AI Output Retention Under GxP: What to Keep, for How Long, and in What Form

When Does an AI Output Become a GxP Record?

Most retention debates about AI start in the wrong place. Teams ask how long they need to keep “the AI logs,” as if the AI system were the unit of record. It is not. The unit of record is still the GxP activity: the batch, the test, the investigation, the study. An AI output becomes a record for the same reason any other piece of data does. It was created to meet a requirement, or a regulated decision relied on it.

Regulators on both sides of the Atlantic already have a test for this, written long before generative AI. The tools are new; the test is not.

The FDA Test: Data Generated to Meet a CGMP Requirement

FDA’s guidance on data integrity puts it in one sentence: “When generated to satisfy a CGMP requirement, all data become a CGMP record.”¹ The same guidance goes on to say the data must be saved at the time of performance, and it rejects the practice of recording data on scraps of paper that are discarded after transcription. It also says it is not acceptable to store electronic records in a way that allows manipulation without creating a permanent record.¹

Apply that to AI. If a model classifies a vial as a reject during visual inspection, the classification was generated to meet an inspection requirement. It is a record. If a language model drafts the narrative of a deviation report and an investigator edits and signs it, the signed narrative is the record, and the question becomes whether the draft that shaped it also needs to be kept. If a data scientist uses a chatbot to brainstorm possible root causes before starting an investigation and none of that output enters the file, the output is probably not a record, though the company should decide that on purpose rather than by default.

The EU Test: Raw Data Is Whatever Quality Decisions Rest On

EU GMP Chapter 4 approaches the same question from the decision end. The current chapter, in force since 2011, tells regulated users to define which electronic data count as raw data, and adds: “At least, all data on which quality decisions are based should be defined as raw data.”² The draft revision of Chapter 4, released for consultation with the draft Annex 11 and the new draft Annex 22, keeps that sentence and adds language written with AI in mind.³

Two additions matter most here. First, the draft states that accountability for the integrity of documents, records, or raw data produced or processed with artificial intelligence “rests with the regulated user.”³ Second, its description of the “contemporaneous” principle says documentation should be an accurate attestation of “what was done, or what was decided and why, i.e. what influenced the decision at that time.”³ An AI output that influenced a quality decision is, by that description, part of what the record has to show.

A note on status. The European Commission’s consultation on the draft Chapter 4, Annex 11, and Annex 22 opened on 7 July 2025 and closed on 7 October 2025.⁴ The consultation page gives no adoption or implementation date, and no final text had been published when this article was written. Treat the drafts as a clear signal of direction, not as requirements in force. The 2011 Chapter 4 and Annex 11 remain the operative EU texts.

Four Kinds of AI Output

In practice, AI outputs in a regulated company fall into four groups. The group decides whether the output is a record, and that decides everything that follows.

The table is a starting point. The final call belongs in the use case’s risk assessment, made by the process owner with quality, and written down. If your company keeps an AI use case register, the output type is a natural column to add.

Discarded Drafts and Regenerated Outputs

The hardest case is the output nobody kept. A user asks a language model to draft a section, does not like it, regenerates twice, and keeps the third version. Are the first two part of the record?

No regulation answers this for AI directly. The nearest existing rule is FDA’s position on reprocessed chromatography. Asked whether it is acceptable to save only the final results from reprocessed laboratory chromatography, FDA answers “No,” and says that if chromatography is reprocessed, written procedures must be followed and each result retained for review.¹ The concern is selection: a person running an analysis several times and keeping only the answer they preferred.

The same concern applies when an AI tool produces a decision input. If a model is run three times on the same image and gives two different calls, keeping only the favorable one is exactly the pattern regulators look for. For decision inputs, our view is that every run should be kept. For drafted text that a qualified person then reviews, edits, and signs, the risk is lower, and a documented rule such as “keep the as-generated version that was accepted for editing, plus a count of regenerations” is a defensible middle ground. Whatever the rule, it should be written in the procedure and enforced by the tool, not left to user habit.

Why You Cannot Regenerate Your Way Out of Retention

A common objection to keeping AI outputs is that they can always be produced again. Keep the inputs and the model, the argument goes, and the output can be recreated on demand. For most AI tools in use today, that argument fails on three counts.

Many Models Do Not Give the Same Answer Twice

Large language models generate text by sampling from a probability distribution. A 2025 study by Shyr and colleagues, which ran repeated trials of several language models on clinical diagnostic cases, states that “the stochastic nature of this process can lead to different outputs even when the input prompt, model architecture, and parameters remain the same.”⁸ The study, released as a preprint and indexed in PubMed Central, is one of several that quantify how often repeated runs disagree.

