Compliance 13 min read

FDA AI Governance: A Practical Guide for Regulated Firms

J

November 14, 2026

Most of what gets written about FDA AI governance is really about one product type. The device people talk about predetermined change control plans, and the drug people talk about credibility assessments. In my view, that split is the first thing to fix, because FDA does not have one AI rule. It has a growing set of guidance documents, statutory provisions, and existing quality regulations that all land on the same question: can you show that you know what your AI does, what it was built on, and what happens when it is wrong?

This guide is for quality, regulatory, and operations leaders who need a governance program that works across product lines. I'll walk through what FDA has actually published, how the pieces differ for devices and drugs, where ISO/IEC 42001:2023 fits, and what an inspector is likely to ask when your quality system touches AI.

What does "FDA AI governance" actually mean?

There is no single FDA regulation titled "AI governance." The phrase describes the internal system a company uses to control AI that touches a regulated product or a regulated decision. That system has to satisfy whichever FDA requirements apply to the use case, and the use case matters more than the technology.

FDA's regulations are technology-neutral. A neural network that reads a chest X-ray is regulated as a device. A model that predicts batch failure in a tablet press falls under the drug CGMP framework. A large language model that drafts deviation reports is regulated by nobody directly, but the deviation report it helped write is a CGMP record, and the quality unit is still responsible for it under 21 CFR 211.22.

That last point is the one I see companies miss. The question to ask is not "is this AI regulated?" It is "what regulated record, decision, or product does this AI influence?"

Which FDA documents shape AI governance today?

Four documents carry most of the weight, and I'd read all of them regardless of which center you answer to. Three are FDA guidance documents; the fourth, the Good Machine Learning Practice principles, is non-binding and jointly issued.

Document Applies to Status Core idea
Predetermined Change Control Plans for AI-Enabled Device Software Functions (Dec. 2024) Devices Final guidance Pre-authorize defined model changes under FD&C Act section 515C
AI-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (Jan. 2025) Devices Draft guidance (as of this writing; confirm on FDA's site) Total product lifecycle expectations for AI submissions
Considerations for the Use of AI to Support Regulatory Decision-Making for Drug and Biological Products (Jan. 2025, Docket FDA-2024-D-4689) Drugs and biologics Draft guidance (as of this writing; confirm on FDA's site) Risk-based credibility assessment tied to context of use
Good Machine Learning Practice Guiding Principles (Oct. 2021) Devices, with wider influence (joint FDA, Health Canada, MHRA) Final, non-binding principles Ten principles for the AI lifecycle

Two other items sit underneath these. Section 520(o)(1)(E) of the FD&C Act defines when clinical decision support software falls outside the device definition, and FDA's clinical decision support software guidance explains how it reads that provision. FDA revised that guidance in early 2026, so rely on the current version on FDA's website rather than the September 2022 edition. And 21 CFR Part 11 still governs electronic records and signatures when an AI system generates or modifies a GxP record.

If you want a deeper walk through the submission side, I covered the technical requirements in our breakdown of FDA machine learning requirements for drug and device submissions.

How does FDA expect device manufacturers to govern AI?

For devices, governance is mostly about the total product lifecycle and the ability to change a model without restarting the clock.

Predetermined change control plans

Section 515C of the FD&C Act, added by the Food and Drug Omnibus Reform Act of 2022, lets a manufacturer get a plan for future modifications authorized as part of the original marketing submission. FDA's December 2024 final guidance describes what that plan should contain: a description of planned modifications, a modification protocol covering data management, retraining, performance evaluation, and update procedures, and an impact assessment.

The practical effect is that you can retrain a model on new data within the boundaries you committed to without a new submission for each update. But the boundaries have to be specific. A plan that says "we may retrain periodically" gives you nothing, and a plan that says what data sources, what performance thresholds, and what triggers a rollback gives you a lot.

Quality system alignment

The Quality Management System Regulation, 21 CFR Part 820, took effect February 2, 2026 and incorporates ISO 13485:2016 by reference. That matters for AI because design controls, risk management, and software validation are now assessed against the ISO 13485 framework. If your AI-enabled software function is a device, your design history needs to show how training data was selected, how it was kept separate from test data, and how you decided that performance was acceptable for the intended use.

The Good Machine Learning Practice principles

The ten GMLP principles are not binding, but they are the closest thing to a shared vocabulary. They include multidisciplinary expertise across the lifecycle, representative clinical study participants and datasets, independence of training and test sets, and monitoring of deployed models. When a reviewer asks how you addressed bias, the GMLP principle on representativeness is the framing they have in mind.

