Training Records for AI-Assisted GxP Work: Who Needs What

Why AI-Assisted Work Changes the Training Question

A note on terms first, because the phrase is ambiguous. In this article, “AI training records” means records of how people were trained to work with AI tools in GxP processes. It does not mean the records of the data used to train a model. Both matter, but they are different records, owned by different people, and they answer different inspection questions.

For most of the history of computerized systems in GxP, training followed a familiar pattern. A system went live, users were trained on the screens and the procedure, and the learning management system (LMS) recorded that each person had completed the course for the current SOP version. The system did what it was validated to do. If a user entered the right inputs and followed the procedure, the output was predictable. Training was about operating the tool correctly.

AI tools break that pattern in one specific way. The output is no longer fully predictable from the inputs, and some of it will be wrong in ways that look right. A large language model can draft a deviation summary that reads well and cites the wrong batch. A classification model can tag a complaint with a plausible but incorrect category. A retrieval tool can quote an SOP that has been superseded. The person working with the tool is no longer only an operator. They are also a checker, and in many workflows they are the main control between the model’s error and the GxP record.

The Control Moved to the Person

When a company decides that a human will review every AI output before it becomes part of a record, it has made a design choice: the human review is the control. That choice is often sound. It is also a choice that makes the reviewer’s competence a central part of the validation argument. If the reviewer cannot reliably spot the errors the tool tends to make, the control does not work, however well the procedure is written.

The draft EU GMP Annex 22 on artificial intelligence says this directly. Where a model gives input to a decision made by a human operator, and the effort to test the model has been reduced because of that human involvement, the draft says the training and consistent performance of the operator should be monitored like any other manual process.⁷ In other words, a regulator’s draft text already treats the human reviewer of AI output as a process step whose performance needs evidence, not just a signature.

Why the Purolea Letter Matters Here

In April 2026, FDA issued a warning letter to Purolea Cosmetics Lab, a manufacturer of homeopathic drug products, that included a section titled “Inappropriate Use of Artificial Intelligence in Pharmaceutical Manufacturing.” The firm told investigators it had used AI agents to create drug product specifications, procedures, and master production or control records. FDA wrote: “If you use AI as an aid in document creation, you must review the AI generated documents to ensure they were accurate and actually compliant with CGMP.”⁸ When investigators pointed out that process validation had not been done, the firm replied that it had not known about the requirement because the AI agent had not told it.

It is worth being precise about what the letter cited. The AI section cites 21 CFR 211.22(c), the quality unit’s responsibility for approving procedures and specifications, and 21 CFR 211.100 for the process validation the firm had not done. It does not cite the training rule in 211.25. But the letter’s closing expectation is, in substance, a training question. FDA wrote that any output or recommendations from an AI agent “must be reviewed and cleared by an authorized human representative of your firm’s QU.”⁸ An authorized reviewer is only as good as their ability to judge the output. The firm’s own defense (the tool never said validation was required) is a clear example of a user who did not have the GxP knowledge to review what the tool produced.

Purolea was an extreme case, with many basic cGMP failures beyond AI. Most pharma and biotech companies are nowhere near it. The lesson for them is narrower and more useful: when AI helps produce a GxP record, the company must be able to show that the people reviewing and approving that record were qualified to do so. That is a training record question.

What the Regulations Already Require

There is no single regulation titled “training for AI users.” There does not need to be one. The existing personnel requirements are written broadly enough to cover any task, and an AI-assisted task is still a task. The work is in reading them against the new activity.

United States

For drug manufacturing, 21 CFR 211.25(a) requires that each person engaged in the manufacture, processing, packing, or holding of a drug product have “education, training, and experience, or any combination thereof, to enable that person to perform the assigned functions.” It also requires that training cover the particular operations the employee performs and cGMP as it relates to their functions, and that cGMP training be conducted by qualified individuals on a continuing basis.¹ Section 211.25(b) applies the same standard to supervisors.

For electronic records, 21 CFR 11.10(i) requires “Determination that persons who develop, maintain, or use electronic record/electronic signature systems have the education, training, and experience to perform their assigned tasks.”² Note the three verbs: develop, maintain, and use. An AI tool that creates or modifies GxP electronic records falls within this, and so do the people who configure it, write its prompts, maintain its knowledge sources, and monitor it.

For nonclinical laboratory studies, 21 CFR 58.29 applies a similar standard and adds a documentation requirement: each testing facility must “maintain a current summary of training and experience and job description for each individual engaged in or supervising the conduct of a nonclinical laboratory study.”³ If AI tools are used in GLP data review or report drafting, that summary should reflect it.

