Computer System Validation (CSV) for AI Workflows: What It Means in Practice

Computer system validation for AI is a validated state you maintain: Annex 11, draft Annex 22, GAMP 5, IQ/OQ/PQ and the drift triggers for re-validation.

Computer System Validation (CSV) for AI Workflows: What It Means in Practice
Written by TechnoLynx Published on 12 Jun 2026

Computer system validation (CSV) is the documented process of showing that a regulated computerised system consistently does what it is meant to do, and does not do what it is not meant to do, in the environment where it is actually used. In EU GMP the proof is not a one-off. Systems are expected to be validated before use and kept in a validated state for their whole lifecycle. For an AI workflow, that second requirement is the hard part, because a model can start behaving differently when production data moves away from its training data, even if nobody changes a line of code.

In practice, then, computer system validation for AI means a validated state you keep up. Intended use and acceptance criteria are fixed before testing, and change control and performance monitoring continue in operation. For the wider context of where AI and generative models are used across the industry, our overview of generative AI in pharma: compliance and innovation covers the regulatory landscape beyond validation.

How does computer system validation work, and what does it actually do in a pharma or life-sciences environment?

CSV produces the evidence that a computerised system is fit for its intended use, so the organisation can rely on it for GMP decisions and answer an auditor’s question, “how do you know this system works?”. The process starts with validation planning, system requirements, testing and verification, and a validation report. The system then goes into operation, and the work continues until the system is retired and its data kept as the regulations require.

The object being validated is wider than the software. EU guidance defines a computerised system as a function carried out by trained personnel and controlled by a computer system. The hardware and software do the controlling, and the equipment plus the operating procedures make up what is controlled. For AI, change control covers the model, the system it runs in, the process it supports and its physical inputs.

Requirements anchor everything. Documented system requirements are what state the intended use of a system, and user requirement specifications set down in writing what the user expects the system to do. The EU glossary separates verification from validation. Verification is confirming, with objective evidence, that specified requirements have been met. Validation is showing that the system is fit for its purpose.

Having a capability is not the same as being validated. A system can be “Part 11 capable”, with an audit trail, role-based access and electronic signatures built in by the vendor, and still be unvalidated until someone proves those controls work as configured in this specific deployment.

How does EU GMP Annex 11 relate to computer system validation for systems that include AI?

Annex 11 is the EU GMP annex for computerised systems, and it applies to every computerised system used in GMP-regulated activities. Its core expectations apply to AI tools in full:

  • the application is validated and the IT infrastructure qualified;
  • a system that replaces a manual operation must not lower product quality, process control or quality assurance;
  • risk management applies across the lifecycle and decides how far validation goes;
  • systems are evaluated periodically to confirm they are still valid.

The EU is revising Annex 11 and Chapter 4 and introducing a new Annex 22 on artificial intelligence. The revised Annex 11 strengthens lifecycle management, asks for Quality Risk Management at every step, and tightens the definition and upkeep of system requirements. It also says requirements must be kept up to date as the system changes, and that those updated requirements are the basis for qualification.

Annex 22 adds to Annex 11. It does not replace it. An AI tool is still a computerised system, so Annex 11 applies in full, and Annex 22 adds what is specific to the model: its intended use, its testing and how its performance is controlled. Its scope is narrow on purpose. It covers AI models in critical applications that directly affect patient safety, product quality or data integrity, such as models that predict or classify data. Within that scope, the draft covers only static models that do not adapt during use. Dynamic, continuously learning models are outside the document and should not be used in critical GMP applications. Generative AI and large language models are also explicitly outside scope for critical applications.

Annex 22 is still a draft, and the July 2025 consultation text is deliberately restrictive about adaptive, non-repeatable and generative models. The draft status does not leave a gap in the meantime. Current Annex 11, the pharmaceutical quality system and existing change-control and data-integrity requirements already apply to AI-assisted work.

How do IQ/OQ/PQ protocols work in CSV, and how do they map onto an AI/ML system?

IQ, OQ and PQ are qualification stages, each producing its own documented evidence. IQ shows the system was installed correctly against a specified version and configuration baseline. OQ shows each function works across its operating range, including boundary and error conditions. PQ shows the system performs correctly under real conditions of use. Each stage runs against an approved protocol that sets out, before testing starts, what will be tested and what counts as a pass.