How does FDA expect drug and biologic sponsors to govern AI?

The January 2025 draft guidance for drugs and biologics takes a different path. Instead of a lifecycle submission pathway, it gives sponsors a seven-step, risk-based credibility assessment framework. The steps run from defining the question of interest and the context of use, through assessing model risk and building a credibility assessment plan, to executing the plan, documenting results, and deciding whether the model is adequate for that use.

Two concepts do the heavy lifting.

  • Context of use is the specific role and scope of the model in addressing the question. The same model can be low risk in one context of use and high risk in another.
  • Model risk is determined by two factors: how much the model output influences the decision, and how serious the consequence of a wrong decision would be.

A model that is the only evidence supporting a safety conclusion sits at the high end. A model whose output is confirmed by independent, well-controlled data sits lower. The guidance expects the depth of your credibility evidence to scale with that risk.

The draft guidance is also explicit about scope.

It addresses AI used to produce information supporting regulatory decisions about safety, effectiveness, or quality. It does not cover AI used purely for drug discovery or for operational efficiencies that do not affect patient safety, drug quality, or the reliability of study results. That boundary is useful, but I'd treat it carefully. A tool that "just" speeds up document drafting can still affect the reliability of a CGMP record.

Where does ISO/IEC 42001:2023 fit?

FDA has not adopted ISO/IEC 42001:2023, and a certificate does not satisfy any FDA requirement. What the standard gives you is an organizing structure for the parts of governance that FDA guidance leaves to you: who owns AI risk, how systems are inventoried, how impacts are assessed, and how changes are controlled.

The clauses I'd map first:

  • Clause 6.1.2 (AI risk assessment) and clause 6.1.4 (AI system impact assessment) give you a repeatable way to document the risk reasoning FDA expects in a credibility assessment or a device risk file.
  • Clauses 8.1 through 8.4 cover operational planning and control (8.1), AI risk assessment (8.2), AI risk treatment (8.3), and AI system impact assessment (8.4), which line up with change control and ongoing monitoring.
  • Annex A controls address data for AI systems, third-party and supplier relationships, and AI system lifecycle management.

If your company already runs ISO 13485 or a CGMP quality system, 42001 is easier to adopt than it looks, because the management system skeleton (policy, objectives, internal audit, management review, corrective action) is familiar. The new work is AI-specific: data provenance, model performance monitoring, and impact assessment.

What controls should a cross-product AI governance program include?

Whatever your product type, I'd build the program around the same six elements.

1. An AI inventory with context-of-use statements

You cannot govern what you have not listed. Record every AI or machine learning tool, including vendor tools embedded in software you already own, along with a plain-language statement of what decision or record it influences. Include the quiet ones: the scheduling optimizer, the complaint-triage classifier, the document summarizer.

2. Risk tiering tied to regulatory impact

Borrow the model-risk logic from FDA's drug guidance even if you make devices. Ask how much the output drives the decision and how bad a wrong answer would be. A tiering of three levels is enough to start. The tier then sets validation depth, review requirements, and monitoring frequency.

3. Data governance

Document where training, tuning, and test data came from, how they were cleaned, and how you prevent leakage between sets. For GxP use, apply ALCOA+ expectations to the data feeding the model, not only to the model's output. Reviewers and inspectors both tend to look here, because data problems are a common root cause of model failures.

4. Validation proportional to risk

Software validation under existing frameworks still applies. For a model, that means defining acceptance criteria before testing, testing on data the model has not seen, and documenting performance in the subpopulations or conditions that matter. A static model can be validated once and controlled through change management. A model that learns continuously needs a defined boundary and a monitoring plan, or it needs to be locked.

5. Change control and monitoring

Retraining is a change. Vendor model updates are changes. A prompt template change in a generative AI workflow can be a change. Route all of them through your existing change control process with a documented impact assessment. For devices, make sure the change falls inside your authorized PCCP or goes through the appropriate submission path.

6. Human accountability

Name an owner for every AI system, and make sure the quality unit retains authority over outputs that become GxP records. Under 21 CFR 211.22, the quality unit's responsibilities cannot be handed to a tool. If a person approves AI-drafted content, the record should show that a person reviewed it and what they reviewed.

What will an FDA investigator ask about AI?

Investigators do not need an AI-specific regulation to ask good questions. They use the ones that already exist. The list below is my inference from existing regulations, not drawn from specific Form 483s or FDA inspection guidance on AI. The questions I would prepare for look like this:

  • Show me how this system was validated for its intended use (21 CFR 211.68 for automated equipment in drug manufacturing, and the design control and software validation requirements under Part 820 for devices).
  • Who reviews and approves the output before it becomes a record or drives a release decision?
  • What happens when the model changes, and where is that documented?
  • How do you know the data it was trained on is representative of what it sees in production?
  • Can you reconstruct what the system did on a given date? This is a Part 11 audit-trail question.