European Union

EU GMP Chapter 2 requires training appropriate to the duties assigned, and adds a point that matters a great deal for AI-assisted work: continuing training should be given, “and its practical effectiveness should be periodically assessed.” The same paragraph says training records should be kept.⁴ Effectiveness, not attendance, is the standard.

The current Annex 11 (2011) says all personnel should have appropriate qualifications, level of access, and defined responsibilities to carry out their assigned duties, and calls for close cooperation among roles such as process owner, system owner, Qualified Persons, and IT.⁵ The July 2025 consultation draft of a revised Annex 11 is more specific. Its section 5.2 reads: “All parties involved with computerised systems used in GMP activities should have adequate system specific training, and appropriate qualifications and experience, corresponding to their assigned responsibilities, duties and access privileges.”⁶ The phrase “system specific” is the one to note. General AI awareness does not meet it.

The draft Annex 22, published for consultation alongside it, covers AI models used in critical GMP applications. It is still a draft with no final text or effective date, so it is guidance on direction rather than a binding requirement. Three of its provisions bear on training:

• Section 2.1 (Personnel): close cooperation among process subject matter experts (SMEs), QA, data scientists, IT, and consultants, and “All personnel should have adequate qualifications, defined responsibilities and appropriate level of access.”⁷
• Section 3.3 (Human-in-the-loop): where testing effort was reduced because a human makes the decision, the intended use should describe the operator’s responsibility, and the operator’s training and consistent performance should be monitored.⁷
• Section 10.5 (Human review): records should be kept of the human review process, which, depending on criticality and the level of model testing, may mean review of every output according to a procedure.⁷

The draft also addresses generative AI in its scope section. It says large language models should not be used in critical GMP applications, and that if they are used in non-critical applications, “personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use.”⁷ Whatever form the final Annex 22 takes on generative AI, the direction on reviewer qualification is clear.

Read together, these texts point to one conclusion. Regulators are not asking for a new category of AI training. They are asking companies to apply the existing personnel rules to a new kind of task, with evidence that fits the task.

Three Roles, Three Training Profiles

The most common mistake we see is a single “AI training” course assigned to everyone who has a license for the tool. That approach produces a tidy completion report and weak evidence. A lab analyst who uses an AI assistant to draft an investigation summary, the QA reviewer who approves that summary, and the system owner who decides which model version is in use have very different jobs. Their training records should look different.

A practical split is three profiles, plus a fourth for the technical people behind the tool.

Profile 1: Users

Users are the largest group and the one most companies already train. The gap is usually in content, not coverage. A good user curriculum for an AI-assisted GxP task covers five things:

1. Intended use and its boundaries. What the tool is approved to do in this process, and what it is not. If the approved use is “first draft of a deviation description for human editing,” the user must know that pasting the output straight into the record is outside it.
2. Known failure modes for this tool. Not generic warnings about hallucination, but specific examples from testing: the tool confuses two product codes, the tool drops units when summarizing tables, the tool cites the previous SOP version when the knowledge base is out of date.
3. The required checks. What the user must verify against source before relying on an output, written as steps in the procedure.
4. Recording AI use. How to show in the record that AI was involved, which version or configuration was used if the procedure requires it, and what the user changed.
5. What to do when the tool is wrong. How to report an error so that it reaches the system owner and feeds performance monitoring.

For users, a read-and-understand of the procedure plus a short knowledge check is often proportionate, as long as the procedure itself contains the failure modes and checks. The training record should link to the procedure version and the tool configuration in use at the time.

Profile 2: Reviewers and Approvers

This is the profile where most records fall short, and it carries the most risk. When a company reduces its testing of an AI tool on the grounds that a human reviews every output, the reviewer’s competence is part of the validation argument. The record needs to show more than course completion. It needs to show that this reviewer, on this kind of output, catches the errors that matter. The next section covers how to build that evidence.

Reviewers also need two things that users may not. First, they need the GxP domain knowledge to judge whether an output is compliant, not just whether it matches the source. The Purolea firm’s reviewer, if there was one, did not know that process validation was required. No amount of tool training fixes that. Second, reviewers need explicit authority and time. A reviewer who is expected to clear fifty AI-drafted summaries a day with no time budget for checking will not do deep checks, whatever the training record says.

