Compliance 16 min read

AI Data Integrity in Pharma: A Practical Guide

J

November 14, 2026

What does AI data integrity mean in pharma?

AI data integrity in pharma means that every record an AI system creates, changes, or relies on still meets the same standard as any other GxP record: attributable, legible, contemporaneous, original, and accurate, plus complete, consistent, enduring, and available. The AI does not get a separate rulebook. It gets the existing one, applied to a system that behaves differently from a spreadsheet or a LIMS.

FDA's Data Integrity and Compliance With Drug CGMP: Questions and Answers (December 2018) defines data integrity as the completeness, consistency, and accuracy of data, and says that data should be reliable and accurate. That definition does not mention who or what produced the data. If an AI model drafts a deviation summary, classifies a complaint, or suggests a CAPA root cause, the output becomes part of your quality record, and the record is held to that definition.

I have come to think the reason this topic feels harder than it is comes from a mismatch in how people picture the risk. They picture a robot making decisions. The more common problem is quieter: a person pastes batch data into a general-purpose chatbot, gets back a tidy paragraph, and nobody can later show what went in, what came out, or which version of the model produced it. That is a data integrity gap before any "AI risk" ever comes up.


Which regulations apply to AI in a GxP environment?

No single regulation is titled "AI in pharma quality," but several existing ones apply directly, and a few newer documents are filling in the AI-specific detail.

Source What it governs Why it matters for AI
21 CFR 211.68(b) Controls over computer and related systems Requires appropriate controls over computer or related systems so that changes to master production and control records or other records are made only by authorized personnel, that input and output are checked for accuracy, and that a backup of entered data is maintained
21 CFR 211.188 and 211.194 Batch production and laboratory records 211.188 requires batch records to document each significant step; 211.194 requires laboratory records to include complete data from all tests. The regulations do not mention AI; our reading is that AI-generated content used to create or support an entry in these records falls under the same requirements
21 CFR Part 11, 11.10(a) Validation of electronic systems Systems must be validated for accuracy, reliability, and consistent intended performance
21 CFR Part 11, 11.10(e) Audit trails Secure, computer-generated, time-stamped audit trails must record operator entries and actions that create, modify, or delete records
EU GMP Annex 11 Computerised systems Risk-based validation, data checks, audit trails, and supplier oversight
EU GMP Annex 22 (draft, consultation opened July 2025) Artificial intelligence in GMP Draft limits critical applications to static, deterministic models and excludes generative AI and LLMs from critical use
FDA draft guidance, January 2025 AI used to support regulatory decision-making for drugs and biologics Seven-step, risk-based credibility assessment built around the "context of use"
ISO/IEC 42001:2023 AI management systems (voluntary standard, not a GxP requirement) Clause 6.1.2 addresses AI risk assessment; Annex A includes controls for data used in AI systems. It can structure your approach but carries no regulatory weight on its own

Two cautions on that table. First, the Annex 22 text was a draft when the consultation opened in July 2025. As of this article's update date (2026-10-04), we have treated it as a draft and have not confirmed that a final text has been issued, so check the current status with the European Commission or EMA before you cite it in a procedure. The same applies to the FDA draft guidance, which we have treated as a draft. Second, FDA's January 2025 draft guidance is aimed at AI that supports regulatory decisions about safety, effectiveness, or quality. It is not a general rulebook for every AI tool in your facility, and it excludes uses like drug discovery and operational efficiencies that do not affect patient safety, drug quality, or study reliability. Where your use sits outside that scope, Part 11, Annex 11, and your own risk assessment still carry the load.


How does ALCOA+ apply to AI-generated records?

ALCOA+ is the clearest way I know to test an AI workflow, because each attribute can be turned into a plain question you can ask a vendor or an auditor can ask you. The MHRA's GxP Data Integrity Guidance and Definitions (March 2018) sets out these attributes in detail, and FDA's 2018 Q&A refers to ALCOA more briefly, as a description of what complete, consistent, and accurate data looks like.

Attributable

Who or what created the record? When AI drafts text and a human approves it, both facts need to be captured: the model or system that generated the content, and the named person who reviewed and accepted it. A signed record that hides the fact that a model wrote the first draft is attributable on paper and misleading in practice.

Legible and enduring

AI output has to remain readable and retrievable for the whole retention period. This sounds trivial until a vendor retires a model version. If the only way to reproduce an output is to call a model that no longer exists, you have an enduring-record problem. Store the output itself, not a prompt you hope will regenerate it.

Contemporaneous

The record should be made at the time of the activity. AI actually helps here when it is used well, for example by capturing a voice entry at the point of work instead of letting an operator reconstruct events from memory at the end of a shift. We cover that pattern in more detail in our piece on voice-first record creation and backfilled batch records.

Original

The original record, or a certified true copy, must be preserved. With AI, the original usually includes the input data, the prompt or configuration, and the raw model output before any human edit. If you keep only the polished final version, you have lost the original.

