How do you validate an AI tool in a GxP environment?
By Savi C., Co-Founder · Published Nov 16, 2026
Start with intended use, not the model
Validation answers one question: is this system fit for the specific job you are using it for? An AI tool that drafts a deviation summary for a person to edit is a different validation problem from one that flags out-of-limit entries in a batch record, even if both run on similar technology. Write the intended use down first, in terms of the GxP process, the decision the output feeds, and who reviews that output before anything happens.
Annex 11 frames it the same way. Its principle says that where a computerised system replaces a manual operation, there should be no resulting decrease in product quality, process control or quality assurance, and no increase in the overall risk of the process. Section 1 asks for the extent of validation to be based on a justified and documented risk assessment. 21 CFR 11.10(a) asks for systems holding electronic records to be validated to ensure accuracy, reliability and consistent intended performance.
What each framework contributes
Framework
ISPE GAMP 5, second edition (2022)
What it gives you
The risk-based lifecycle for GxP computerised systems, critical thinking over documentation volume, and Appendix D11 on AI and machine learning
Framework
ISPE GAMP Guide: Artificial Intelligence (2025)
What it gives you
A dedicated guide for AI-enabled computerised systems, meant to be used alongside GAMP 5
Framework
EU GMP Annex 11 (2011, in force)
What it gives you
Expectations for risk management, suppliers, validation, audit trails, change control and periodic evaluation for medicinal products
Framework
FDA Computer Software Assurance guidance (February 2026)
What it gives you
A risk-based approach for software used in medical device production or the quality management system, which explicitly includes AI and machine learning tools
Framework
Draft EU Annex 11 revision and draft Annex 22
What it gives you
Direction of travel only. Both were consulted on in 2025 and are not in force
Reading the FDA CSA guidance correctly
FDA issued Computer Software Assurance for Production and Quality System Software as a final guidance on September 24, 2025, then replaced it on February 3, 2026 with Computer Software Assurance for Production and Quality Management System Software, renamed to match the QMSR. It was prepared by CDRH and CBER in consultation with CDER, and its scope is software used in medical device production or the quality management system under 21 CFR Part 820. Drug manufacturers working under Part 211 are not its direct audience, although many use its reasoning as a reference.
Two parts of it matter for AI. It says its risk framework can be applied to artificial intelligence and machine learning tools. And it treats software that measures, inspects or determines the acceptability of product or process with limited or no additional human review as generally high process risk. An AI tool whose every finding is reviewed by a qualified person sits in a different risk position from one that decides on its own, and your validation should say which one you have.
A validation package for an AI tool
Define
- Intended use and context of use, tied to a named GxP process and decision
- Risk assessment per function, including what happens if the output is wrong and who would catch it
- User requirements that are traceable to testing, as Annex 11 section 4.4 expects
- Supplier assessment and a written agreement setting out responsibilities, as Annex 11 section 3 expects
Test
- A test set of real, representative records, kept separate from anything used to configure or tune the tool
- Acceptance criteria set before testing, for both missed findings and false alarms
- Results compared with experienced human reviewers on the same records
- Behaviour on poor inputs: illegible scans, missing pages, unexpected formats. These should be reported as unverified, not passed
Operate
- Model, prompt, rule and configuration changes under change control, with regression testing against the same test set
- Performance monitored in routine use, including findings reviewers overturn
- Audit trail of what the tool flagged and what the reviewer decided
- Periodic evaluation that the system remains in a valid state, as Annex 11 section 11 requires
How Litewave fits
Litewave reads paper or electronic records, checks every entry against your master record and specifications, and shows QA only the exceptions, each linked to the page it came from. Litewave proposes, qualified people approve, and it never dispositions a batch on its own, which keeps a human review step inside every decision. It supports 21 CFR Part 11 electronic records, signatures and audit trails. Your validation stays yours: we supply the validation documentation you need, scoped with you during procurement, and you test the tool on your own records against your own reviewers.
Questions we get asked
Does FDA’s CSA guidance apply to drug manufacturers?
Not directly. It is a CDRH and CBER guidance for software used in medical device production and quality management systems under Part 820, prepared in consultation with CDER. Drug manufacturers still validate under Part 211 and Part 11, and many borrow its risk-based reasoning.
Can you validate a tool built on a large language model?
You validate the system for its intended use, whatever is inside it. That means a fixed, version-controlled configuration, testing on representative records, human review of output where the risk calls for it, and change control on every model update. Note that draft Annex 22, which is not in force, would keep generative AI out of critical GMP applications.
Is a supplier’s validation enough?
No. Supplier documentation and testing can reduce your effort, and both Annex 11 and the CSA guidance encourage using it, but the regulated company remains responsible for showing the tool is fit for its own intended use.
Sources
- EudraLex Volume 4, Annex 11, Computerised Systems
- FDA, Computer Software Assurance for Production and Quality Management System Software (February 2026)
- ISPE, GAMP 5: A Risk-Based Approach to Compliant GxP Computerized Systems, Second Edition (2022)
- ISPE GAMP Guide: Artificial Intelligence (July 2025)
- 21 CFR 11.10, Controls for closed systems (eCFR)
- European Commission and PIC/S consultation on draft Chapter 4, Annex 11 and Annex 22 (2025)