Profile 3: Owners

Process owners and system owners make decisions that shape everyone else’s risk: what the intended use is, which outputs require full human review and which can be sampled, when a model or prompt change needs revalidation, and when performance trends call for action. The draft Annex 22 gives process SMEs specific responsibilities, including the adequacy of the intended use description and the definition of acceptance criteria.⁷

Owners do not need to be data scientists. They do need enough understanding to ask the right questions: what the test data covered, which subgroups performed worse, what the error rate was, and how the reviewers are performing. Their training record should show that they were trained on the tool’s validation summary and on the governance procedure they are responsible for, and it should be refreshed when either changes.

Profile 4: Builders and Maintainers

The people who configure prompts, manage retrieval sources, tune thresholds, or deploy model updates are “persons who develop, maintain, or use” the system under 11.10(i).² In many companies, these people work in IT or a data science group that has never been part of the GxP training matrix. Their technical skill may be excellent. What they often lack is training in change control, data integrity principles, and good documentation practice. A prompt edit made on a Friday afternoon without a change record is a likely data integrity finding, and it usually traces back to someone who was never told that the prompt is part of the validated configuration.

Evidencing Competence to Review AI Output

Training someone to review AI output is not hard. Proving that they can do it is. The difficulty comes from a well-documented human tendency: people tend to accept suggestions from automated systems, even when they have the knowledge to spot the error. This is called automation bias, and it is the main reason a signature alone is weak evidence.

What the Research Shows

A 2023 study in Radiology gives one of the clearest examples. Twenty-seven radiologists read 50 mammograms with help from what they were told was an AI system. For 12 of the mammograms in the test set, the system suggested an incorrect category. The share of mammograms rated correctly dropped sharply when the AI suggestion was wrong, for readers at every experience level.⁹

Experience helped, but it did not remove the effect. The most experienced readers still lost a large share of their accuracy when the tool was wrong. A GxP reviewer with fifteen years of QA experience is in the same position when an AI tool presents a confident, well-formatted, and incorrect summary.

An older review of the human factors literature by Parasuraman and Manzey reached a sobering conclusion about training specifically. Automation bias, they found, occurs in both naive and expert participants and “cannot be prevented by training or instructions.”¹⁰ That finding does not mean training is useless. It means training alone is not a sufficient control, and a training record alone is not sufficient evidence.

A systematic review by Lyell and Coiera adds a useful detail. Of 890 papers screened, 40 met their criteria, and they found automation bias in single tasks, not only in multitasking settings, and particularly in tasks with high verification complexity. They concluded that automation bias appears to be associated with cognitive load.¹¹ For GxP work, this points to a practical lever: make the reviewer’s checking task simpler and more specific, and automation bias becomes easier to manage.

The generative AI research points the same way. A 2025 Microsoft Research study surveyed 319 knowledge workers who shared 936 examples of using generative AI at work. Higher confidence in the AI tool was associated with less critical thinking, while higher self-confidence in the task was associated with more. The authors also found that generative AI shifts the worker’s effort toward verifying information, integrating responses, and overseeing the task.¹² That shift describes the GxP reviewer’s job exactly, and it is the skill most training programs do not assess.

ISPE’s Pharmaceutical Engineering made the same point in a GxP context. Writing about language models used in regulated operations, the authors noted that many applications include a human in the loop, but “the danger lies in the inherent risk that humans may not possess the expertise or diligence to properly evaluate the LLM’s output.”¹³

A Competence Check That Produces Real Evidence

The most practical way to evidence reviewer competence is a task-based assessment built from the tool’s own test results. The approach below works for most AI-assisted review tasks, from deviation summaries to complaint classification.

Ongoing Performance Is Part of the Record

Step five deserves more attention than it usually gets. A reviewer who passed a seeded assessment in March may be approving outputs without real checks by September. The research on automation bias suggests this is likely as people come to trust a tool that is right most of the time. The Lee study’s finding that higher confidence in AI goes with less critical thinking points the same way, although that study looked at one point in time and did not track people over months.¹²

Some useful signals are cheap to collect. If the tool logs whether a reviewer edited an output before approving it, a reviewer whose edit rate falls toward zero while colleagues on the same work still make edits is worth a conversation. If a second-level check (a QA spot check, a periodic review, an audit) finds an error that the first reviewer approved, that event should be linked back to the reviewer’s training file, not only to the deviation system. Occasional re-seeding, where a known-bad output is placed into the live queue under a controlled procedure, is a stronger test still, though it needs careful design so that seeded items never reach a real GxP record.

