← Back to All Blogs
Pharma AI

Responsible AI in Pharma: A Governance Framework

By Ankit Singh — Social Media Marketing Specialist, Multiplier AI  ·  Published November 18, 2026  ·  ✎ Updated November 18, 2026

Pharmaceutical AI governance occupies an unusual regulatory position, and naming it precisely is the first useful thing a framework can do. Three major regulators have published their thinking. None of them has published a rule. The FDA document is marked not for implementation and described as containing non-binding recommendations. The EU GMP annex on artificial intelligence is a consultation draft that has not been adopted. The EMA's contribution is a reflection paper, which is the agency's own word for a document that describes current thinking rather than creating a requirement.

That combination produces a specific failure mode, and most pharmaceutical companies are in it. A compliance function asked to govern AI finds three authoritative documents, none of which it is obliged to follow, and responds either by treating all three as mandatory, which over-governs low-risk systems into paralysis, or by treating none of them as relevant, which leaves high-consequence systems with no control at all. Both responses are reasonable reactions to genuine ambiguity and both are wrong.

There is a second gap that matters more commercially and is discussed less. All three frameworks address regulated activity: regulatory submissions, manufacturing, the medicinal product lifecycle. Commercial AI is outside every one of them. The models that score physicians, generate promotional copy, recommend next best actions and forecast territory performance are not covered by the FDA guidance, not covered by the draft manufacturing annex, and not covered by the EMA paper. They are also, in most companies, the majority of AI in production and nearly all of the generative AI.

This page is an operating model rather than a map. Which law applies to which dataset in which market is covered on our HIPAA, GDPR and DPDP comparison page, including the EU AI Act classification question and the amended high-risk dates. What follows here is how to decide, document and review a deployment when the rule you would like to comply with does not yet exist.

How reliable is this page, stated up front

No primary instrument was read in writing this page. The FDA page we reached gave the guidance's title, status and abstract but not its body, and the guidance text itself was not fetched. The draft GMP annex, the EMA reflection paper, ISO 42001 and the NIST framework were all read through secondary summaries.

Two consequences follow. The seven FDA step names come from a consultancy's reproduction rather than from the FDA's own wording. The draft annex's section numbers come from one vendor's paraphrased numbering and are not reproduced here at all.

One source gave requirements that two others did not, including training data rules, parallel running and model rollback. Those are flagged where used.

Verify against the primary documents before building a programme on this page. The structure is well attested. The wording is not.

Four frameworks, and what each one actually is

Knowing the status of each document prevents both over-governance and neglect, and the status differs more than the content does. The FDA's credibility assessment guidance is a draft published in January 2025 under docket FDA-2024-D-4689, with comments closed in April 2025 and no public commitment to a finalisation date as of the sources we consulted in late 2026. The EU GMP annex on artificial intelligence went to public consultation alongside a revised Annex 11 and a revised Chapter 4 on 7 July 2025, closed on 7 October 2025, and had not been adopted as of September 2026, with an agency workshop in late June 2026 reportedly considering whether to widen its scope. The EMA's reflection paper was finalised on 9 September 2024 after more than 1,300 consultation responses and describes expectations rather than obligations. The two voluntary standards are different in kind: one is a certifiable management system standard with an external audit regime, and the other is a free, self-attested risk framework. None of them is a rule you can be fined under today, and the EU AI Act, which is binding, reportedly does not classify most commercial pharmaceutical AI as high-risk at all.

The status table

Four documents, four different kinds of authority, and the differences change how each should be used.

 

DocumentStatus as reportedWhat it coversHow to use it
FDA draft guidance on AI supporting regulatory decision-makingDraft, January 2025. Marked not for implementation and containing non-binding recommendations. Comments closed April 2025 with no finalisation commitment reportedAI producing information to support regulatory decisions on safety, effectiveness or quality. Reportedly excludes drug discovery and operational efficiencyAdopt the risk model and the credibility plan structure now. They are the most transferable things in any of these documents
EU GMP draft annex on artificial intelligenceConsultation draft. Opened 7 July 2025, closed 7 October 2025, not adopted as of September 2026AI models in critical GMP applications with direct impact on patient safety, product quality or data integrity. Static and deterministic models onlyTreat as the direction of travel for manufacturing. The acceptance criteria rule is worth adopting everywhere
Revised EU GMP Annex 11 on computerised systemsSame consultation. Not adopted; the 2011 version reportedly remains in forceComputerised systems generally, with supplier oversight, audit trails, electronic signatures and securityThe AI annex is additional to this, not a replacement. Your existing validation framework still governs
EMA reflection paper on AI in the medicinal product lifecycleFinal, 9 September 2024. A reflection paper, not binding guidanceDiscovery through post-authorisation, including pharmacovigilance. Risk framed as high patient risk and high regulatory impactUse its two-category risk vocabulary. It was reportedly chosen specifically to avoid confusion with the AI Act
ISO/IEC 42001Published 2023. Voluntary and certifiable, with external audit and annual surveillance reportedAn AI management system, structured like other management system standardsChoose it if you need external assurance. It gives a certificate, not a conformity position
NIST AI Risk Management Framework 1.0Published January 2023. Voluntary, self-attested, with no audit layer unless boughtFour functions: govern, map, measure and manageChoose it if you need engineering traction. Free, flexible, and no certificate at the end

The second row carries a point most summaries miss. The draft annex is additional to the computerised systems annex rather than a replacement for it, and one analysis was explicit that it does not displace computerised system validation or the established industry guidance. An organisation with a mature validation framework already has most of the structure, and what the annex adds is specific to the fact that a trained model is not explicitly programmed logic.

