Compliance 14 min read

AI CSA in Pharma: A Practical Guide to Assurance

J

November 14, 2026

Most pharma quality teams I talk to are asking the same question in slightly different words: if we want to put AI inside a GxP workflow, do we validate it the old way, or do we use computer software assurance? The honest answer is that CSA gives you a better way of thinking about it, but the document that made CSA famous was not written for pharma, and it was not written for AI. Knowing where it applies and where you are borrowing its logic is most of the job.

This guide walks through what CSA actually is, what the pharma-specific rules say, how AI changes the picture, and what I would put on paper so an inspector can follow your reasoning.

What is computer software assurance (CSA)?

Computer software assurance is a risk-based approach to establishing confidence that software used in production or the quality system does what you intend. FDA's CDRH issued the final guidance, "Computer Software Assurance for Production and Quality System Software," on September 24, 2025, after a draft in September 2022 (confirm the date and wording against the version on FDA.gov before citing it in a policy). It replaces Section 6 of the 2002 guidance "General Principles of Software Validation," the section that dealt with automated process equipment and quality system software.

The core idea is simple. You spend your effort in proportion to the risk that a software feature could compromise product quality or patient safety, and you stop producing documentation that proves nothing. The guidance can be summarized in four steps: identify the intended use, determine whether the feature is high process risk, choose assurance activities that fit that risk, and record the result in a way that shows what you did and why.

The shift most people notice is from scripted testing toward unscripted and exploratory testing for lower-risk features, plus a clear permission to use the vendor's own evidence. A screenshot of every step in a passing test case is not the point. A person with judgment looking at the software behave and writing down what they concluded is the point.

Does CSA apply to pharma, or only to medical devices?

This is where a lot of articles blur things. The CSA guidance comes from CDRH, and it is framed around device manufacturers and the quality system regulation, which became the Quality Management System Regulation (QMSR) effective February 2, 2026 (21 CFR Part 820 as amended by FDA's QMSR final rule). The QMSR incorporates ISO 13485:2016, and that standard's clause 4.1.6 is where the software validation requirement for quality system software now sits for device firms.

For drugs and biologics, the binding requirements are different:

  • 21 CFR 211.68 requires that automatic, mechanical, and electronic equipment, including computers, be routinely calibrated, inspected, or checked, and that controls exist over changes to computer systems.
  • 21 CFR 11.10(a) requires validation of electronic record systems to ensure accuracy, reliability, consistent intended performance, and the ability to discern invalid or altered records.
  • EU GMP Annex 11 covers computerised systems, with risk management across the lifecycle written into its general principles.

None of those texts prescribe scripted IQ/OQ/PQ with screenshots. FDA's 2003 guidance, "Part 11, Electronic Records; Electronic Signatures – Scope and Application," already said in its validation discussion that approaches should be based on a justified and documented risk assessment. So a pharma company using CSA is not doing something exotic. It is returning to what the regulations always allowed, with a recent FDA document that gives the approach a clearer vocabulary.

The practical position I recommend: treat CSA as an accepted risk-based method, cite your own regulatory basis (Part 11, 211.68, Annex 11) in your validation policy, and cite the CDRH guidance as supporting reasoning, not as the legal authority. ISPE's GAMP 5: A Risk-Based Approach to Compliant GxP Computerized Systems (Second Edition, 2022) takes a compatible position, encouraging critical thinking and assurance effort scaled to risk; check the edition's own text for the wording before quoting it.

How does AI change the assurance question?

Traditional software is deterministic. You give it the same input and get the same output, so a test you ran last year still tells you something about what it does today. AI, and particularly generative AI, breaks that assumption in several ways at once:

  1. The output can vary between runs with identical input.
  2. The behavior depends on training data you may never see.
  3. A vendor can update the model underneath you without a release note you would recognize as a change.
  4. The failure is often plausible rather than obvious, which is harder for a reviewer to catch than a crash.

CSA was built for the first kind of software. Its logic of intended use, risk, and proportionate evidence still holds, but the specific techniques need adjusting. You cannot assume that a handful of passing tests characterizes a probabilistic system, and you cannot assume the thing you tested is the thing running next quarter.

I think the most useful reframing is this: with AI you are not only assuring the software, you are assuring the combination of software, human reviewer, and the process around them. A model that drafts a deviation summary which a qualified person reviews and signs is a different risk than a model that decides batch disposition on its own. The assurance effort should follow that difference.

What do regulators say about AI in GxP right now?

Three documents matter, and none of them is the finished answer. The descriptions below reflect the documents as understood at the time of writing; verify current status before citing.

EU GMP draft Annex 22 on artificial intelligence. The European Commission and EMA published a draft Annex 22 for consultation in July 2025, alongside a proposed revision of Annex 11. The draft is deliberately narrow. For critical GMP applications, those that affect patient safety, product quality, or data integrity, it expects static models with deterministic output. It does not accept dynamic models that keep learning in use, and it excludes generative AI and large language models from critical applications. For non-critical uses, it contemplates such models only with a qualified human in the loop. It is a draft, so check the current status on the European Commission's EudraLex Volume 4 page before you cite specific wording in a procedure.