Detailed sampling plans for human review are a topic of their own. The point for training records is simpler: whatever monitoring you do on reviewer performance, it should be traceable to the individual and visible when someone asks whether that person is still qualified.

Reduce the Cognitive Load, Not Just Train Harder

Because training cannot fully prevent automation bias, the procedure and tool design must do part of the work. The Lyell and Coiera finding on verification complexity points to practical steps.¹¹ Give reviewers a short, specific checklist tied to the tool’s known failure modes instead of a general instruction to “review for accuracy.” Have the tool show its sources next to each claim, so checking is a comparison rather than a search. Limit the number of outputs a reviewer clears in one sitting. Each of these makes the reviewer’s job easier to do well, and each is easier to defend in an inspection than a claim that training alone made reviewers vigilant.

What a Good Training Record Contains

Most LMS platforms can hold the evidence described above. The gap is usually in how the curriculum and the records are structured, not in the software. Here is what a training record for AI-assisted GxP work should let you show, for any named person, on any date.

Tie Training to Change Control

AI tools change more often than most validated systems. A vendor updates the underlying model, a system owner revises a prompt, the knowledge base behind a retrieval tool is refreshed with new SOPs. Each of these can change what errors the tool makes, which means it can change what reviewers need to look for.

The fix is procedural. Every change control for an AI-assisted GxP tool should include a training impact assessment that asks two questions. Does this change alter the tool’s intended use, its outputs, or its known failure modes? If so, which roles need retraining, and does the reviewer assessment need a new seeded set? When the answer is yes, the change should not close until the retraining is complete for everyone who will use the updated tool. This is the same logic companies already apply to procedure changes. It simply needs to extend to model and prompt changes.

Do Not Forget Contractors and Service Providers

Contract staff, consultants, and service-provider personnel often do AI-assisted GxP work: a contract QA reviewer, a consultant drafting validation documents with an AI assistant, a CRO team member summarizing monitoring visit reports. The personnel rules apply to them too. The draft Annex 22 names consultants among the parties who need adequate qualifications.⁷ Your quality agreements and onboarding procedures should say whose training records cover these people, and you should be able to produce those records on request.

Keep Records for as Long as the Work They Support

A training record’s value comes from being able to show that a named person was qualified on the date they did a piece of work. That means training records for AI-assisted work should be kept at least as long as the GxP records that person produced or approved. If a batch record reviewed with AI help must be kept for years, the reviewer’s qualification evidence for that period must still be retrievable. This is easy to miss when training data lives in one system, AI usage logs in another, and the GxP record in a third.

Where EU AI Act Article 4 Fits

Many companies first heard about AI training through the EU AI Act, and some have treated its AI literacy requirement as the whole answer. It is not, and the July 2026 amendment makes the distinction clearer. A later article in this series covers how to design an AI literacy program. Here we only place Article 4 next to GxP training records.

What Article 4 Requires Today

Article 4 of the AI Act has applied since February 2, 2025. In its original form, it required providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and others dealing with the operation and use of AI systems on their behalf.¹⁵

The Digital Omnibus on AI, Regulation (EU) 2026/1744, changed that. It was published in the Official Journal on July 24, 2026, and entered into force on July 27, 2026.¹⁶¹⁷ The European Commission’s AI literacy Q&A, updated the same week, summarizes the amended Article 4 this way: AI literacy remains an obligation for providers and deployers, but no specific or “sufficient” level is mandated; providers and deployers must take measures to support the development of AI literacy; and the obligation does not require them to guarantee any specific level of AI literacy of any individual.¹⁴ Commentators have described the change as a move from an obligation of result to an obligation of means.¹⁶

Three other points from the Commission’s Q&A are useful for GxP teams:¹⁴

• No certificate is required. The Commission says organizations can keep an internal record of training and other guidance activities.
• Scope is broad. It covers anyone in the organization directly dealing with an AI system, and “other persons” can include contractors and service providers.
• National authorities enforce it. National market surveillance authorities supervise and enforce Article 4, starting in August 2026.

The Q&A also notes that a separate, stricter duty is unchanged: deployers of high-risk AI systems must ensure that staff dealing with those systems are sufficiently trained to handle them and to ensure human oversight, under Article 26.¹⁴ The high-risk obligations themselves were deferred by the same Omnibus, to December 2, 2027 for Annex III systems and August 2, 2028 for Annex I systems.¹⁷ Many GxP AI tools used in manufacturing and quality may fall outside the Act’s high-risk categories, but each company should confirm that for its own portfolio.

