Are we there yet? Jack Prior is a bioprocess engineer who has spent his career in manufacturing science and biopharma manufacturing, using process data and process models to monitor, understand and control process variation — for yield improvement, tech transfer and troubleshooting. He began in the late 1980s at MIT looking at how to apply artificial intelligence to bioprocessing, perhaps a little too soon; forty years on, in 2026, the time for AI seems finally to have arrived. The promise has never been greater, and neither have the pitfalls and the challenges. The question this podcast asks is: are we there yet — in our data readiness, in our technology, and in our regulatory frameworks? Season 1 is an experiment. There are at least forty regulatory guidances and white papers to understand, so Jack uses AI to work through them in a narrative format, one question per episode, sized for a walk. The voices are AI characters, not Jack; the views are not his or his employer's, and none of it is regulatory advice. Each episode is fact-checked, claim by claim, against the source documents, with the results in the show notes.
Every episode reads documents from the AI-in-CMC Regulatory Map, an interactive index of the whole rulebook with its tiers and change-control instruments.
Season one read draft EU GMP Annex 22 as the one text that drew a hard line on model type: static and deterministic machine learning only in critical GMP applications, with adaptive models, probabilistic models and generative AI told they should not be used there. On 2 October 2026 EMA published the report of its 30 June expert workshop, and its introduction says the drafting group has discussed opening that scope, provided the models comply with the annex and sit under a fully documented, risk-based control strategy. Sam and Sarah read the fifteen pages closely: industry's case against technology labels, the ICH Q9 vocabulary of uncertainty, importance, complexity and detectability, where the limits of mitigation lie, the report's distinction between human oversight and human-in-the-loop, layered controls that prevent, detect and contain, the lifecycle and supplier questions, and what the drafting group says it may now write. Then the four recurring systems go through the proposed test instead of the old gate, the agent longest of all, and the episode closes on what a site should build now because it is the same evidence package either way.
157 factual claims in this episode were checked against the source documents by an independent AI pass: 139 are supported by the primary text, 0 rest on secondary sources or are hedged as such (listed in the report), 18 open. Full report.
Open items — treat these lines with caution:
HOST: “Under draft Annex 22, dynamic/probabilistic/generative models should not be used in critical GMP applications, and in non-critical applications a human in the loop is expected (the model drafts and a person decides), so the model-type restriction does not bar such non-critical use.” — PARTIAL (The critical-use bar covers all three types (dynamic lines 13-15, probabilistic 17-19, generative/LLM 20-21), but the non-critical HITL sentence is attached only to Generative AI/LLM ('such models'). For a generative deviation assistant, which is what the cold open describes, the claim holds. 'The model drafts and a person decides' is the script's gloss; the text says personnel are responsible for ensuring outputs are suitable.)
HOST: “The third page of the report says the drafting group has discussed allowing generative AI in GMP applications, critical or not, provided it is supported by a documented control strategy.” — PARTIAL (Page 3 is correct. The phrase 'whether critical or non-critical' belongs to the preceding sentence about public-consultation comments; the drafting-group sentence says only 'in GMP applications' (no criticality qualifier, though it drops the draft's 'critical' restriction). It also adds a second condition: compliance with the annex's requirements.)
SARAH: “Draft Annex 22 is the only text in the corpus that sorts AI by model type.” — NOT-CHECKABLE (Corpus-wide characterisation carried over from episode 4; not checkable against a single document. Grep of the corpus: only the Annex 22 draft, the workshop report (quoting the draft) and the BioPhorum guidance (discussing the draft) pair 'static' with 'deterministic' as model sorting; no other corpus document gates AI by model type. Hedged by 'in the corpus'.)
SARAH: “The Annex 22 draft dates from July 2025.” — UNSUPPORTED
HOST: “This report is the first official text in the corpus that moves the answer on adaptive models or generative AI in critical GMP.” — NOT-CHECKABLE (Scoped to the corpus on disk; not checkable against a single document. Grep of the corpus: the only other official texts mentioning the Annex 22 revision are the Feb 2026 HMA-EMA minutes (industry request forwarded; 'undergoing revision'), the IWG work plan (Q4 2026 target) and the June 2026 event page ('EMA is still considering the implications'); none records the drafting group discussing opening the scope.)