Accurate

Accuracy is where AI behaves differently. A deterministic calculation gives the same answer every time; a generative model can produce fluent, confident text that is simply wrong. For this reason, I would not allow a generative model to be the sole source of a number, a specification limit, or a conclusion in a GxP record without verification against a primary source.

Complete, consistent, available

Complete means the full context was captured, including inputs, outputs, edits, and approvals. Consistent means timestamps and sequence make sense across systems. Available means the records can be produced for inspection in a readable form. An AI tool that keeps its logs in a vendor dashboard you cannot export from fails the last test no matter how good the model is.


What audit trail does an AI system need?

Under 21 CFR 11.10(e), the audit trail must be secure, computer-generated, and time-stamped, and record changes must not obscure previously recorded information. For AI, I would extend the practical content of that trail to include the following:

  • The model name and version that produced each output
  • The input data or a reference to its exact location and version
  • The prompt, template, or configuration used
  • The raw output as generated
  • Every human edit, with user ID and timestamp
  • The reviewer's acceptance, rejection, or override, with the reason

Part 11 does not list every one of those fields by name. They follow from the requirement to record who did what and when, and from the need to reconstruct how a record came to be. When an inspector asks how a CAPA root-cause statement was drafted, you want to answer from the trail instead of from memory.

One point I would state flatly: an audit trail that a user with ordinary privileges can edit or switch off is not an audit trail. That is true of any system, and it matters more when the system is generating content.


Can you use ChatGPT or other general-purpose AI tools with GxP data?

In my view, not for regulated records, at least not without a controlled arrangement that you have assessed and documented. The reasons are specific. A consumer-grade tool typically offers no validated configuration, no change control over model updates, no audit trail you can retain as a GxP record, and no supplier agreement that addresses your data. Each of those maps to a requirement you already carry under Part 11 and Annex 11.

We wrote about this in more depth in why ChatGPT is not safe for quality management records. The short version: the problem is rarely malicious. It is a well-meaning analyst saving ten minutes, and a record trail that cannot be defended six months later.

The draft EU Annex 22 reflects the same caution. As published for consultation, it treats generative AI and large language models as unsuitable for critical GMP applications, and where they are used in non-critical settings it expects a qualified person to remain responsible for the result. Even if the final text changes, the direction of travel is clear enough to plan around.


How should you validate an AI system under GxP?

Validation of AI follows the same logic as validation of other computerized systems, with a few additions. Part 11 at 11.10(a) asks for validation to ensure accuracy, reliability, consistent intended performance, and the ability to discern invalid or altered records. Annex 11 asks for a risk-based approach. Neither document asks you to treat the model as a mystery.

Define the intended use first

Write down what the AI does, what decisions it informs, and what a human still decides. FDA's January 2025 draft guidance calls this the context of use, and I find it the most useful single idea in the document. A model that suggests a document category for a human to confirm carries a different risk from a model that dispositions a batch.

Match the validation effort to the risk

A simple way to think about it:

AI use Typical risk Reasonable control approach
Drafting meeting summaries, no GxP decision Low Supplier assessment, user guidance, human review
Suggesting NCR or deviation categories Medium Documented intended use, accuracy testing on representative data, mandatory human confirmation
Trending and flagging out-of-trend results Medium to high Challenge testing, defined thresholds, periodic performance review
Influencing product release or disposition High Full risk assessment, formal validation, locked model version, independent verification of outputs

Test with data that looks like your data

Acceptance testing should use representative records, including awkward ones: misspellings, abbreviations, incomplete entries, and edge cases. Set acceptance criteria before you run the test, and keep the raw results.

Lock the version and control changes

If the vendor updates the model, that is a change to a validated system. Your change control procedure has to capture it, assess its impact, and decide whether revalidation is needed. Ask vendors directly whether you can pin a model version and how much notice you get before changes.

Monitor after go-live

Models can drift as the data they see changes. Periodic review of performance, with defined metrics and a documented decision on what you will do if performance drops, belongs in the validation plan. For practical help with protocol drafting, see our article on validation and qualification protocol generation using AI QMS tools.


What are the most common AI data integrity failures?

These are the failure modes I consider most likely, based on my professional judgment rather than a formal data set, and none of them are exotic.

  1. Shadow use. Staff use public AI tools on their own to summarize records or draft investigations, and the company has no inventory of it. You cannot control what you do not know about.
  2. Missing human accountability. The record says "approved," but the reviewer clicked through without reading. Review has to be a real act with enough time and information to be meaningful.
  3. No retention of inputs and raw outputs. Only the final text is kept, so the origin of a statement cannot be shown.
  4. Unvalidated model updates. The vendor changes the model and nobody notices until results shift.
  5. Shared logins. A shared account defeats attribution regardless of how sophisticated the AI is.
  6. Training data of unknown quality. If a model learned from historical records that were incomplete or backfilled, it can reproduce those habits at scale.