How Article 4 and GxP Training Records Relate

The comparison makes the relationship plain. Article 4 sets a baseline for everyone who works with AI, across the whole company, including functions with no GxP role. GxP personnel requirements are a stricter and narrower standard for the people whose AI-assisted work becomes part of a regulated record. A company can use one curriculum structure for both: a general AI literacy module for all staff, which supports the Article 4 duty, followed by role-specific GxP modules and assessments for users, reviewers, owners, and maintainers of each AI-assisted process. What it should not do is point an inspector to the AI literacy module as evidence that a QA reviewer is qualified to approve AI-drafted batch record reviews.

Gaps to Look For in Your Own Records

When we review training programs for AI-assisted GxP work, the same gaps tend to appear. None of them needs a new system to fix. Most need a clearer decision about roles and a change to how curricula are assigned.

The Generic Course

One “Responsible AI” or “Using Our AI Assistant” course is assigned to all licensed users, with no role differentiation and no link to specific GxP procedures. It may support the Article 4 duty. It does not show that anyone is qualified for a specific GxP task under 211.25 or the “system specific” language in draft Annex 11.¹⁶

The Invisible AI Use

Staff are using general-purpose AI tools to help draft GxP documents, but no procedure covers it and no training exists for it. This is the Purolea pattern at smaller scale.⁸ The first fix is not training. It is deciding which uses are allowed, writing them into procedures, and then training against those procedures.

The Reviewer Who Was Never Assessed

The validation package for an AI tool says human review is the key control, but the reviewers’ training records show only a read-and-understand of the SOP. Given what the automation bias research shows, this is the gap with the most risk and the one most worth fixing first.⁹¹⁰

The Stale Record

The tool’s model or prompt was changed through change control, but no one assessed the training impact. Reviewers are checking for last quarter’s failure modes while the tool now makes different ones.

The Missing Builders

The data scientists and IT staff who configure and maintain the tool are not in the GxP training matrix. They have never had data integrity or change control training, even though 11.10(i) names people who develop and maintain the system.²

The Disconnected Evidence

Training records are kept in the LMS, AI usage and edit logs in the tool, and GxP records in the eQMS or MES. No one can join them to show who was qualified for what on a given date. Each piece exists, but the story cannot be told on request.

The Unowned Contractor

Contract reviewers or consultants use AI tools on GxP deliverables, and neither the company nor the service provider can say whose training program covers them. The quality agreement is silent.

A Sequenced Plan for the Next Two Quarters

Most companies do not need a large program to get this right. They need a clear inventory, a role model, and a few well-designed assessments. The sequence below is written for a pharma or biotech quality organization with a handful of AI-assisted GxP processes, and it can scale up or down.

A note on sequencing. It is tempting to start with the LMS because it produces visible progress. Resist that. Training built before the procedure states the failure modes and checks will need to be rebuilt, and the records it produces will show completion of content that did not match the task. Procedures first, then curricula, then assessments.

How to Size the Effort

Not every AI-assisted task needs a seeded reviewer assessment. Use the same risk logic you apply to validation. A tool that drafts internal meeting notes for a quality team is low risk, and a procedure plus a knowledge check is enough. A tool that pre-classifies deviations by severity, where a misclassification could delay an investigation or a CAPA, is higher risk, and the reviewers need assessed competence. A tool whose output feeds batch disposition is the highest tier, and many companies will decide that it belongs in the scope of the draft Annex 22 controls rather than being treated as a human-in-the-loop convenience. The training evidence should scale with that tiering, and the tiering rationale should be written down.

Conclusion

The rules on GxP personnel did not need to change for AI, and they have not. What has changed is the work. When an AI tool drafts, classifies, or summarizes, the person who reviews its output becomes a control, and the evidence that the control works lives in that person’s training record. A course completion and a read-and-understand signature cannot do that job on their own, because the research on automation bias shows that knowledgeable, experienced people still follow confident, wrong suggestions. Records that inspectors can rely on split the work into roles, tie training to the procedure and tool version, assess reviewers on realistic errors, and link ongoing performance back to the individual. EU AI Act Article 4 is a useful baseline for everyone who works with AI, especially after the July 2026 amendment clarified that it is about supporting literacy rather than certifying it, but it is not the standard for GxP work.

Sakara Digital works with pharma and biotech organizations on exactly this kind of problem: turning AI adoption in quality and manufacturing into something that inspectors can follow and accept. If you are deciding how to structure training and qualification for AI-assisted GxP work and want an independent view on where to start, we are happy to have that conversation.

For Further Reading