What each one has to prove changes.

Stage Classical scope What it has to cover for an AI/ML component
IQ Installed components match the specified version and configuration baseline The model artefact and version, its configuration and input pre-processing, and the inference runtime, all recorded as the baseline that change control protects
OQ Functions behave correctly in isolation, including at boundaries and on errors Results below the confidence threshold are flagged, sent for manual review and kept out of critical decisions
PQ The system works under real workflows, users and data volumes Performance against acceptance criteria fixed before testing, on test data kept independent of training data and representative of the site’s real inputs, including rare cases
After PQ Periodic review on a defined schedule Ongoing monitoring of the validated metrics, and of whether input data stays within the validated sample space

Why CSV for an AI workflow differs from CSV for a deterministic system

A deterministic clinical or manufacturing system does the same thing with the same inputs every time, so its qualification record stays true until someone changes the system. An AI model breaks that assumption. Its performance can change without any change to the code, purely because production data drifts away from the training data. Risk also depends on context: the same architecture can be low risk in a document classifier and high risk when it informs batch disposition.

The draft Annex 22 requires critical models to be static and deterministic, so that the same input always produces the same output. ISPE describes a related pattern in which AI is used during method development, and the parameters are then locked so that what enters validation is a fixed, deterministic, reviewable method.

Determinism helps repeatability, but reconstructing a result later depends on keeping the input, the output, the model version, the configuration and the audit-trail evidence.

What fills the gaps the classical model leaves

  1. Acceptance criteria are set before testing. A process subject matter expert defines them, and they are documented and approved before acceptance testing begins. They must be at least as high as the performance of the process the model replaces, which means the current manual or rule-based process has to be measured first.
  2. Test data is independent. Data used to test a model must not have been used in development, training or validation. It is either captured after training or split off before training starts, and developers should not have access to it.
  3. Confidence thresholds are part of the acceptance criteria. Each threshold is justified during validation, matched to the intended use, and documented with the acceptance criteria.
  4. Human review is measured as a control. A reviewer who never disagrees with the model is not really controlling it. Override rates and seeded-error checks show whether review has turned into a rubber stamp.

What does a risk-based CSV approach scope in versus scope out, and how is intended use defined?

A risk-based approach sets validation strategy and effort according to the system’s intended use and the risks it poses to product quality, patient safety and data integrity. Manufacturers should be able to justify their protocols, acceptance criteria and records from their risk assessment. The practical result is depth where failure matters. A vendor’s standard, unmodified date-stamping routine needs lighter verification than a custom calculation that decides batch release.

For an AI model, intended use has to be written down before testing starts. It covers what the model does, which process step it supports, and what its input data looks like, including expected variation, limitations and edge cases. Whether an application counts as “critical” depends on its intended use and decision pathway, not on the department it sits in or what the software is called. A tablet-grading vision system that decides accept or reject is the kind of application Annex 22 targets.

Scoping cuts both ways. Our position is that the first question to ask about a proposed AI application is whether it falls under regulatory requirements at all, because many parts of a life-sciences pipeline sit outside the regulated portion even when the organisation assumes everything is constrained. A model that drafts internal summaries and a model that classifies deviations warrant different validation effort. How much testing a component needs also depends on its GAMP software category, which our guide to GAMP software categories explains.

How does GAMP 5 apply when validating an AI/ML workflow, and what does the FDA’s current guidance expect?

GAMP 5, published by ISPE, is the risk-based framework that pharmaceutical, biotech, device and CRO validation programmes refer to. Its central argument is that validating every function to the same exhaustive depth wastes effort. For AI, Appendix D11 of GAMP 5 Second Edition covers static and dynamic machine-learning subsystems. It sets out an AI/ML lifecycle aligned with the GAMP phases of concept, project and operation, and it stresses data integrity and the iterative way models are developed. The operation phase includes deployment and continuous monitoring. ISPE maps the requirement to describe changes through the life cycle onto GAMP’s operational change and configuration management and periodic review activities. ISPE has also published a separate GAMP Guide on artificial intelligence alongside GAMP 5. For the framework itself, see our GAMP guide for validation of automated systems.