SARAH: “For each topic the report gives the drafting group's questions, the industry perspective, a Q&A, key messages, and a paragraph headed potential impact on Annex 22 drafting.” — PARTIAL (Five of six topics carry all five headings. Topic 2 lacks the 'Industry Perspective' and 'Q&A session' headings (the content is present in running prose: 'This led to important Q&A'). Key Messages heading varies: 'from the Session' (Topic 1) vs 'Emerging from the Discussion' (others).)
SARAH: “Industry at the workshop leaned on the Q9 sentence that formality should be commensurate with uncertainty, importance and complexity.” — PARTIAL (The report invokes the Q9(R1) formality concept with those three factors but never quotes or paraphrases a 'commensurate with' sentence; 'commensurate' appears only in the EU AI Act reference on p.7. The attribution of that specific Q9 sentence to industry is the script's gloss.)
SARAH: “The reply at the workshop went back to ICH Q9's text: risk analysis is about likelihood and severity, and in some risk management tools the ability to detect the harm (detectability) also factors in the estimation of risk.” — PARTIAL (Substance matches, but the report does not quote ICH Q9's text; it paraphrases ('severity and probability') and names FMEA specifically rather than 'some risk management tools'. The Q9 sentence the script reads is Q9's own (verified against ICH Q9(R1) 4.3), not a quotation recorded in this report.)
HOST: “The report comes from EMA's GMP inspectors (an EMA inspectors' report).” — PARTIAL (The page never names the report's author; it describes a workshop report 'with expert input' reviewed by 'the Annex 22 drafting group at EMA'. 'Inspectors' report' is an inference from the organiser.)
SARAH: “'Outsourced Activities' is the title of Chapter 7 of Part I of the EU GMP guide.” — NOT-IN-CORPUS (EU GMP Part I Chapter 7 is not among the primary documents; hedged in-script ('isn't in our corpus, so take that as my reading'). The report's own words are verified (p.12 'particularly Chapter 7 on outsourced activities') and the Annex 11 draft's chapter 7 is verified as 'Supplier and Service Management'.)
SARAH: “The draft's lines 12-25 still read as they did in July 2025 (no amended draft exists).” — NOT-CHECKABLE (Absence claim about the world (no amended draft exists); the report describes the exclusion as what the draft 'currently states' (p.3) and the revision as ongoing (p.14); no revised draft is on disk; not checkable against a single document.)
SARAH: “No revised Annex 22 text has been published.” — NOT-CHECKABLE (Absence claim about the world; the report describes the revision as ongoing and commits only to review of the input (p.14); no revised draft is on disk; not checkable against a single document.)
HOST: “The report's assessment dimensions are consequence, influence, uncertainty, complexity and detectability, while the July draft sorted by technology.” — PARTIAL (The dimensions and the draft's technology-based exclusion are supported; the report never dates the draft as 'July' (it says only 'the draft version ... currently states').)
HOST: “ICH M15 uses the vocabulary of consequence and influence (model influence, decision consequence) in its risk framework.” — PARTIAL (Substance right: M15's risk framework combines 'model influence' and 'consequence' into 'model risk'. M15's exact term is 'consequence of wrong decision' (2.1.4), not 'decision consequence', which appears nowhere in the document.)
HOST: “Under the draft, a soft sensor that retrains itself nightly (a model that adapts performance) should not be used in critical GMP applications.” — PARTIAL (The soft-sensor / nightly-retraining example is the script's own; the draft's dynamic-model wording is 'continuously and automatically', so whether a scheduled nightly retrain counts is an inference from the frozen-model definition, not a statement in the text.)
SARAH: “The report refers to change triggers and re-optimisation or retraining criteria as lifecycle controls for adaptive models.” — PARTIAL ('Change triggers' and 're-optimisation' are the report's terms; 'retraining criteria' is not a phrase in the report (retraining appears only in the drafting-group question 'updates, retraining, drift' and 'continuously retrains'). The report's criteria language is 'rollback or intervention criteria'.)
SARAH: “The report says supplier agreements should give the regulated user visibility of changes when a third-party or foundation model is updated.” — PARTIAL (Substance supported; the term 'foundation model' does not appear anywhere in the report (it says 'third-party model', 'cloud/provider', 'supplier').)
SARAH: “No revised draft of Annex 22 exists.” — NOT-CHECKABLE (Absence claim about the world; same basis as lines 63-64: the report describes the revision as ongoing (p.14) and no revised draft is on disk.)