One disagreement to flag now rather than later. Our sources differ on whether the draft annex finalises soon: one written in April 2026 expected all three documents to be final by mid-2026, and two later sources confirm they were not. Treat any published timeline for these documents as an expectation rather than a schedule, which is the same lesson our comparison page draws from every deadline in this area moving during 2026.

AI governance framework template for pharma

A workable framework has five parts and fits on four pages, and most of what makes it work is borrowed from the regulator drafts rather than invented. Part one is a register with one row per AI system, carrying a named owner, because a committee that owns everything owns nothing. Part two is a classification gate scoring each system on two axes: how much the model influences the decision, and how badly a wrong decision ends, which is the FDA's model risk formulation and is the single most transferable idea in any of these documents. Part three is a tiered evidence requirement, so that a low influence, low consequence system needs a paragraph and a high influence, high consequence system needs a credibility plan. Part four is a pre-deployment gate whose hardest requirement is the one the draft GMP annex supplies: acceptance criteria agreed in advance, by a subject matter expert, set at least as high as the performance of the process being replaced. Part five is in-life monitoring with a defined response when performance drifts below the floor. Nothing in that list requires a principles document, an ethics board or a maturity model, and a company that has those five parts running is better governed than one with a published framework and no register.

Part one: the register

Every governance programme that works starts with a list of what exists, and the absence of that list is the most common finding in this area. The columns below are the minimum. A register with fewer columns cannot answer a regulator's first question, and a register with more columns stops being maintained.

 

ColumnWhat it holdsWhy it is non-negotiable
System name and business purposeIn plain language, as the business would describe itTechnical names make the register unreadable to the people who must approve entries
Named accountable ownerOne person, not a function and not a committeeThe single highest-value column. The EMA paper reportedly places accountability on named parties rather than on the organisation in general
Regulated or commercialWhich framework, if any, reaches itDecides whether the regulator drafts apply or whether you are governing voluntarily. Most entries will be commercial
Model influenceHow much weight the output carries in the decision, relative to other evidenceHalf of the FDA's risk formulation, and the half organisations consistently overstate
Decision consequenceSeverity and probability of an adverse outcome if the decision is wrongThe other half. A high consequence decision with an independent check is not automatically high risk
Risk classThe product of the two, in whatever bands you adoptDrives the evidence tier. Bands matter less than consistency
Performance floorThe agreed acceptance criterion, and the measured performance of the process it replacesThe control that prevents retrospective justification
Provider or deployerWhether you built it, bought it, or materially changed a bought oneFine-tuning a vendor model can change your status, which our comparison page covers
Data inputs and their lawful basisWhat the model consumes and under what permissionLinks the governance register to the data register. Most organisations maintain these separately and they disagree
Last review date and next review dateDates, not intervalsAn interval is a policy. A date is a commitment and it shows up as overdue

The ninth row is where two programmes usually collide. Data governance maintains a record of what personal data exists and on what basis, AI governance maintains a record of what models exist, and in most companies neither points at the other. A model consuming physician data under a basis that has since been withdrawn is invisible to both registers, which is exactly the failure a regulator would find first. The data side of that question is covered in our DPDP Act compliance checklist for India and in our cross-market comparison elsewhere.

Part two: the classification gate

Classification is the step that makes everything downstream proportionate, and the FDA's two-axis model is better than the single risk score most frameworks use. Model influence is the weight of the model's output relative to other evidence in the decision. Decision consequence is the significance of an adverse outcome if the decision is wrong, taking account of both severity and probability. The product of the two is model risk, and the insight is that risk is a property of the decision rather than of the model.

A worked pair of examples from our sources makes the point better than a definition. A model determining which trial participants can skip inpatient monitoring after dosing carries high influence and high consequence, and is high risk. A model inspecting fill volumes where an independent release test will catch a miss carries high consequence and low influence, and is reportedly medium risk, because the independent check absorbs the error. The same accuracy, a different decision architecture, a different evidence requirement.

Decision consequenceLow model influenceMedium model influenceHigh model influence
High: patient harm, product quality, regulatory decisionMedium risk. An independent control absorbs the errorHigh riskHighest risk. Full credibility evidence, independent review, human sign-off on each output
Medium: material commercial or operational harm, reputational exposureLow to medium riskMedium risk. The commonest band for commercial AIHigh risk. Most generative content systems sit here and are governed as if they were low
Low: recoverable, visible, cheap to correctLow risk. A register entry and an owner is proportionateLow riskMedium risk. Automation without a check raises the band even when the stakes are small

The middle cell of the second row is where most pharmaceutical AI actually lives and where governance is thinnest. A system generating promotional copy for review, scoring physicians for targeting, or recommending a next best action carries medium influence over a decision with real commercial and reputational consequence. It is almost never classified, almost never has a performance floor, and is frequently the only AI in the company that a regulator outside the medicines system could complain about, because the output is published rather than internal.

One adaptation worth making for European readers. The EMA reportedly replaced the term high risk with two separate categories, high patient risk and high regulatory impact, specifically to stop people conflating a medicines judgement with the EU AI Act's product safety classification. Using those two words in your register alongside the influence and consequence scores is free and prevents a recurring argument about whether a system is high risk, in which the two sides mean different things.

Part three: evidence tiers

Proportionality is the whole point of classification, and a framework that demands the same evidence from a forecasting tool and a dosing model will be ignored by the people who have to use it. The tiers below are our construction, assembled from what the regulator documents ask for at the top end and from common sense at the bottom.