FDA draft guidance on AI to support regulatory decision-making for drugs and biologics. The guidance, "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products," was issued in January 2025; it proposes a seven-step, risk-based credibility assessment framework built around a defined "context of use." Its scope is AI used to produce information supporting decisions about safety, effectiveness, or quality. It explicitly leaves out drug discovery and operational efficiencies that do not affect patient safety, drug quality, or the reliability of results. So it is a useful model for thinking about credibility evidence, but it does not directly govern every AI tool in your QMS.

Existing GxP rules. Part 11, 211.68, and Annex 11 apply to an AI tool the same way they apply to any computerized system. Audit trails, access control, change control, and records integrity do not get waived because a model is involved.

A statement I would be comfortable defending in an inspection: under current rules, an AI feature that touches GxP records is a computerized system, and the regulatory question is whether you can show it is fit for its intended use and under control, not whether it is labeled AI.

A comparison: traditional CSV, CSA, and CSA adapted for AI

Aspect Traditional CSV CSA (CDRH 2025 logic) CSA adapted for AI
Starting point Validate the whole system Identify intended use and risk of each feature Define context of use and the human review step
Testing style Mostly scripted, with screenshots Scripted for high risk, unscripted or exploratory for lower risk Representative test sets with acceptance thresholds, plus adversarial and edge cases
Reuse of vendor evidence Often repeated by the user Encouraged where credible Useful, but model version and update behavior must be confirmed
Documentation Heavy, per-step Record what was done and the conclusion Record intended use, test data rationale, results, and reviewer role
Change handling Revalidate on release Risk-based reassessment Monitor performance continuously, trigger reassessment on model or data change
Typical failure Paper that proves little Under-documenting a high-risk feature Treating a probabilistic output as deterministic

How do you apply CSA to an AI tool step by step?

Here is a sequence that follows the CSA steps and adds what AI needs. A deviation summarizer serves as the running example.

Step 1: Write the intended use in one paragraph

Be specific about what the AI does, who uses it, and what decision or record it influences. "Summarizes deviation investigation narratives for a QA reviewer, who edits and approves before the record is finalized" is useful. "Supports quality operations" is not. The paragraph is the anchor for every later decision, and if you cannot write it clearly, the tool is not ready for GxP use.

Step 2: Classify the risk of failure

Ask what happens if the AI is wrong and nobody notices. Use your existing quality risk management process, which for most pharma firms means ICH Q9(R1), and keep the scale you already have. It helps to separate three questions: does the output directly create or alter a GxP record, does it drive a decision about product or patients, and is there an independent human check before it takes effect?

A rough guide:

  • Lower risk: drafting assistance where a qualified person rewrites and owns the final record, such as first-draft SOP language or meeting summaries not used as records.
  • Medium risk: classification or triage that a person confirms, such as suggesting a deviation category or routing a complaint.
  • Higher risk: anything that makes or effectively makes a quality decision, such as automated disposition, release recommendations people tend to accept, or data review that replaces a second-person check.

Higher-risk uses are where the draft Annex 22 posture bites hardest, and where I would want a static, locked model and substantial evidence.

Step 3: Choose assurance activities that fit

For lower-risk uses, exploratory testing by someone who knows the process, a review of vendor documentation, and a clear human-review requirement may be enough. For medium risk, add a defined test set drawn from your real documents, with acceptance criteria you set before running it. For higher risk, build a more formal protocol, with statistically reasoned sample sizes, edge cases, and a documented rationale for why the model's behavior is stable enough to rely on. As an illustration for a deviation summarizer, you might pre-define a critical error as a summary that omits or misstates a root cause, product impact, or batch number, and require zero critical errors in a sample of 59 representative deviations. Under the usual zero-failure (success-run) calculation, that supports at least 95% reliability at 95% confidence. Set the sample size and threshold from your own risk tolerance, and justify them in the file.

If you are using AI to help generate validation documents themselves, apply the same logic to that use. See validation and qualification protocol generation with AI QMS tools for more detail.

Step 4: Qualify the supplier and the model version

CSA allows leverage of vendor evidence, but with AI you need to ask different questions. What model is running? Can the vendor change it without telling you? Is your data used for training? Where are prompts and outputs stored, and who can read them? A general-purpose public chatbot generally cannot give you satisfactory answers, which is covered in why ChatGPT is not safe for quality management records. Put model versioning and change notification in your quality agreement.

Step 5: Define the human review and make it real

A reviewer who clicks approve on every AI output is not a control. If you rely on the human check to reduce risk, show that the reviewer has the qualification, the time, and the information to disagree. I would capture the reviewer's role in the intended use statement and in training records, and I would periodically sample reviewed outputs to see whether people are actually catching errors.

