pre-trained on and how the model was developed or obtained; and if the context of use involves a human in the loop, evaluate the human-AI team, not the model alone.
19:23 Sam Notice the two criticality tests are almost the same sentence. FDA: patient safety, drug quality, reliability of results. Annex Twenty-two: patient safety, product quality, data integrity. Two regulators, one question, which is whether the model's output touches quality. And a different consequence when the answer is yes: FDA says now do the credibility assessment; Europe says now the language model should not be there.
19:51 Sarah EMA's reflection paper does use the words, in two places. In its section on product information, it says AI used for drafting, compiling, editing or reviewing product information documents should be used under close human supervision, and gives the reason: quoting, generative language models are prone to include plausible but erroneous or incomplete output, so quality review mechanisms need to ensure model-generated text is factually and syntactically correct. And in its integrity section it says large language models, often containing billions of parameters, are at particular risk of memorising training data. On manufacturing it says model development and lifecycle management should follow quality risk management principles and ICH Q8, Q9 and Q10, quoting, awaiting revision of current regulatory requirements and GMP standards. That waiting is Annex Twenty-two.
20:50 Sam And EMA on learning in use?
20:52 Sarah Context-dependent, which is worth knowing. For pivotal clinical trials: quoting, incremental learning approaches are not accepted, and any modification of the model during the trial requires a regulatory interaction. For pharmacovigilance it says the opposite may be allowed: incremental learning can continuously enhance models for classifying adverse event reports, with the holder responsible for validating and monitoring. So EMA's own paper treats adaptive models as acceptable in one setting and unacceptable in another, by consequence. And a small lineage point: its glossary defines a frozen model as one where all parameters have been finally set, not allowing further adaption to new data. Annex Twenty-two's glossary defines static with the same words.
21:45 Sam Now NIST, which is the only one of tonight's documents written about generative AI on purpose.
21:51 Sarah AI six hundred dash one is a profile of NIST's AI Risk Management Framework for generative AI, approved in July twenty twenty-four, standards layer, voluntary. Section two lists twelve risks unique to or exacerbated by generative AI. Four of them are the plant's. Confabulation, which it defines as, quoting, the production of confidently stated but erroneous or false content, colloquially hallucinations, and it says confabulation is especially important to monitor when generative AI is integrated into consequential decision-making. Human-AI configuration: automation bias and over-reliance, humans perceiving generated content as higher quality than it is. Information integrity: content that may not distinguish fact from inference or acknowledge its uncertainties. And value chain and component integration: third-party components, pre-trained models and datasets that are not transparently vetted, so that issues in a system's behaviour are hard to attribute to any one source.
23:04 Sam Put those four next to a deviation investigation and they're not abstract. A confident paragraph citing a batch record entry that doesn't exist. A reviewer who stops reading closely by the fortieth draft. A root cause stated as fact that was an inference. And a vendor who swaps the model under you between Tuesday and Wednesday. NIST names them; nobody in GMP has yet said what evidence answers them.
23:28 Sarah Which leaves the question the landscape puts at the centre: whether a given documentation use is critical. Under Annex Twenty-two that is the only question that matters for a generative model, because criticality is the property of the application, not the model, and there is no middle tier. Under FDA it is the same question in different words, and the answer decides whether you owe a credibility assessment.
23:55 Sam One more thing about the anchor before the examples. It's six pages. It can't be six pages by accident.
24:01 Sarah It isn't. Annex Twenty-two was published in the same consultation as a revised Annex Eleven on computerised systems, which grows from a few pages to seventeen chapters covering supplier oversight, audit trail, identity and access management and cybersecurity, and a revised Chapter Four on documentation. The AI annex says in its first paragraph that it provides additional guidance to Annex Eleven. Its no-decrease clause, four point three, which says the model's acceptance criteria should be at least as high as the performance of the process it replaces, cites Annex Eleven for how you know that performance. And its operation section, ten, is Annex Eleven vocabulary: change control on the model, the system and the whole process it automates before deployment; configuration control with measures to detect unauthorised change; performance monitoring against the defined metrics; and monitoring whether the inputs are still inside the model's sample space, with metrics for drift.
25:06 Sam So you read it as chapter twenty-two of a book whose other chapters are being rewritten at the same time. The data integrity, the access control, the audit trail: those aren't in the six pages because they're in the seventeen chapters next door. The six pages are only what's new about a model.