 

TierEvidence requiredApproval
Low riskA register entry with owner, purpose, inputs and lawful basis. One paragraph on why the risk class is lowThe owner. Nothing else, and resisting the urge to add a committee is part of the design
Medium riskWritten intended use. Documented performance floor and the measured baseline it was set against. A held-out evaluation and its result. A named human check in the workflow. A monitoring planOwner plus one independent reviewer who did not build it
High riskAll of the above, plus a credibility plan agreed before execution, documented data provenance and independence, explainability evidence appropriate to the model, limitations and known biases stated, and a defined rollbackGovernance forum, with the business owner present. Not delegated to a technical review
Highest riskAll of the above, plus independent review of the evidence, human sign-off on individual outputs rather than on the system, and regulatory engagement where the activity is regulatedExecutive accountability, recorded. If nobody is willing to sign, that is the answer

The top-tier regulatory engagement line is worth taking seriously where it applies. Both the FDA and the EMA frameworks reportedly encourage early engagement for high-impact systems, with the FDA's minimum package described as the first three steps of its framework plus the proposed credibility activities, and the EMA pointing to qualification of innovative development methods and scientific advice. Neither route exists for commercial AI, which is one more reason the commercial population needs its own discipline rather than a borrowed one.

Part four: the pre-deployment gate

One requirement in the draft GMP annex does more work than any other control described on this page, and it can be applied to any system in any part of the business. Acceptance criteria are defined in advance, by the subject matter expert responsible for the intended use, before test execution, and reportedly must be at least as high as the performance of the process being replaced.

 

Gate checkWhat it requiresWhat it prevents
Intended use is written and approved before testingA description of the task, the inputs, their expected variation, rare cases and known limits, approved before acceptance testingScope creep after deployment, where a model approved for one task quietly starts doing another
The baseline is measuredHow well does the current process actually perform? Measured, not estimatedThe most common gap. Teams cannot say how well the humans do, so any model looks like an improvement
The floor is set at or above the baselineThe acceptance criterion written down before results existRetrospective justification. After the fact, every model passes
The evaluation data is independentNever used in development. Where split from a shared pool, access controlled and audit trailed, with uses recordedContamination, which is the quiet reason many internal model evaluations are meaningless
Labels are defensibleProduced by independent experts or a validated methodA model that reproduces one person's opinion and is reported as accuracy
Confidence and abstention are designed inA confidence score with a threshold, and the option to return undecided rather than a forced answerSilent low-confidence outputs, which is how an averagely good model produces a catastrophically wrong single decision
The human check is specifiedWho checks what, on which outputs, with what authority to overrideOversight that exists on a slide, where the reviewer approves a hundred outputs a day and reads none
Rollback is definedWhat happens when the floor is breached, and who decidesA degraded model staying in production because removing it has no owner

The second and third rows together are the heart of it. A performance floor set against a measured baseline converts a subjective question, is this model good enough, into an objective one, is it better than what we do now. Most organisations skip the baseline because measuring the current process is tedious and politically awkward, and that omission is precisely why their model approvals cannot be defended afterwards.

The sixth row deserves emphasis because it is cheap and rarely done. A confidence threshold that lets the system decline to answer, routing the case to a human instead, converts a model's weakest outputs from errors into escalations. One source describes exactly this in the draft annex, flagging low-confidence results as undecided, and there is no reason a commercial system should not do the same.

Part five: in-life controls

Deployment is the start of the governance obligation rather than the end of it, and the drafts are consistent on what in-life means.

ControlWhat it involvesTrigger
Performance monitoringThe same metric as the acceptance criterion, tracked continuously, not a different metric chosen because it is easier to collectBreach of the floor. Defined in advance
Input drift monitoringWatching whether the data arriving still resembles the data the model was evaluated onA shift in the input distribution, independent of output quality
Change controlModel, system and process under change control, with any decision not to retest justified in writingAny change to the model, its inputs or the process around it
Deviation handlingProblems recorded and investigated through the existing quality process where one appliesAny floor breach or unexpected behaviour
Periodic reviewA dated review of the register entry, the floor, the classification and the ownerThe date in the register
Removal from useTaking the model out of the decision until it is investigated, retrained and revalidatedDescribed by one source only for the draft annex, and sensible regardless of its provenance

The first row contains a trap worth naming. Teams frequently monitor a metric that is convenient rather than the one the model was accepted against, which means the monitoring cannot detect a breach of the acceptance criterion. If the floor was precision at a fixed recall, that is what the dashboard must show, even when a simpler number is available.

What controls do regulators expect on commercial AI?

Almost nothing specific, and that is the finding rather than the relief. No framework described on this page reaches commercial AI. The FDA guidance covers AI producing information to support regulatory decisions on safety, effectiveness or quality, and reportedly excludes operational efficiency and drug discovery. The draft GMP annex reaches AI in critical manufacturing applications, and one analysis states it does not apply to administrative functions, business intelligence, or research not directly linked to manufacturing. The EMA reflection paper addresses the medicinal product lifecycle and, in the sources we read, does not discuss promotional content, commercial analytics or marketing claims at all. The EU AI Act is the one binding instrument, and the sources we consulted for our comparison page listed commercial analytics among the systems in neither high-risk annex. So a model that scores physicians, drafts promotional copy or recommends a next best action is, as far as AI-specific regulation goes, governed by nothing. What governs it instead are the instruments that never mention AI: data protection law over its inputs, the pharmaceutical marketing codes and professional conduct rules over its outputs, and ordinary liability over its consequences. Those apply in full, they apply today, and none of them accepts a model as an excuse.

What actually binds a commercial AI system

Four instruments reach commercial AI, none of them written with AI in mind, and all of them in force now. The useful mental move is to stop asking which AI rule applies and start asking what the system does, because every one of these regimes regulates an activity rather than a technology.