21 CFR Part 11 sets the controls that make electronic records and signatures trustworthy, and CSV is how a site proves those controls work in its own configuration. A 2026 FDA warning letter shows that US authorities can enforce existing CGMP requirements against AI-assisted work that is not adequately controlled. Engineering teams working through the FDA general principles of software validation will find them covered there, and the choice between computer software assurance and full CSV has its own page on when to use CSA vs full CSV for AI systems.

Does mapping CSV to ISPE GAMP make the validation evidence stronger?

Yes, in a specific way. Annex 11 asks manufacturers to justify their standards, protocols and acceptance criteria from a risk assessment. Tying each validation decision to a named, widely referenced framework turns that justification into something an inspector can check against a structure they already know, instead of a bespoke argument they have to reconstruct. ISPE notes that GAMP-based quality activities rest on good engineering practice and can serve as the basis for meeting the draft EU AI regulation’s requirements for high-risk AI systems, even though those are not assessed as GxP systems. The framework does not do the work for you. It gives the evidence a shape that others recognise.

Re-validation triggers for computer system validation of AI models

Under Annex 11, any change to a computerised system goes through a controlled procedure. The Annex 11 consultation draft adds that a significant change that may affect product quality, patient safety or data integrity should be re-qualified and re-validated. Annex 22 goes further for AI. The model, the system it runs in and the process it supports all go under change control before deployment, and that includes physical inputs such as cameras or sensors. Any change is evaluated to decide whether the model needs retesting, and a decision not to retest must be fully justified.

A common misconception is that any change means full re-validation. ISPE’s position is that it does not, provided the response to drift was planned before deployment. ISPE tracks three kinds of drift. Data drift is a shift in the statistics of the inputs. Concept drift is a change in the relationship between inputs and outputs. Performance drift is a measurable fall in effectiveness. Each needs thresholds, a review frequency and an escalation route.

Event Expected response
Model retrained or updated Treat it as a new model: a controlled change, retested before use
Change to platform, infrastructure, process or physical input device Evaluate the impact on model performance. Re-validate partially or fully where relevant, and document any decision not to retest
Minor drift within acceptable limits Log it and keep monitoring
Moderate drift with contained impact Run targeted tests against the affected criteria
Major drift, architecture change or pipeline modification Re-validate the affected scope, or pause the system until resolved
Input data leaving the validated sample space Take corrective action: retrain, re-validate, or restrict use until alignment is restored
Scheduled periodic review Confirm the system is still fit for intended use, at a frequency justified by risk

This is how a model update becomes a scoped delta rather than a full re-qualification. The difference lies in defining, before go-live, which tests each kind of change re-runs. Periodic review also has to check the combined effect of several small changes, including changes in other connected systems, and to find undocumented changes, for example through configuration auditing. Decisions not to re-validate need a trace just as much as decisions to re-validate.

Where CSV evidence lands in the GxP evidence pack

CSV is the work that produces the validation-evidence section of a GxP evidence pack. That section includes the signed validation summary report recording what was tested, what passed, deviations, and the fit-for-use conclusion. The operating evidence sits next to it: change-control records for each post-validation modification, assessed for re-validation impact before the change is applied, and drift assessments, approvals and test results kept in an auditable trail. An auditor reads the audit trail first, because it decides whether the rest of the pack is credible.

How these documents fit together is covered in how the evidence pack relates to validation evidence, and the pack as a whole is described in the HIPAA / GxP workflow evidence pack. Reconstructing a result depends on keeping inputs, outputs, model versions and audit records, so the store holding them matters too: see Cassandra DB performance for validated AI workflows.

Who can engineer the validated-state controls for an AI workflow?

TechnoLynx builds the engineering side of a GxP-ready AI workflow. TechnoLynx describes this work under its AI engineering for life sciences and healthcare page.

The validation decision itself, which engineering cannot sign off, belongs to the regulated company and its quality function.

What stays open while Annex 22 is still a draft

The final Annex 22 text may change the scope boundaries, especially how tightly adaptive and generative models are excluded from critical use. The obligation underneath already applies: the system must stay in a valid state and every change must follow a defined procedure. We would ask whether a current AI deployment could show today, under Annex 11 alone, what has changed since qualification and why each change did or did not trigger a retest.

Back See Blogs
arrow icon