Step 6: Record the conclusion, not a mountain of evidence

CSA's documentation expectation is modest by design: the intended use, the risk determination, the assurance activities and results, any issues found and how they were resolved, and the conclusion that the software is fit for use. For AI, add the model identifier or version, the test data description, and the acceptance thresholds. Someone reading it a year from now should understand why you were comfortable.

Step 7: Monitor and reassess

This is the step most teams skip. Deterministic software can sit unchanged for years. An AI feature should have a defined monitoring plan: periodic sampling of outputs, tracking of error or override rates, and triggers for reassessment such as a vendor model update, a change in the document types it handles, or a rise in reviewer corrections. Route reassessments through your normal change control.

Worked example: the deviation summarizer end to end

This is an illustrative walk-through, not a record from a specific site. Intended use: the tool drafts a summary of a deviation investigation for a QA reviewer, who edits and approves it. Risk: medium-low, because a qualified person owns the final record. Assurance: a test set of 59 real, de-identified deviations, a pre-set acceptance criterion of zero critical errors, and exploratory testing by a QA investigator. Supplier: model version pinned in the quality agreement, with change notification. Review: reviewer role named in the intended use statement. Monitoring: monthly sampling of approved summaries and tracking of reviewer corrections, with reassessment through change control on any model update. The whole file fits in a few pages.

What evidence will an inspector want to see?

Inspectors do not need a specific format, but they will look for a line of reasoning that connects the pieces. In practice I would expect questions like these:

  • Where is the intended use written, and does the way people actually use the tool match it?
  • How did you decide the risk level, and who was involved?
  • What did you test, with what data, and what would have counted as a failure?
  • How do you know the model has not changed since you tested it?
  • What does the human reviewer do, and what evidence shows they do it?
  • How are audit trails, access, and data retention handled for prompts and outputs that become part of a record?

If you can answer those from a short file, you are in good shape. If the answers live in people's heads, fix that before the inspection does it for you.

Common mistakes

Treating CSA as permission to do less without saying why. CSA reduces low-value documentation. It does not reduce the need to justify your risk call. A thin file with no reasoning is worse than a heavy one.

Applying the same effort to every AI feature. A grammar-assist function and an automated record classifier do not deserve the same protocol. Proportionality works in both directions, and underinvesting in the high-risk feature is the more expensive error.

Validating once and forgetting the model can move. Version control for the model, not just the application, belongs in your change control process.

Ignoring data integrity basics. ALCOA+ principles still apply. If an AI-generated summary becomes part of a GxP record, you need to know who or what created it, when, and from what inputs.

Waiting for perfect regulatory clarity. Annex 22 is a draft and FDA's drug AI guidance is a draft, but the underlying obligations are not new. A documented, risk-based, human-supervised approach is defensible today and will adapt more easily than a rushed rebuild later.

Where does this leave a quality team starting out?

Begin with a low-risk, high-volume use where a qualified person already owns the output, build the assurance file as if an inspector will read it, and use the experience to shape your policy before you approach the harder cases. Writing the intended use and the monitoring plan will teach you more about your own process than any vendor demo.

Next step: pick one candidate AI tool, write its intended use in one paragraph, assign a risk tier, and have QA review both before any testing starts.

Last updated: 2026-11-01

Frequently Asked Questions

Does FDA's CSA guidance apply to pharmaceutical manufacturers?

The final CSA guidance (September 24, 2025) comes from CDRH and is aimed at medical device firms under the QMSR. Pharma firms are governed by 21 CFR Part 11, 21 CFR 211.68, and in the EU Annex 11. Those rules allow risk-based validation, so many pharma companies adopt CSA principles and cite the guidance as supporting reasoning, not as the legal basis.

Can you use generative AI in a GxP process?

It depends on the use. The July 2025 draft EU Annex 22 would exclude generative AI and large language models from critical GMP applications, and would allow them in non-critical uses only with a qualified human in the loop. Draft status means you should check current wording, but low-risk drafting assistance with documented human review is the most defensible starting point.

What is the main difference between CSV and CSA?

Traditional CSV tends to apply heavy, scripted testing and documentation across the whole system. CSA starts from intended use and risk, uses scripted testing where risk is high, allows unscripted or exploratory testing where it is lower, and records the conclusion rather than every screenshot.

How often should an AI tool be reassessed after validation?

There is no fixed interval in the regulations. Set a monitoring plan based on risk, with periodic output sampling and defined triggers such as a vendor model update, new document types, or rising reviewer corrections, and handle reassessments through change control.

What documentation should an AI CSA file contain?

Include the intended use, risk determination, assurance activities and results, issues and resolutions, the model identifier or version, a description of test data and acceptance thresholds, the human reviewer's role, and a conclusion that the tool is fit for its intended use.

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.