 

What the system doesWhat binds itWhy AI makes no difference
Consumes physician or patient dataData protection law in each marketA model is a processing purpose like any other. Lawful basis, notice, retention and rights all apply unchanged
Produces promotional contentThe pharmaceutical marketing code, and professional conduct rules where a practitioner is involvedThe code governs the claim, not the author. A generated claim that fails substantiation fails identically
Sends a messageElectronic communication rules, which differ by marketRegistered templates and consent registries do not care how the copy was written. Generated variants may not fit a registered template at all
Decides who gets contacted or prioritisedData protection rules on profiling, plus fairness and anti-discrimination exposureAutomated decision provisions attach to the decision, not the method
Informs a clinical or safety judgementThe regulated frameworks, and it has stopped being a commercial systemThe commonest scope error. A commercial tool that starts touching safety has changed category
Generates text a customer seesAdvertising standards, consumer protection and ordinary liabilityNo regime accepts the model as the author. The company published it

 

The third row is the one that catches generative content systems operationally rather than legally. Our DPDP Act compliance checklist describes the Indian requirement that promotional messages run on registered content templates, and a system generating a different message for every recipient is in direct tension with a regime built on pre-registered copy. That is not an AI ethics problem. It is a campaign that gets stopped, and it is the kind of control that surfaces only when somebody maps the generated output against the regime that governs sending.

The sixth row is the one to put in front of a marketing leadership team. No regulator, advertising authority or court will accept that a model wrote the claim. The pharmaceutical marketing code, covered in full on our UCPMP 2024 compliance guide, requires substantiation of claims and reportedly prohibits using healthcare professionals' names in promotional material, and the professional conduct rules on our NMC ethical advertising guidelines page bar testimonials outright for Indian audiences. A generative system that produces a plausible testimonial-shaped sentence has produced a compliance incident, and the fact that no human wrote it makes it worse rather than better, because it means nobody read it.

The controls we would impose on commercial AI anyway

Absence of a rule is not absence of risk, and the governance a commercial system needs is largely the same as the regulated one with the regulatory engagement removed. The list below is what we would require, with the reason stated in commercial rather than ethical terms, because that is what gets a budget.

ControlWhat it looks likeThe commercial reason
A register entry with an ownerOne row, one nameWhen something goes wrong, the first question is who approved this. An unanswerable question becomes an escalation
A performance floor against a measured baselineIs the model better than the targeting, the copy or the forecast it replaced?Most commercial AI is never compared against what it replaced, which means nobody can say whether it works
Human review before anything is publishedA named reviewer with authority to reject, and a realistic review volumeThe output is published under your name. The code applies to it whoever wrote it
Provenance of the data the model consumesField level source and lawful basis, linked to the data registerThe input is where the legal exposure is, and it is governed by instruments that are already in force
Template and channel compatibility checkDoes the generated output fit the registered template and the consent state for that channel?An operational control that prevents a campaign being stopped, which is the likeliest real consequence
Logging of what was generated and sentRetained records of outputs, recipients and approvalsYou cannot investigate a complaint about a message you did not keep, and generated content is not reproducible
An abstention pathThe system declines and routes to a human when confidence is lowConverts the worst outputs from published errors into internal escalations
A defined kill switchWho can stop it, how fast, without a committeeThe control you will need at the worst possible moment, and the one nobody rehearses

The sixth row is specific to generative systems and is routinely missed. A deterministic model's output can be reproduced by rerunning it; a generative system's cannot, which means that if the output was not logged at the moment of sending, it is gone. An organisation investigating a complaint about a message it generated six months ago, with no log, has no way to establish what was actually sent, and that is a worse position than having sent something indefensible and being able to show it.

The uncomfortable question for a steering committee

How many AI systems are in production in the commercial organisation, and who approved each one?

In most pharmaceutical companies the honest answers are more than anybody thinks and nobody in particular. Commercial AI arrives through marketing platforms, agency tooling, embedded vendor features and individual experimentation, none of which passes a governance gate because no gate exists for it.

The regulated side has a validation culture, a quality system and decades of practice. The commercial side has a budget and a deadline. That asymmetry, rather than any gap in the frameworks, is the actual governance problem in this industry.

Start by counting. The count is the finding.

The mismatch: frameworks assume a model you can test

Every framework described on this page was designed around a particular kind of model, and it is not the kind pharmaceutical companies are currently buying. The draft GMP annex covers models that are static, meaning they do not adapt during use, and deterministic, meaning identical inputs produce identical outputs. Dynamic, adaptive and probabilistic models are reportedly outside its scope, and two of our three sources say such models should not be used in critical GMP applications at all, with generative AI and large language models treated the same way.

The FDA's framework shows the same assumption in a different way. One analysis notes that the draft's vocabulary dates from 2024, that its worked architecture is a convolutional neural network, and that nothing in it tells a sponsor how to define a test set, a reference standard or a confidence interval for free text output. Commenters reportedly asked the agency for clarity on generative and foundation models. The EMA reflection paper, finalised in September 2024, does not address generative AI in any of the sources we read.

So the position in late 2026 is that three regulator frameworks describe how to validate a classifier, and the technology arriving fastest in pharmaceutical organisations produces prose. That is not a criticism of the regulators, who are reportedly considering the question, with an agency workshop in mid-2026 examining whether dynamic, adaptive and probabilistic models could be addressed under risk-based safeguards. It is a practical problem for anybody who has to govern a generative system this quarter.

Testing and evaluation are different disciplines

The distinction that resolves most of the confusion is between testing a model and evaluating one, and conflating them produces governance documents that cannot be executed. Testing asks whether a system meets a specification on a held-out set with known correct answers. Evaluation asks how a system behaves across a space of inputs where correctness is a judgement rather than a label.