25:23 Sarah And what is new about a model, in the draft's view, is intended use with an input sample space, acceptance criteria fixed before testing, independent test data, explainability, confidence, and the operation clauses. The evidence episode reads those in full. Tonight what matters is that all of it presumes a frozen model: you cannot put a model under configuration control and detect unauthorised change if the model is changing itself every night. The static requirement isn't an extra rule bolted onto the annex. It's the assumption the rest of the annex stands on.
25:58 Sam Which brings us to the four systems, and the cold open. Soft sensor first, with nightly recalibration switched on.
26:04 Sarah Under Annex Twenty-two: it adapts its performance during use by incorporating new data, so it is dynamic, and if it feeds a critical step, in the tiers we set last episode a tier two or tier three use, it should not be used, whatever its accuracy. Under FDA's draft: anticipated. It is a self-evolving model; you monitor its metrics on an ongoing basis, you evaluate model-directed changes through the quality system's change management, and you report the ones that affect performance. The same object, on the same day, is told not to exist in one jurisdiction and told to keep a maintenance plan in the other.
26:45 Sam And the practical answer for a site inspected by both.
26:49 Sarah The documents point at one construction. Switch the nightly recalibration off in the deployed model. Freeze it, version it, put the version under configuration control per clause ten point two. Run the learning loop somewhere else: a development environment where the candidate model is trained on the new batches, tested against an independent test set held under the controls in section six, and compared against the acceptance criteria fixed under section four. When it passes, the new version is released through change control under ten point one as a deliberate, documented change. The deployed model is static between releases; the improvement still arrives, on a schedule you control, with a test record for every version. In the change-control vocabulary this season uses, you have turned a model-directed change into a deliberate one.
27:44 Sam Which is what most sites will do, and it costs you the thing the adaptive model was for. You wanted it to track the process as it drifted; now it tracks the process on your release cadence. That's a real loss and I'd say so in the risk assessment rather than pretend the frozen version is just as good. Batch monitoring model.
28:05 Sarah The multivariate model is the easy one, and it is worth saying why. It is fitted once on a set of reference batches, frozen and versioned; identical inputs give identical scores; it is not generative. Static, deterministic, in scope. Annex Twenty-two applies in full, which means the evidence clauses apply in full, but the gate is open. If a site re-fits it after a campaign, that is a new version through change control, exactly the construction we just described. It only becomes dynamic if someone builds it to update its own reference set as batches complete, and the draft would then say the same thing it says to the soft sensor.
28:47 Sam Fill-volume camera.
28:48 Sarah A trained image classifier, frozen: static and deterministic, in scope and permitted, and FDA's own example of a manufacturing model. Two things to watch. Vision vendors increasingly offer on-line learning, the model refining itself on the production stream; under Annex Twenty-two that feature stays off in a critical application. And a vendor's model update is a change to the model under ten point one, so it is evaluated for retest before it goes live, and a decision not to retest must be fully justified. That is the same clause whether the vendor calls it a patch or a new version.
29:26 Sam And the agent. Take it slowly.
29:28 Sarah An assistant on a large language model that reads batch records and drafts deviation investigations. Under Annex Twenty-two it meets all three bars at once. It is generative and a language model, so it is excluded by name. Left at default settings it samples, so it is probabilistic. And if it is set up to learn from the engineer's corrections, it is dynamic. In a critical application, should not be used, three times over. Pinning the weights and setting temperature to zero removes two of the three and leaves the name.
30:01 Sam So the only route the draft offers is non-critical.
30:04 Sarah Non-critical, with a qualified person always responsible for the output. That means the site has to be able to say, and defend, that drafting the investigation has no direct impact on product quality or data integrity, because the engineer re-derives the facts from the source records and the quality unit's decision does not rest on the draft. The crosswalk from last episode calls exactly that argument the open question. Under FDA, the same design decides whether you are inside the operational-efficiency carve-out or inside the seven steps; and if you are inside them, step four wants to know which foundation model, pre-trained on what, obtained how, and it wants the human-AI team evaluated, not the model. NIST tells you what the failure modes are called. No document tells you what test passes.