If you cannot answer these from documents rather than memory, that is where a Form 483 observation comes from. For a sense of how that process unfolds, see what to expect during an FDA inspection.

Common governance gaps I keep seeing

Four patterns come up often enough that they're worth naming.

Shadow AI. Staff adopt a general-purpose chatbot to draft SOPs, summarize complaints, or write CAPA narratives. Nobody inventoried it, nobody validated it, and the output flows into controlled records. The fix is a clear policy on permitted uses, not a blanket ban that people quietly ignore.

Vendor opacity. A supplier embeds a model in a platform and updates it without notice. Your supplier controls should require change notification for any AI component, and your quality agreement should say so in plain words.

Validation at launch only. The model passes validation in month one and drifts by month nine because the input population shifted. Monitoring plans with defined thresholds and escalation steps close this gap.

Governance by committee with no owner. A cross-functional AI council meets monthly and produces minutes, while no individual is accountable for any single system. Committees help with policy; named owners make the system work.

How do you start if you have nothing in place?

I'd begin with a four-week sequence rather than a large program.

  1. Build the inventory and write a context-of-use sentence for each tool.
  2. Tier each tool by regulatory impact and rank the top few.
  3. For the highest-tier tools, confirm that validation, change control, and record review are documented, and fill the gaps.
  4. Write a short AI policy that references your existing quality manual, assigns ownership, and defines how new AI tools get approved.

Then decide whether to formalize against ISO/IEC 42001:2023. For many firms the right order is to get the controls working first and pursue the standard later, if customers or partners ask for it.

If your AI program touches submissions, inspections, or a quality system that is already under strain, our FDA compliance team can help you scope what needs to be in place first.

Where is FDA AI governance heading?

As of this writing, the two January 2025 documents are the ones I have treated as drafts, so check FDA's site for any finalization before relying on this. FDA has said in them that the framework will develop with experience. I expect the credibility-assessment logic from the drug side and the lifecycle logic from the device side to keep converging, because both are asking for the same thing in different words: evidence that scales with risk. Companies that build one governance program around that idea will have less to rework when the drafts are finalized.

I'd also be cautious about reading silence as permission. The absence of a specific rule for generative AI in a quality unit does not mean the use is unregulated. It means the existing rules apply and you are expected to work out how.

Author: Jared, Certify Consulting. Last updated: 2026-10-25

Frequently Asked Questions

Does FDA have a single regulation for AI governance?

No. FDA applies existing statutes and regulations to AI based on the product and use case. Key documents include the December 2024 final guidance on predetermined change control plans for AI-enabled devices, and the January 2025 draft guidances on AI-enabled device software lifecycle management and on AI to support regulatory decision-making for drugs and biologics.

What is a predetermined change control plan?

A predetermined change control plan (PCCP) is authorized under section 515C of the FD&C Act. It lets a device manufacturer pre-specify planned modifications to an AI-enabled device software function, along with the protocol and impact assessment for making them, so qualifying updates do not each need a new marketing submission.

What is the credibility assessment framework in FDA's drug AI guidance?

It is a seven-step, risk-based process in the January 2025 draft guidance (Docket FDA-2024-D-4689). Sponsors define the question of interest and the context of use, assess model risk based on model influence and decision consequence, plan and execute credibility activities, and document whether the model is adequate for its context of use.

Is ISO/IEC 42001:2023 required by FDA?

No. FDA has not adopted it and certification does not satisfy any FDA requirement. It is useful as a management system structure, particularly clause 6.1.2 on AI risk assessment and clause 6.1.4 on AI system impact assessment.

Do generative AI tools used in quality systems fall under FDA oversight?

The tool itself is usually not regulated directly, but records it helps create are. Under 21 CFR 211.22 the quality unit remains responsible for CGMP records, and 21 CFR Part 11 applies when electronic records are created or modified, so AI-assisted content needs validation, review, and an audit trail.

J

Jared Clark

Principal Consultant, Certify Consulting

Jared Clark is the founder of Certify Consulting, helping organizations achieve and maintain compliance with international standards and regulatory requirements.

Need Help Now? Don't Wait for the Blog.

If you're facing an FDA inspection, responding to a warning letter, or navigating a regulatory submission — schedule a free consultation with Jared Clark. Expert guidance, no obligation.