That last one deserves a pause. If your historical records contain years of retrospective entries, an AI trained on them will learn what a "normal" record looks like, including the flaws. Cleaning the source data is part of AI data integrity, and it is often the least glamorous part of the project.


How do you build an AI data integrity control framework?

Here is the sequence I would follow for a company starting from scratch. It is deliberately modest, because a small framework that people actually follow beats an elaborate one that sits in a binder.

Step 1: Inventory every AI use

Survey departments, including quality, QC, manufacturing, regulatory, and supply chain. Ask what tools people are using today, officially or not. Record the use case, the data touched, and who owns it.

Step 2: Classify by GxP impact

For each use, ask whether the output feeds a GxP record or a decision affecting product quality, patient safety, or data reliability. Anything that does goes into the controlled group. Anything that does not should still have clear rules about what data can be entered.

Step 3: Set an AI use policy

Keep it short. Name approved tools, prohibited data types, review requirements, and who to call with questions. ISO/IEC 42001:2023 gives you a management system structure if you want a formal framework, and clause 6.1.2 on AI risk assessment is a sensible anchor for the risk assessment piece.

Step 4: Assess suppliers

For any vendor in the controlled group, review how they manage model changes, data segregation, audit trail export, security, and incident notification. Put the key points into a quality agreement. Annex 11 expects supplier and service provider oversight, and AI vendors are not an exception.

Step 5: Validate proportionately

Follow the risk-based approach in the validation section above.

Step 6: Train and verify

Training records should show who learned what, and the quality unit should spot-check that practice matches procedure.

Step 7: Review periodically

Add AI systems to your periodic and management reviews, using the monitoring approach described under post-go-live monitoring above.


What will inspectors ask about AI?

I do not have a list from any agency of standard AI inspection questions, so treat the following as my expectation based on how inspectors already examine computerized systems. If I were on the other side of the table, I would ask:

  • Show me your inventory of computerized systems that use AI.
  • What is the intended use of this system, and where is that documented?
  • How was it validated, and with what data?
  • How do you know the output is correct today, not just at go-live?
  • Who reviews the output, and how do I see that they did?
  • What happens when the vendor changes the model?
  • Can you reproduce how this specific record was created?

If you can answer all seven from documents instead of from recollection, you are in good shape.


Where does this leave a small or mid-sized manufacturer?

Smaller companies sometimes assume AI data integrity is an enterprise problem. My judgment is that the gap is likely wider for small teams, because one person often owns quality, validation, and IT at once, and there is no spare capacity to run a governance committee. The answer is not to avoid AI. It is to choose tools built for regulated records from the start, so the audit trail, access control, and version history are part of the product instead of an add-on you have to bolt on.

Whatever tool you consider, evaluate it against these criteria:

  • Can it show the model name and version for every output?
  • Does it retain inputs, prompts, and raw outputs, and can you export them?
  • Is the audit trail secure, time-stamped, and not editable by ordinary users?
  • Does it use individual logins with role-based access?
  • Can you pin a model version, and does the vendor give notice of changes?
  • Does the vendor provide validation documentation and a quality agreement?
  • Is human approval a required step in the workflow?

A purpose-built QMS is one way to meet these criteria. You can see how Nova QMS approaches them on the platform page. Start with your AI inventory either way.


Frequently asked questions

Does 21 CFR Part 11 apply to AI systems? Yes, when the AI creates, modifies, maintains, or transmits electronic records that are required under FDA predicate rules, or are submitted to FDA. Requirements such as validation (11.10(a)), audit trails (11.10(e)), and access controls (11.10(d)) apply to the system regardless of whether it uses AI.

Can generative AI be used in GMP environments? Possibly, in non-critical uses, with risk assessment, validation, and documented human review. See the discussion of general-purpose tools and draft Annex 22 above for the limits on critical uses.

What is ALCOA+ and why does it matter for AI? ALCOA+ is the set of data integrity attributes regulators use to judge whether records are reliable. AI-generated or AI-assisted records have to satisfy each one; the ALCOA+ section above turns each attribute into a question you can ask of an AI workflow.

Do I need to validate an AI tool that only drafts text for human review? It depends on whether the text enters a GxP record or informs a quality decision. If it does, you need a documented intended use, a risk assessment, and testing proportionate to that risk. If it does not touch GxP records at all, a lighter policy-based control may be enough.

What should an AI audit trail capture? The field list is in the audit trail section above. One practical addition: confirm you can export the trail in a readable form for inspection, not only view it in a vendor dashboard.


Last updated: 2026-10-04

Jared Clark, JD, MBA, PMP, CMQ-OE, CQA, CPGP, RAC, is Principal Consultant at Certify Consulting (https://certify.consulting).

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.