The draft Annex 22 draws the same line from the regulatory side. It applies only to static models with deterministic output, meaning models that “when given identical inputs, provide identical outputs.” It says models with probabilistic output are not covered and “should not be used in critical GMP applications,” and it names generative AI and large language models specifically.⁵ Where such models are used in non-critical GMP work, the draft expects qualified personnel to be responsible for checking that outputs are suitable, a human-in-the-loop arrangement.⁵

So for any generative tool, the original output is the only copy there will ever be. Rerunning the prompt produces a new output, not a copy of the old one.

Models Change and Retire

Even a deterministic model only gives the same answer if it is the same model. Hosted models are versioned, updated, and eventually retired by their providers. A tool configured today against one model version may be pointed at a newer version next year, and the old version may no longer be available at all. FDA’s draft guidance on AI credibility notes that “a model’s performance can change over time or across deployment environments” and describes life cycle maintenance as a set of planned activities to keep a model fit for its context of use.⁹ That guidance remains a draft, issued in January 2025, and is not for implementation, but the underlying point is a technical fact rather than a policy position.

What happens to records when a model is retired is covered in our article on decommissioning a GxP AI system. The point for retention during use is simpler: a plan that depends on rerunning a model years from now depends on a model that may not exist.

The Precedent in 21 CFR 211.68

US CGMP has a rule that shows what regulators accept when output is not kept. Section 211.68(b) requires that input to and output from computer systems “shall be checked for accuracy.” It then allows certain data, “such as calculations performed in connection with laboratory analysis,” to be eliminated by computerization, but only if “a written record of the program shall be maintained along with appropriate validation data.”⁶

That exception was written for deterministic calculations, where the program and its inputs fully determine the result. It is a reasonable model for a static, deterministic AI model under strict configuration control: keep the exact model version, the inputs, and the validation evidence, and the output can be justified. It does not fit a generative model whose output varies between runs. And even for a deterministic model, most companies will find it simpler and safer to store the output than to prove, years later, that they could reproduce it.

What to Keep: The Seven Parts of an AI Output Record

An AI output on its own is rarely enough. A classification that says “reject” means little without knowing which image was classified, by which model, at what confidence, and what the operator did next. FDA’s definition of metadata makes the same point about all data: it calls metadata “the contextual information required to understand data,” and gives the example that the number 23 means nothing without the unit.¹ The draft Chapter 4 adds: “A complete record of data generated electronically includes relevant metadata.”³

For AI outputs, the relevant metadata falls into seven parts. Not every use case needs every part at the same depth, but each should be considered and the choice documented.

The Human Decision Needs Its Own Record

Where a person reviews an AI output before acting on it, the review is itself a GxP activity and needs its own record. The draft Annex 22 is direct about this. When a model gives input to a decision made by a human operator, and the testing of the model has been reduced on that basis, “records should be kept from this process.” Depending on criticality, the draft says this may mean a consistent review or test of every output, according to a procedure.⁵

The draft Annex 11 sets out what an audit trail should capture when data changes: the user, what was changed including the old and new value, the date and time, and a reason for the change.¹⁰ An AI output that a reviewer edits is a change from an old value to a new one. Capturing the as-generated output alongside the approved version, with the reviewer and the reason, gives an inspector the same who, what, when, and why they expect from any other audit trail.

In the laboratory, this also lines up with an existing requirement. Section 211.194(a)(8) requires the initials or signature of a second person showing that original records were reviewed for accuracy, completeness, and compliance.⁷ Where an AI tool performs or supports a check that a person used to perform, the record should show which person reviewed the result and on what basis. FDA’s draft AI guidance makes a related point for evaluation: where the context of use involves a human in the loop, evaluation should consider “the performance of the human-AI team, rather than just the performance of the model in isolation.”⁹ You cannot evaluate the team later if the human’s decisions were not recorded. Our earlier article on human-in-the-loop requirements covers the review design itself; the retention point here is that the review leaves a record.

Keep the As-Generated Output, Not Only the Edited Version

A common design stores only the final, approved text. That loses the evidence of what the reviewer changed. If an AI draft of a deviation summary named the wrong root cause and the investigator corrected it, the correction is evidence that the review worked. If the draft and the approved text are identical across hundreds of records, that pattern is evidence of something else, and a quality team will want to see it before an inspector does.