30:56 Sam Here's my one opinion on the gate itself, labelled, and then I'll leave it. The draft is right about practice and wrong about the tool. Freeze and version is the right practice: a frozen model is the only thing I can hand an inspector a test record for. But writing the exclusion as a name rather than a property means a pinned, deterministic language model with a full independent test set is out, while a poorly tested random forest is in. The eight companies behind the BioPhorum guidance said much the same in June. That's my one. The rest tonight is what the text says.
31:31 Sarah And what the text says is stable. Thirteen hundred comments and a workshop that looked squarely at adaptive and generative models produced no change to the draft. If the final text moves on this, the landscape's watchlist will record it. Until then the gate is what a European inspector reads.
31:49 Sam So what for biomanufacturing. Opinions labelled. First, for any site an EU or PIC/S inspectorate walks, model type is now a compliance decision and it belongs in the risk assessment before the architecture is chosen. Three questions, in the draft's own words: does it adapt in use, does it give identical outputs for identical inputs, and is it generative. Write the answers down. If the application is critical and any answer is the wrong one, the model doesn't go there, and no accuracy figure changes that.
32:20 Sarah The draft supports the ordering. Its personnel clause puts process experts, quality, data scientists and IT together during algorithm selection, before training, validation and testing. Algorithm selection is where the three questions get answered, so the draft has already told you who is in the room when the model type is chosen.
32:40 Sam Second: freeze, version, and put the learning loop outside the GMP boundary. Every model that matters gets a version number, a configuration record, and an independent test result per version. Retraining is a project, not a process; it happens in a development environment and re-enters production through change control. That construction satisfies the European draft and it is also what FDA's lifecycle plan wants documented, so it is the one design that works on both sides of the ocean. The cost is that you give up tracking the process in real time, and the honest risk assessment says so.
33:22 Sam Third, for the agent: decide now which documentation uses you'll defend as non-critical, and design them so the claim stays true. The engineer re-derives every fact from the source record. The agent proposes and never executes. The signature is human, and so is the decision. Keep an audit trail of what the agent read and what it wrote, because a draft you can't trace is a data integrity question, and data integrity is one of the three words in the criticality test. And pin the model version, because a vendor update is a model change under clause ten point one whether or not the vendor told you.
34:00 Sarah Where the teams feel it. MSAT and process science own the intended use and the criticality call for each application; the draft makes a process expert responsible for that description. Data science owns the three model-type answers and the frozen-version discipline. Quality owns the change control and configuration control that make the freeze real, and the monitoring of any human in the loop as a manual process. Regulatory CMC has to carry the divergence: the same model may need a lifecycle-plan summary in a US application and a static declaration for an EU inspection.
34:39 Sam And fourth, the docket habit from episode one applies with force here. The gate is a draft. The workshop happened. If your site has an adaptive model doing real good in a critical step, with evidence, the place to say so is the consultation on the final text and the industry groups feeding it, not the inspection.
34:58 Sam Three things. Sarah.
35:00 Sarah Draft Annex Twenty-two is the first binding-track GMP text on AI, six pages, and its scope is deliberately narrow: static, deterministic machine learning in critical applications. Dynamic models, probabilistic models and generative AI are not merely uncovered; the draft says they should not be used in critical GMP applications, and generative models in non-critical use always get a qualified human owning the output. Thirteen hundred comments and a workshop; the published text is unchanged, and EMA says a revision is under way.
35:38 Sam Second, mine: FDA and the EU diverge on exactly the model that's new. FDA doesn't endorse or exclude any technique, tells you to anticipate self-evolving models under a lifecycle plan, and lets you change the approach if credibility falls short; the device side pre-authorises continuous learning. Europe gates before the evidence. A global site resolves it the same way every time: freeze, version, and keep the learning loop outside the boundary, and admit what that costs.
36:09 Sarah Third: on generative AI, FDA is silent and the only handle is the operational-efficiency carve-out, which holds until the output touches quality. EMA says generative text is plausible but erroneous and wants close human supervision. NIST names confabulation, automation bias, information integrity and the value chain. Whether a documentation use is critical is the question every one of them hands back to the site.
36:37 Sam Next time: evidence. What do I have to show? What credibility looks like on paper, and how validation turned into assurance. FDA's step four and Annex Twenty-two's clauses three to ten, read clause by clause, and the fill-volume camera's evidence package built in front of you. Are we there yet? We know what we're allowed to build. Now we find out what we have to prove.