DimensionA model you can testA model you can only evaluate
Correct answerExists, as a label, produced by an expert or a validated methodOften a judgement. Several outputs may be acceptable and several unacceptable ones plausible
DeterminismIdentical input gives identical output, which the draft annex requiresIdentical input may give different output. Temperature, context and model version all move it
Performance metricPrecision, recall, accuracy, with confidence intervalsA rubric score, a pass rate against criteria, a human preference rate. All noisier and all defensible if defined in advance
The evaluation setA held-out sample of the real input distributionA curated set of cases including adversarial and edge cases, built deliberately rather than sampled
Failure modeMisclassification, usually visible in aggregatePlausible, fluent and wrong. Invisible in aggregate and obvious in the single case that reaches a regulator
What a confidence score meansA calibrated probability, usable as a thresholdFrequently not calibrated. Treat a self-reported confidence with suspicion unless you measured it
Regression on changeRerun the test setRerun the evaluation set, and accept that the result moves. Define the tolerance in advance
Governance consequenceAcceptance criteria and a confusion matrix workAcceptance criteria still work. The matrix does not. Write the criteria in the language of the evaluation

The last row is the practical reconciliation and it is more optimistic than it sounds. The draft annex's requirement, that acceptance criteria be agreed in advance and set at least as high as the process being replaced, survives the move from testing to evaluation completely intact. What changes is the metric, not the discipline. A generative system for promotional copy can perfectly well be held to a pre-agreed pass rate against a published rubric, measured against the pass rate of the copy the agency currently produces, on a curated evaluation set built before anybody sees results.

What does not survive is the pretence. A governance document promising validation of a large language model, with a confusion matrix and a precision target, describes an exercise that cannot be performed. Teams write those documents because the template demanded them, and the result is a control that looks rigorous and is never executed, which is worse than an honest evaluation plan with softer numbers.

Generative AI in a GxP environment

Generative systems and regulated manufacturing are currently in an unresolved relationship, and the responsible position is conservative and documented rather than clever. The draft annex covers static deterministic models and, on two of our three sources' reading, says dynamic and probabilistic models should not be used in critical applications. One source puts it more strongly as a prohibition, which is a wording difference worth noting rather than relying on.

UsePosition we would takeReasoning
A generative model making or informing a batch release decisionDo not, and record that you considered and declined itCritical GMP application, and the draft annex reportedly covers only static deterministic models. Arguing the point after an incident is not a position
A generative model drafting a deviation investigation narrative, reviewed and approved by a qualified personDefensible with controls, and it is the human who signsThe output is a draft and the decision stays human. Model influence is low even where consequence is high
A generative model summarising documents for an internal readerProportionate governance. Register entry, owner, and a note on known limitationsLow influence and recoverable consequence
A generative model drafting regulatory submission contentTreat as high regulatory impact and apply the full gateThe EMA vocabulary fits here precisely. A substantial impact on regulatory decision-making without necessarily touching patient safety
A generative model producing promotional copy for publicationHuman review before publication, with a reviewer who can rejectNo AI framework reaches it and the marketing code does. The reviewer is the control
A conversational system answering a medical information queryThe hardest case on this list. Treat as high risk or do not deployA plausible wrong answer to a clinical question is the specific failure mode generative systems produce best. Pharmacovigilance capture is a second obligation layered on top

The second row is where most of the honest value sits and it is worth generalising. Generative systems are strongest where a human remains the decision-maker and the model reduces effort rather than exercising judgement. In the FDA's vocabulary, that is a system with low model influence, which means the evidence burden is low even where the decision consequence is high. Designing deployments deliberately into that quadrant is a governance strategy and not a compromise.

The last row is the one to argue about internally before a vendor demonstration rather than after. A conversational system that answers a clinician's question about a product has high influence over a decision with high consequence, produces outputs that cannot be exhaustively tested, and sits inside a regulated activity with an adverse event reporting obligation attached. Every property that makes generative AI attractive makes that specific deployment harder to defend, and companies that have thought about it carefully have mostly concluded that retrieval with citation and a human in the loop is the only version they would sign.

A note on the direction of travel

The frameworks are moving, and the movement is toward accommodation rather than prohibition. An agency workshop reportedly held on 30 June and 1 July 2026 considered whether dynamic, adaptive and probabilistic models, including generative AI and large language models, could be addressed under risk-based safeguards, and one source reports that consultation responses supported enabling them.

One of our sources is explicit that this is not an agreed change of scope, and nothing has been adopted.

What that means practically is that a conservative position today is unlikely to look foolish later, because the likely change is a route to permission with conditions rather than a reversal. Build the evidence you would need under those conditions now.

GET ACTIONABLE AI AND HEALTHCARE INSIGHTS THAT HELP YOU STAY AHEAD IN A FAST-CHANGING PHARMA WORLD

Who owns this, and how the standards fit

Accountability is the part of AI governance that organisations most often design badly, usually by creating a committee and declaring the problem solved. The regulator documents point the other way. The EMA paper reportedly places responsibility on named parties, the marketing authorisation applicant and holder and, in the final version, explicitly the clinical trial sponsor and the manufacturer. Responsibility attaches to the entity accountable for the activity, not to a governance function that oversees it.