A site head wants one slide by Friday: four AI systems, three columns, do now, wait for, do anyway. Sam and Sarah close the season by reading the two documents that say where the rulebook is heading: the joint EMA–FDA Guiding Principles of Good AI Practice (January 2026), ten principles on two pages that put context of use, model risk, fit-for-use data, lifecycle and the human–AI team under both logos for the first time; and ICH's October 2025 reflection paper on advanced manufacturing, which names process modelling and continuous process verification as topics, lists the gaps in the 2011 Points to Consider in its own words, and proposes a process-models guideline as the first step, with no topic adopted yet. Then the calendar as the landscape verified it (CDER's promised manufacturing guidance, the Annex 22/11/Chapter 4 finals, the FDA draft, the AI Act's deferred dates), the EU machinery behind it (EMANS 2028, the NDSG and QIG workplans), the sandboxes that exist and the GMP sandbox that doesn't, the six open questions as the season found them, a scorecard for the four recurring systems, and Sam's verdict: nearly there on vocabulary, not on instruments, and the plant floor decides.
88 factual claims in this episode were checked against the source documents by an independent AI pass: 68 are supported by the primary text, 9 are confirmed by official web sources, 3 rest on secondary sources or are hedged as such (listed in the report), 8 corrected before release, 0 open. Full report.
A vision vendor pushes a new model version at two in the morning and a language-model provider deprecates the version a site validated. Sam and Sarah split 'change' into four kinds (the environment drifts, the owner retrains, the model learns, the vendor pushes) and read the instruments inside out, anchored on ICH Q12: Q10 and the Annex 11 and Annex 22 drafts inside the quality system; Q12's reporting categories, Established Conditions and the post-approval change management protocol in the filing; FDA's 2025 draft tying them together with a lifecycle maintenance plan; and the device-side predetermined change control plan as the template nobody has married to the PACMP. Then the two questions FDA asked in 2023 and ICH repeated in 2025, a tier-by-tier rule for which instrument applies, the vendor update no document covers, and the four recurring systems taken through all four kinds of change, ending with what an AI PACMP for the soft sensor would actually say.
68 factual claims in this episode were checked against the source documents by an independent AI pass: 63 are supported by the primary text, 2 are confirmed by official web sources, 1 rest on secondary sources or are hedged as such (listed in the report), 2 corrected before release, 0 open. Full report.
Every document that lets AI into GMP lets it in on one condition: a human in the loop. Sam and Sarah read what that phrase actually means across six documents, anchored on EMA's 2024 reflection paper: BioPhorum's three oversight settings and the grade each one earns, the EU AI Act's Article 14 list of what an overseer must be able to do, draft Annex 22's operator monitored like any other manual process, FDA's and the GMLP principles' human-AI team, and the paradox the documents state themselves, that oversight is both a control and a risk. Then Sam's question, parked since episode four: which human, with which competence, evaluating which inputs, measured how. The four recurring systems get the three questions, the agentic assistant gets the night-shift test, and Sam's position is offered for the listener to test: encode every check a computer can run, name the human judgment by competence, and if that competence is not on shift, the control is fiction and the grade must not get credit for it.
47 factual claims in this episode were checked against the source documents by an independent AI pass: 43 are supported by the primary text, 2 are confirmed by official web sources, 0 rest on secondary sources or are hedged as such (listed in the report), 2 corrected before release, 0 open. Full report.
Every framework this season assumes the data under the model are trustworthy. Sam and Sarah read what the rulebook actually requires: ALCOA and its pluses from FDA's 2018 Data Integrity Q&A, PIC/S PI 041 and the Annex 11 revision; FDA's 2025 'fit for use', which adds relevance to reliability; and the data-governance clauses of the AI Act, EMA's reflection paper and draft Annex 22. The Annex 11 revision turns the training-data pipeline into a GMP computerised system, with validated transfers, unique accounts, audit trails and backup. Then the gap: nobody says how to get from a historian, a batch record and a LIMS to a fit-for-use dataset. The soft sensor's training set is interrogated attribute by attribute, the batch model's golden batches become exclusion decisions, the camera's labels and vials get their clauses, and the agentic assistant's data turn out to be documents, versions and all.
69 factual claims in this episode were checked against the source documents by an independent AI pass: 65 are supported by the primary text, 4 are confirmed by official web sources, 0 rest on secondary sources or are hedged as such (listed in the report), 0 open. Full report.