Section 211.194(a) requires laboratory records to include “complete data derived from all tests,” and 211.194(a)(4) asks for a complete record of all data secured in the course of each test.⁷ Where an AI tool contributes to a test result, the as-generated output is part of that data. Outside the lab, the same principle follows from the requirement that records show what influenced a decision.

Reference, Do Not Copy, When the Source Is Already Controlled

Storing every input and every prompt in full for every output multiplies storage and creates new copies of data that already exist under control. Where an input is itself a controlled record, such as a batch record in the EBR or a result in the LIMS, a stable reference with a version is enough, provided the referenced record has an equal or longer retention period. The same goes for prompt templates. If templates live in a governed library with version history, each output can point to the template version used rather than carrying a copy. Our article on governing prompt libraries covers how to set one up.

Copy rather than reference when the source is volatile: a user’s free-text prompt, a retrieved passage from a document that may be revised, an image that is not otherwise stored. If the source can change or disappear before the output’s retention period ends, the output record needs its own copy.

How Long: Tying AI Outputs to the Retention Clock You Already Have

Once an output is identified as part of a record, its retention period follows. No GxP regulation sets a separate retention period for AI outputs, and none is needed. The principle is simple: an AI output inherits the retention period of the regulated record it supports. An output that informed a batch release decision is kept as long as the batch record. An output that fed an adverse event assessment is kept as long as the adverse event record.

The Retention Periods That Already Apply

The table below maps common AI use cases to the rule that already governs the underlying record. It is not a complete list of retention rules, and local law, product type, and company policy can extend these periods. It shows how the mapping works.

Two details in these rules deserve attention. First, the GLP rule is unusual in setting the shortest of several periods rather than the longest, and it notes that it does not override retention requirements in other regulations.¹² Second, the EMA clinical trials guideline asks sponsors and investigators to maintain an inventory of “all essential data and documents and corresponding retention periods.”¹⁴ If AI outputs are part of the trial record, they belong in that inventory.

When One Output Supports Several Records

AI outputs often serve more than one record. A trend analysis of environmental monitoring data might support a batch release decision, a deviation investigation, and an annual product review. Section 211.180(e) requires that records be kept so their data can be used to evaluate quality standards at least annually.¹¹ In these cases the output takes the longest applicable period. The simplest way to make that work is the seventh part of the output record, the link: if the output carries references to every record it supports, the retention engine can compute the latest end date instead of relying on someone to remember.

Validation evidence follows its own logic. Chapter 4 says raw data supporting process validation should be kept at least as long as the records for all batches released on the basis of that validation.² The draft Annex 22 asks that AI model test documentation be retained along with the intended use description, the test data, and the access and audit trail records for that data, “similarly to other GMP documentation.”⁵ Applied to AI, the model’s validation package lives at least as long as the last output that relied on that validated model version.

The AI Act’s Six-Month Floor Is Not Your GxP Clock

The EU AI Act includes its own record-keeping rules. Article 12 requires high-risk AI systems to allow automatic recording of events (logs) over the lifetime of the system.¹⁶ Article 19 requires providers to keep those logs, to the extent they control them, for a period appropriate to the system’s intended purpose “of at least six months, unless provided otherwise in the applicable Union or national law.”¹⁵

Whether a given pharma AI tool is a high-risk system under the AI Act is a separate question, and many GxP tools will not be. Either way, six months is a floor, and the Article itself defers to other applicable law. Where GxP rules require a longer period, the GxP period governs. The danger is a platform team that reads “six months” in an AI governance document and configures log deletion to match, not realizing that some of those logs are GxP records with a retention period measured in years.

Balancing Retention Against Data Protection

Keeping more context makes outputs easier to understand, but inputs and prompts often contain personal data: patient information in safety cases, employee names in deviation reports, free text typed by users. The EMA clinical trials guideline notes that retention periods “should respect the data protection principle of storage limitation.”¹⁴ The OpenTelemetry semantic conventions for generative AI, an open standard for AI telemetry that is still marked as in development, mark the capture of input and output messages as opt-in and warn that these attributes are likely to contain sensitive information, including personal data.¹⁷

The answer is to keep what the record needs and no more. Pseudonymize where the process allows it, reference controlled sources instead of copying them, and set operational telemetry to a short retention period separate from the GxP output record. Do not solve the privacy question by deleting outputs that are part of a record.

In What Form: Static, Dynamic, and True Copies

Retention is not only a question of time. A record kept for ten years in a form that has lost its meaning has not been retained in any useful sense. GxP rules give a fairly clear framework for form, built around the difference between static and dynamic records.