RoleWhat they ownWhat they must not own
System owner, named individualThe register entry, the classification, the performance floor, the review date and the decision to keep it runningThe technical build. Owning both is how a model gets approved by the person who wants it approved
Independent reviewerChallenging the classification and the evidence, with authority to send it backNothing else. The moment the reviewer has delivery accountability the review stops being independent
Governance forumSetting the bands, resolving disputes, approving high and highest risk systems, and seeing the registerOwning individual systems. A forum that owns everything owns nothing
Data protection officer or equivalentThe lawful basis for inputs, rights handling, and the link between the data and model registersThe model performance question, which is not a data protection judgement
QualityWhere the activity is GxP: change control, deviations and the existing validation frameworkCommercial AI, unless the organisation deliberately extends the quality system to it
The commercial functionIts own systems, under the same gate. This is the population most likely to be unregisteredAn exemption. The absence of an AI rule is not an exemption from the other rules
Executive sponsorSigning for the highest tier, and the budget for the register to existRoutine approvals. Escalating everything to an executive is how a gate becomes a rubber stamp

The second row is the structural test of whether a framework is real. An independent reviewer who reports to the person delivering the project is not independent, and in most organisations that is exactly the reporting line. The cheapest fix is to draw reviewers from a different function and rotate them, which costs nothing and changes the conversation in the room.

ISO 42001 or the NIST framework, and what each actually buys

The two voluntary standards are frequently presented as alternatives and they answer different questions, which makes the choice easier than the comparison content suggests. One is a certifiable management system standard published in 2023, structured like other management system standards, with reported external audit and annual surveillance and a certificate at the end. The other is a free, voluntary risk management framework published in January 2023, organised around four functions, self-attested, with no audit layer unless an organisation buys one.

QuestionISO/IEC 42001NIST AI Risk Management Framework
What is it?A certifiable management system standard for an AI management systemA voluntary risk management framework built around govern, map, measure and manage
What do you get at the end?A certificate, reportedly via auditors qualified under a companion standard, with annual surveillanceA way of working. No certificate, and self-attestation unless you commission assurance
Who is it for?The board, the customer and the partner who want external assuranceThe engineering and risk teams who need a usable structure
Cost profilePurchase plus audit. One vendor estimate put certification at six to twelve monthsFree to download. The same vendor estimated implementation at six to nine months
PrescriptivenessClauses and an annex of controls across reported areas including policy, impact assessment, lifecycle, data and third partiesFlexible and sector agnostic, tailorable to size and maturity
Does it satisfy the EU AI Act?No source we read made that claim, and none described any presumption of conformityNo. One source describes it as complementing rather than replacing legal obligations
Does it address generative AI?Not specifically in any source we readNot specifically in any source we read
Does it address life sciences?No. None of our sources discussed pharmaceutical or GxP adoption at allNo, same position. Sector mapping is left to the organisation

The sixth row is the one to take to anybody proposing certification as a compliance strategy. No source we read claimed that conformity with the management system standard creates any presumption of conformity with the EU AI Act, and one source was explicit that the voluntary framework complements rather than replaces legal obligations. Certification is assurance, not conformity, and treating a certificate as regulatory cover is a mistake that only becomes visible under scrutiny.

The last two rows together explain why this page builds its framework from the regulator drafts rather than from either standard. Neither voluntary standard addresses generative AI specifically and neither addresses life sciences, which means an organisation adopting one still has to answer the two questions that actually matter in pharmaceutical AI. The regulator documents, for all that none of them binds you, were written by people thinking about patient harm, and that shows in the specificity of what they ask for.

The review cadence that survives contact with a calendar

Periodic review is in every framework and happens in very few organisations, because an annual review of everything arrives as one unmanageable quarter. Tying cadence to the risk class spreads it and makes it proportionate.

Risk classReview cadenceWhat the review actually checks
LowAnnually, or on changeIs the classification still right, and is the owner still here?
MediumEvery six months, or on changePerformance against the floor, drift in inputs, and whether the human check is being performed rather than recorded
HighQuarterly, plus on any change to model, inputs or processAll of the above, plus the evidence base, the limitations statement and the rollback path
HighestQuarterly with an independent reviewer present, and on every changeAll of the above, plus a sample of individual outputs and their sign-offs
Any class, on triggerImmediatelyA floor breach, an input distribution shift, a complaint, a vendor model version change, or a regulatory development

The vendor model version change trigger in the last row deserves its own sentence because it is the one organisations have no process for. A bought system whose underlying model is updated by the vendor has changed, often without notice and sometimes without a version number the customer can see. An organisation that cannot tell which model version produced last month's outputs cannot meet any of the change control expectations in any of these frameworks, and the time to raise that is during procurement.

What to measure

Governance reported as maturity produces maturity scores. The measures below are counts and dates, which are harder to present well and considerably harder to fake.

 

Stop reportingStart reportingWhy
Maturity score against a frameworkNumber of AI systems in the register with a named ownerMaturity is a judgement. A register entry exists or it does not
Policies publishedShare of production systems with a classification and a documented performance floorThe floor is the control. A policy without floors governs nothing
Training completedShare of systems whose floor was set against a measured baselineThe baseline is the step everybody skips, and the number will start near zero
NothingCount of commercial AI systems, reported separately from regulated onesThe ungoverned population. Reporting it separately is what gets it governed
NothingOverdue reviews, and the age of the oldest oneA dated commitment that visibly ages is worth more than a cadence policy
NothingFloor breaches, and what happened nextThe only evidence that the monitoring is connected to a decision rather than to a dashboard
NothingSystems with no identifiable model versionCatches the vendor update problem, which is invisible until somebody counts

The fourth row is the measure we would introduce first in a pharmaceutical company. Reporting the commercial AI count separately from the regulated AI count makes the asymmetry visible to an executive in one line: a regulated population that is small, documented and validated, and a commercial population that is larger, growing faster and governed by nothing. That single comparison does more to fund a governance programme than any maturity assessment.