Once a model is allowed through the gate, what do you owe on paper? Sam and Sarah read FDA's 2025 draft for its credibility assessment plan and report, sized to model risk, and then read draft EU GMP Annex 22 sections three to nine as the most concrete evidence recipe in the corpus: an intended use owned by a process expert, acceptance criteria fixed before testing and no lower than the process being replaced, a designed test set with verified labels, test-data independence enforced by access control and staff separation, and explainability and confidence recorded as evidence. FDA's Computer Software Assurance guidance, the ten Good Machine Learning Practice principles and the 2023 credibility vocabulary explain how validation became assurance. The fill-volume camera's evidence package is built clause by clause, then the soft sensor, the batch model and the agentic assistant go through the same recipe.
58 factual claims in this episode were checked against the source documents by an independent AI pass: 52 are supported by the primary text, 4 are confirmed by official web sources, 1 rest on secondary sources or are hedged as such (listed in the report), 1 corrected before release, 0 open. Full report.
For twenty years the model you picked was an engineering choice. Draft EU GMP Annex 22 makes it a compliance decision: static and deterministic machine learning only in critical GMP applications, with dynamic models, probabilistic models and generative AI told they should not be used there at all. Sam and Sarah read the six-page draft in full, set it against FDA's 2025 draft, which tells sponsors to anticipate self-evolving models and never mentions language models, and against the device world's predetermined change control plans. EMA's reflection paper and NIST's generative AI profile supply the vocabulary for the risks. The four recurring systems go through the gate, and the episode ends with how a site keeps a retraining loop outside the GMP boundary.
64 factual claims in this episode were checked against the source documents by an independent AI pass: 58 are supported by the primary text, 4 are confirmed by official web sources, 2 rest on secondary sources or are hedged as such (listed in the report), 0 open. Full report.
Seven documents grade models seven ways, and they grade different things. Sam and Sarah start with the 2011 ICH Points to Consider, which graded models low, medium and high by their role in assuring quality and made 'sole indicator' the line, then set beside it FDA's and M15's model risk, Annex 22's critical-or-not test and model-type gate, the AI Act's product tiers, and BioPhorum's redefinition of model influence as maturity. Four tiers emerge for the season. All four recurring systems are placed and moved, and the agentic assistant's tier turns out to be an argument about criticality that has to be won first.
69 factual claims in this episode were checked against the source documents by an independent AI pass: 62 are supported by the primary text, 6 are confirmed by official web sources, 1 rest on secondary sources or are hedged as such (listed in the report), 0 open. Full report.
The one idea in nearly every document: context of use. Sam and Sarah read FDA's January 2025 draft closely, from question of interest to role and scope to model risk as influence times consequence, trace the vocabulary from ASME's device standard through FDA's 2023 device guidance to ICH M15, and set EMA's two axes beside it. FDA's fill-volume camera is medium risk only because a sample test stays in place. All four recurring systems go through steps one to three, and the agentic assistant turns out to be the one whose scope sentence is hardest to write honestly.
74 factual claims in this episode were checked against the source documents by an independent AI pass: 70 are supported by the primary text, 4 are confirmed by official web sources, 0 rest on secondary sources or are hedged as such (listed in the report), 0 open. Full report.
Before reading any single document, you need the map. Sam and Sarah sort the AI-in-manufacturing rulebook into its four layers, replace "binding or not" with maturity and who-will-ask, and trace five trends across the documents: from models to AI, from model-agnostic to Annex 22's line on model type, from validation to lifecycle, from regional to the joint EMA-FDA principles and ICH's reflection paper, and from silence on generative AI to a gap everyone can see. The recurring cast is introduced, and the agentic assistant that drafts deviation investigations is walked down the maturity axis to show how little the documents yet say about it, and how a site reasons about it anyway.
148 factual claims in this episode were checked against the source documents by an independent AI pass: 121 are supported by the primary text, 20 are confirmed by official web sources, 6 rest on secondary sources or are hedged as such (listed in the report), 1 corrected before release, 0 open. Full report.
Jack Prior introduces the show: a bioprocess engineer who started out in the late eighties at MIT looking at how to apply AI to bioprocesses, on why 2026 is the year that time has arrived — and the question the podcast asks of our data readiness, of machine learning and agentic AI, and of the regulatory frameworks: are we there yet? Season 1 works through the forty-odd guidances and white papers that make up the regulatory landscape, in a narrative format produced with AI and fact-checked against the source documents. Everything at https://jackprior.ai/podcast
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: chat@jackprior.ai.