Static and Dynamic Records

FDA uses “static” to mean a fixed-data record such as a paper record or an electronic image, and “dynamic” to mean a record format that allows interaction between the user and the record content.¹ MHRA’s data integrity guidance uses the same distinction and gives the example that chromatography records, once printed or converted to a static format, lose the ability to be reprocessed.¹⁹ MHRA also states that information originally captured in a dynamic state should remain available in that state, and the draft Chapter 4 carries the same sentence into its description of the “original” principle.¹⁹³

Where does an AI output fall? The output text or classification on its own is closer to static. A drafted paragraph does not change when you look at it again. But the AI output record described above, with inputs, model version, retrieved context, reviewer decision, and links, is a structured set of data. Its value comes from being able to search, sort, and trend it: which outputs did reviewers overturn, which model version produced them, which input types caused low confidence. That makes the record as a whole dynamic, and it should be kept in a form that keeps those abilities.

The draft Annex 11 supports this reading. It says it should be possible to obtain a complete electronic copy of system data including audit trail data, and that “Flat and locked files are not acceptable, it should be possible to search and sort data.”¹⁰ For archived data, it asks that retrieval be possible in a format that allows searching and sorting, or export to a tool where this is possible.¹⁰

True Copies and Format Changes

US rules allow records to be kept as originals or true copies. Section 211.180(d) says records “may be retained either as original records or as true copies such as photocopies, microfilm, microfiche, or other accurate reproductions of the original records.”¹¹ FDA’s Part 11 scope and application guidance goes further: FDA does not intend to object if companies archive required electronic records to paper, microfilm, or a standard electronic file format such as PDF or XML, provided predicate rule requirements are met and the records “should preserve their content and meaning.”²⁰ The same guidance says FDA intends to exercise enforcement discretion on the Part 11 requirement for protecting records throughout the retention period, while companies must still meet predicate rule retention requirements.²⁰

The phrase that matters is “content and meaning.” FDA’s data integrity guidance defines it further: electronic true copies are acceptable if they preserve the content and meaning of the original record, “which includes all metadata required to reconstruct the CGMP activity and the static or dynamic nature of the original records.”¹ For an AI output, the meaning includes which model produced it, from what input, and what a person did with it. A PDF of a chat window that shows the text of an answer but not the model version, the source documents, or the reviewer’s decision does not preserve that meaning.

Printing a Chat Is Not Retention

It is tempting to meet retention needs by printing or exporting a conversation and attaching it to the deviation or the batch record. The draft Chapter 4 addresses the practice directly: data recorded or processed electronically “should not be converted to or stored in a paper form” unless it meets the hybrid system requirements or the conversion is validated or verified for accuracy.³ Even under current rules, a static printout of a dynamic record does not meet the requirement to keep the original or a true copy, as FDA has said of laboratory instrument data.¹

A printed or PDF export can still be useful as a human-readable view attached to a record. It should not be the only copy.

Protect Completed Records From Change

Once the regulated activity is complete, the output record should become read-only. The draft Annex 11 says that after a process is complete, such as the release of a product, GMP data and metadata including audit trails “should be protected from deletion and changes throughout the retention period,” either by setting the data to read-only where it was created or by moving it to a dedicated archive through a validated interface.¹⁰ When data are moved, the draft asks for integrity verification with a high degree of certainty, for example by checksum, before anything is deleted.¹⁰

The current Annex 11 already requires that stored data be checked for accessibility, readability, and accuracy, that access be ensured throughout the retention period, and that archived data remain retrievable when the system changes.²¹ For AI tools, where the application layer changes often, that last point is the one most likely to be missed. An output record that can only be read by the current version of a vendor’s interface is at risk the first time the interface is redesigned.

For long retention periods, format durability becomes its own design problem. Our article on designing GxP archives for 30 years covers format choice and migration planning in more depth.

Where AI Output Records Should Live

The form question leads to a location question. Most AI tools keep some history of their own, and it is easy to assume that history is the record. It usually is not.

The Chat History Is Not the Record

Chat-style tools keep conversation history for the user’s convenience. Enterprise platforms often add compliance retention on top. Microsoft’s documentation for Purview, for example, explains that retention policies can cover user prompts and responses for Microsoft 365 Copilot and other AI apps, and that the retained data is stored in a hidden folder in the user’s mailbox that compliance administrators search with eDiscovery tools.¹⁸ That is useful for legal holds and investigations. The same documentation warns that “Messages visible in your AI apps are not an accurate reflection of whether they are retained or permanently deleted for compliance requirements.”¹⁸

A store of this kind is organized by user and conversation, not by batch or deviation. It does not know which conversation supported which regulated decision, it does not capture the reviewer’s decision in a structured way, and its retention is set by an IT policy rather than by the GxP record. It can be a useful second copy. It is not a substitute for a GxP output record linked to the regulated activity.