Limitations, and what would change our mind

LimitationWhy it mattersHow it is handled here
No primary document was readThe FDA guidance text, the draft annex, the EMA paper and both standards were all read through secondary summaries. The FDA page we reached gave title, status and abstract but not the bodyStated in a callout after the introduction and again here. Structure is well attested across sources; wording is not. Verify before building a programme on it
The FDA step names are a reproductionThe seven steps come from a consultancy's list, not from the FDA's own wording, and a second source gave slightly different headingsUsed for structure rather than quoted as the agency's language. Take the names from the guidance itself
Draft annex section numbers are omittedOne source supplied numbers and described them as its own paraphrased numbering; two others gave noneDeliberately not reproduced. The requirements are described without citing sections we cannot verify
One source gave requirements the others did notTraining data rules, parallel running and model removal appear in a single vendor source, and cannot be cross-checkedFlagged at the point of use. The removal control is recommended on its merits rather than on its provenance
Sources disagree on the strength of the generative AI exclusionOne says not permitted and calls it a prohibition; two say should not be used. The difference matters to anybody arguing a case internallyThe weaker wording is reported alongside the stronger, and the recommended position is conservative either way
Status is verified only to September 2026The latest status statement we found for the draft annex was 24 September 2026, and for the FDA guidance early September 2026Dated explicitly. Both should be re-checked before publication, since the whole page turns on nothing having been finalised
The evidence tiers are oursThe four tiers are our construction, assembled from what the regulator documents require at the top end and judgement at the bottomLabelled as our construction at the point of use. They are a proposal, not a standard
Standards clause references conflictedThree sources gave three different numbering conventions for the same management system standard, and one contradicted itselfNo clause numbers are reproduced. Control areas are described in words
We sell AI services to pharmaWe benefit if organisations build AI governance programmes, and we supply commercial AI of exactly the kind this page says is ungovernedDeclared in the closing box. Note that the page's conclusion is that the biggest gap is in the commercial systems our own category sells

What would change our mind. Adoption of the GMP annex with dynamic and probabilistic models brought inside it under safeguards would convert the most conservative recommendation on this page into an obsolete one, and the mid-2026 workshop suggests that is the likelier direction. A final FDA guidance that addresses generative and foundation models would close the largest substantive gap in the regulator frameworks and would change what a credibility plan looks like for the systems being deployed fastest. An official statement that conformity with the management system standard carries any weight under the EU AI Act would change the build-or-certify calculation materially. And any regulator addressing commercial pharmaceutical AI directly would remove the central argument of this page, which is that the largest AI population in the industry is governed by instruments that never mention it.

Conclusion: govern the decision, not the technology

Responsible AI in pharmaceutical companies has been discussed as an ethics question for several years and is better treated as a decision architecture question. The regulator frameworks that exist all point the same way. The FDA's formulation makes risk a property of the decision rather than of the model, scaling the evidence owed with how much the model influences the outcome and how badly a wrong outcome ends. The draft manufacturing annex makes the standard explicit by requiring acceptance criteria to be set in advance and to be at least as high as the process being replaced. The EMA frames risk as patient risk and regulatory impact, which are both consequences rather than technical properties.

None of those documents is binding, and the useful response is to borrow rather than to wait. A register with named owners, a two-axis classification, a performance floor set against a measured baseline, a specified human check and a monitoring plan with a defined response constitute better governance than any principles document, and all five parts come straight out of frameworks written by people thinking about patient harm.

The gap that matters most is not in the frameworks but in what they do not reach. Commercial AI, which in most pharmaceutical companies is the largest population of deployed models and nearly all of the generative systems, sits outside every framework on this page. What governs it is data protection law over its inputs, the marketing codes and professional conduct rules over its outputs, and ordinary liability over its consequences, none of which mentions AI and all of which apply in full. The absence of an AI rule is not an exemption from the rules that exist.

And the technology has moved past the frameworks, which has a practical rather than a philosophical consequence. The drafts assume a model you can test against a held-out set with known answers, and pharmaceutical organisations are deploying models that can only be evaluated against a curated set with a rubric. The discipline survives that transition intact and the vocabulary does not, so write acceptance criteria in the language of the evaluation you can actually perform rather than promising a confusion matrix for a system that produces prose. A governance document describing an exercise nobody can execute is worse than an honest one with softer numbers.

Frequently Asked Questions For Responsible AI in Pharma: A Governance Framework

A decision architecture rather than an ethics statement. The working definition that survives contact with a regulator is that every AI system has a named owner, a classification based on how much the model influences a decision and how badly a wrong decision ends, a performance floor agreed before testing and measured against the process it replaces, a specified human check, and monitoring with a defined response when performance falls below the floor. Those five things come from the published regulator frameworks rather than from an ethics literature. A company with all five running is better governed than one with a principles document and no register of what it has deployed.

None of the healthcare-specific ones are binding yet, which is the central fact to plan around. The FDA's credibility assessment guidance was published in draft in January 2025, is marked not for implementation with non-binding recommendations, and was still draft as of the sources we checked in September 2026 with no public finalisation commitment. The EU GMP annex on artificial intelligence went to consultation from 7 July to 7 October 2025 and had not been adopted as of late September 2026. The EMA's reflection paper, finalised on 9 September 2024, describes expectations rather than creating obligations. The EU AI Act is binding, and its high-risk obligations were reportedly deferred to December 2027 and August 2028, with commercial analytics listed among systems in neither high-risk annex.

Nothing AI-specific, because no framework reaches commercial AI at all. The FDA guidance reportedly excludes operational efficiency and drug discovery, one analysis says the draft GMP annex does not apply to administrative functions, business intelligence or research unconnected to manufacturing, and the EMA paper does not discuss promotional content or commercial analytics in any source we read. What binds commercial AI instead are instruments that never mention it: data protection law over the inputs, the pharmaceutical marketing codes and professional conduct rules over the outputs, electronic communication rules over the sending, and ordinary liability over the consequences. None of those accepts a model as the author of a claim, and the absence of an AI rule is not an exemption from the rules that exist.