The Record Belongs With the Regulated Activity

The better pattern is to write the AI output record, or a reference to it, into the system of record for the activity. The classification belongs with the inspection record in the MES or EBR. The drafted narrative and its review history belong with the deviation in the eQMS. The trend flag belongs with the environmental monitoring data in the LIMS. The full trace, with inputs, prompts, and retrieved context, can live in a dedicated AI record store, as long as the two are linked in both directions and both are under the same retention and access controls.

Chapter 4 already says: “It should be clearly defined which record is related to each manufacturing activity and where this record is located.”² For AI outputs, that definition has to include the trace store. Our reference AI audit trail shows one way to structure the trace so it can be joined back to the regulated record.

Backup Is Not Archive

AI platforms are often backed up as a whole, and that can lead teams to assume outputs are retained. FDA uses “backup” in 211.68(b) to mean a true copy of the original record maintained securely throughout the retention period, and states that temporary backup copies would not satisfy that requirement.¹ MHRA separates the two plainly: backups for recovery purposes “do not replace the need for the long term, retention of data and metadata in its final form for the purposes of verification of the process or activity.”¹⁹ The draft Annex 11 adds that backup retention should be set by risk, with examples running from a week for hourly backups to years for monthly ones.¹⁰ A platform backup rotated after 90 days is disaster recovery, not retention.

Hosted Services and Accountability

Many AI tools run as hosted services, and some keep outputs only on the provider’s side. The draft Chapter 4 is clear about who is responsible: if the regulated user relies on hosted services, it is the regulated user’s responsibility “to understand, approve and justify the control measures of the hosted service provider based on a service level agreement.”³ Combined with the draft’s statement that accountability for AI-produced records rests with the regulated user, this means a supplier’s retention settings cannot be treated as someone else’s problem.³

In practice, the supplier agreement for any AI tool that produces GxP records should answer four questions: where outputs and their context are stored, for how long, how the company can export them in a searchable format, and what happens to them if the contract ends or the model is retired. If the answers are “on our side,” “30 days,” “not easily,” and “deleted,” the company needs its own capture before the tool goes live.

A Five-Step Retention Design for AI Outputs

None of this requires a new program. It requires applying the company’s existing records management discipline to a new source of records. The five steps below can be run for one AI use case in a few working sessions, then repeated for each new tool as it enters the use case register.

On disposal, the draft Chapter 4 asks for a documented process to ensure that “the correct original records or true copies are disposed only after the defined retention period,” with measures to reduce the risk of deleting the wrong documents and controlled access rights for disposal.³ For AI outputs, the risk is less often premature deletion by a person and more often deletion by a platform setting nobody connected to GxP. The fifth step is where that setting gets found.

Where Retention Designs Usually Go Wrong

What Good Looks Like in an Inspection

The simplest way to check the design is to rehearse the question an inspector would ask. Pick a released batch, or a closed deviation, where an AI tool played a part. Ask the team to show, from the regulated record, the AI output that informed the decision, the input it was based on, the model version that produced it, the person who reviewed it, and what that person decided. If the team can do that in minutes, from the system of record, without asking IT to search logs, the retention design works. If it takes a ticket to the platform team, the design has a gap, even if every piece of data exists somewhere.

The rehearsal also shows whether the answer will still be available years from now. Ask the same question about the oldest AI-supported decision in the system. If the model version it references has already been replaced, can the team still read the record and understand it? That is the test the retention period is there to pass.

Conclusion

Retention of AI outputs is not a new regulatory category. It is the existing records framework meeting a new source of data. The FDA test for when data becomes a CGMP record, the EU rule that data behind quality decisions is raw data, the predicate rule retention periods, and the long-standing guidance on true copies and dynamic records already answer most of the questions. What they do not do is make the design choices for you. Deciding which outputs are records, what context turns an output into something an inspector can follow, and where that record lives are choices each company has to make and write down. The draft Annex 22, Annex 11, and Chapter 4 set out the same expectations: records from human review of AI outputs, searchable and complete metadata, read-only archives, and clear accountability with the regulated user, whatever the hosting model.

Sakara Digital works with pharma and biotech organizations on the records, validation, and governance side of AI adoption. If you are putting AI tools into GxP processes and want an independent view on which outputs your procedures need to capture and how to keep them usable for their full retention period, we are happy to have that conversation.

For Further Reading