On two axes rather than one, using the FDA's formulation: model risk is model influence multiplied by decision consequence. Model influence is how much weight the output carries relative to other evidence in the decision; decision consequence is the severity and probability of an adverse outcome if the decision is wrong. The useful consequence is that risk is a property of the decision, not of the model, so a high-consequence decision with an independent downstream check can be medium risk while a low-stakes fully automated decision can be higher than it looks. European readers should also adopt the EMA's two words, high patient risk and high regulatory impact, which were reportedly chosen specifically to avoid confusion with the EU AI Act's own high-risk classification.
On two axes rather than one, using the FDA's formulation: model risk is model influence multiplied by decision consequence. Model influence is how much weight the output carries relative to other evidence in the decision; decision consequence is the severity and probability of an adverse outcome if the decision is wrong. The useful consequence is that risk is a property of the decision, not of the model, so a high-consequence decision with an independent downstream check can be medium risk while a low-stakes fully automated decision can be higher than it looks. European readers should also adopt the EMA's two words, high patient risk and high regulatory impact, which were reportedly chosen specifically to avoid confusion with the EU AI Act's own high-risk classification.

A performance floor agreed before testing and set at least as high as the measured performance of the process being replaced. That requirement comes from the draft EU GMP annex, applies to any system in any part of the business, and defeats the failure mode that makes most model approvals indefensible: after results exist, every model looks acceptable. The hard part is the baseline, because most organisations cannot say how well their current human process performs and measuring it is tedious and politically awkward. Doing it converts a subjective question, is this model good enough, into an objective one, is it better than what we do now.

Not for critical GMP decisions on the current draft, and the conservative position should be documented rather than assumed. The draft annex covers models that are static and deterministic, and two of our three sources report that dynamic, adaptive and probabilistic models, including generative AI and large language models, are outside its scope and should not be used in critical applications, with a third describing it as a prohibition. Generative systems are defensible where a human remains the decision-maker and the model reduces effort, such as drafting a deviation narrative that a qualified person reviews and signs, because model influence is low even where consequence is high. An agency workshop reportedly held on 30 June and 1 July 2026 considered whether such models could be brought in under risk-based safeguards, and one source is explicit that this is not an agreed change of scope.

You evaluate it rather than validate it, and the distinction is not pedantic. Validation in the regulatory sense assumes a held-out test set with known correct answers and a deterministic system that returns the same output for the same input; a generative system satisfies neither condition. What transfers intact is the discipline: acceptance criteria agreed in advance, an evaluation set built before anybody sees results including adversarial and edge cases, a rubric with defined pass criteria, and a measured baseline to beat. What does not transfer is the vocabulary. A governance document promising a confusion matrix and a precision target for a system that produces prose describes an exercise that cannot be performed, and teams write those documents because the template demanded them.

Only if you need external assurance, and never as a substitute for a conformity position. The standard is a certifiable AI management system standard published in 2023, with reported external audit, annual surveillance and a certificate; the NIST framework is free, voluntary, self-attested and closer to engineering practice. No source we read claimed that conformity with the management system standard creates any presumption of conformity with the EU AI Act, and one source described the voluntary framework as complementing rather than replacing legal obligations. Neither standard addresses generative AI specifically and neither addresses life sciences in any source we read, which means an organisation adopting either still has to answer the two questions that matter most in pharmaceutical AI.

A named individual per system, with a forum above them that owns the bands and the disputes rather than the systems. The common failure is a committee that owns everything, which in practice owns nothing and slows every decision without improving any of them. The EMA paper reportedly places responsibility on named parties, including the marketing authorisation applicant and holder and, in the final version, explicitly the clinical trial sponsor and the manufacturer, which is the same principle: accountability attaches to whoever is accountable for the activity. The structural test of whether a framework is real is whether the independent reviewer reports to someone other than the person delivering the project, and in most organisations today they do not.

Tie the cadence to the risk class, because an annual review of everything arrives as one unmanageable quarter and does not happen. Low risk annually or on change, medium every six months, high quarterly, and the highest tier quarterly with an independent reviewer present and a sample of individual outputs examined. The more important part is the trigger list: a floor breach, a shift in the input distribution, a complaint, a regulatory development, and a vendor model version change. That last one is the trigger organisations have no process for, because a bought system whose underlying model is updated by the vendor has changed without notice, and a company that cannot say which model version produced last month's outputs cannot meet any change control expectation in any of these frameworks.

Nobody measured the baseline, nobody logged the outputs, and nobody counted the commercial systems. Without a measured baseline there is no defensible floor and every approval is retrospective justification. Without logs there is no way to investigate a complaint about generated content, which cannot be reproduced by rerunning the system, so an organisation six months later has no record of what it actually sent. Without a count of commercial AI systems, the largest and fastest-growing population of models in the company is invisible to the governance function. All three are cheap to fix and none of them requires a framework, a standard or a consultant.

Count the AI systems in the commercial organisation and find out who approved each one. In most pharmaceutical companies the honest answers are more than anybody expects and nobody in particular, because commercial AI arrives through marketing platforms, agency tooling, embedded vendor features and individual experimentation, none of which passes a governance gate because no gate exists for it. The count is the finding, and reporting it beside the regulated AI count makes the asymmetry visible to an executive in one line. The second action is to add a named owner to every entry, and the third is to require a measured baseline before any further model is accepted. None of those three requires a policy to be written first.

GET ACTIONABLE AI AND HEALTHCARE INSIGHTS THAT HELP YOU STAY AHEAD IN A FAST-CHANGING PHARMA WORLD

Let's Discuss Your Requirements

+91
Contact Multiplier AI