Back to Notes

Regulated work · Life sciences

Validate the workflow, not the model: AI agents under GxP and Part 11

An AI model's answer changes from run to run. The workflow and the checks around it do not. For quality and regulatory teams, that is the part to validate.

Nils Widal · · 8 min read

Quality and regulatory teams in life sciences ask a fair question about AI agents. If the model can give a different answer each time, how can a system that uses it be validated? This note gives our answer. It is our approach, not an FDA position. Each company's quality team decides how it applies to their own procedures.

Some terms first

  • GxP is the group of "good practice" rules for regulated work, such as good manufacturing practice (GMP), good clinical practice (GCP) and good laboratory practice (GLP).
  • 21 CFR Part 11 is the FDA rule for electronic records and electronic signatures. It asks, among other things, for audit trails and for signatures that are tied to one named person.
  • CSV (computer system validation) is the traditional way to show that a system does what it should. CSA (computer software assurance) is the FDA's newer, risk-based approach, which puts the effort where the risk to patients and product quality is.
  • GAMP is the industry guide from ISPE for validating computerized systems. In 2025 ISPE published a GAMP guide on artificial intelligence.
  • An agent is an AI program that does a task in several steps. A workflow is the written list of steps it must follow. A check is a small program that tests a result and gives the same answer every time.

The problem with validating the model

Traditional validation expects the same input to give the same output. An AI model does not promise that. Its wording changes between runs, and its maker updates it. A validation report that depends on one exact answer would be out of date quickly.

The FDA's draft guidance from January 2025 on AI in drug and biological product submissions does not ask for that either. It asks sponsors to define the question, the context of use (what the model is used for, and how its output is used), the risk, and a plan to show the model is credible for that use. The level of effort follows the risk. That guidance is still a draft. ISPE's GAMP guide on AI also takes a risk-based approach.

The FDA's guidance on computer software assurance, now titled "Computer Software Assurance for Production and Quality Management System Software" (final in September 2025, updated in February 2026), says its risk-based approach can be applied to AI and machine learning tools and to automated workflows. So the question for a quality team is not "is the model validated", but "which parts of this process carry risk, and how do we assure them".

What does not change: the workflow and the checks

Around the model, a lot can be fixed. The steps of the job and their order can be written down and locked. The rules a result must meet can be written as checks. The person who must approve can be named. None of these change between runs.

Validate the workflow and its checks once. Check every AI draft against them on every run. A named person signs.

This changes what the AI is allowed to be in the process. It drafts. It does not decide.

  • The workflow decides which steps happen and in which order.
  • The checks decide whether a draft meets the written rules.
  • A person decides whether the draft is right, and signs.
  • The record shows every step, every check result and every signature.

A worked example: a deviation and its CAPA

A deviation is a departure from an approved procedure, such as a temperature out of range during a batch. A CAPA (corrective and preventive action) is the plan to fix the cause and stop it from happening again. Here is how the work could run.

  1. A quality specialist opens a deviation and attaches the evidence: batch record, alarm log, notes.
  2. An agent drafts the deviation report and a proposed CAPA from that evidence and the approved procedures.
  3. Fixed checks test the draft. Every required field is filled. Every procedure it cites exists and is the current approved version. Every statement about the batch points to a line in the evidence. The risk class is one of the allowed values. The due dates fall inside the limits your procedure sets.
  4. If a check fails, the draft goes back to the agent with the reason. After a set number of tries, it goes to the specialist with the list of what failed.
  5. When all checks pass, the specialist reviews the content: is the root cause right, is the action enough. The checks cannot judge that. The specialist edits and approves.
  6. The quality approver signs. The record keeps the draft, the check results, the edits and both approvals.

What gets validated here is the workflow (steps 1 to 6), each check script, and the signature step. The model is treated as a drafting tool whose output is always checked and always reviewed. The risk assessment then asks a narrower question: what happens if the draft is wrong and a person misses it? That is a question quality teams already know how to answer.

The limits, stated plainly

  • A check tests only what it can measure. It can test that a cited procedure exists. It cannot test that the root cause is correct. Scientific and medical judgment stays with people.
  • The checks must be out of the agent's reach. If the agent could edit a check, it could change the rule it is tested against. Checks must ship as a locked, versioned part of the workflow, outside the agent's working folder.
  • "Could not check" is not "passed". If a check cannot run, the work waits. It never passes by default.
  • A person can still miss something. Review must stay real. The record helps here: it shows how often people change drafts, and where.

Where we are

Widal.Factory runs today for engineering teams, on AWS. Its workflow from ticket to merge is fixed, tests run on every change, and a person merges. Workflows written as files, check scripts on any step, approval gates in the hub, and Part 11 signatures are being tested or planned. Quality is the first department we want to build with a design partner. If that could be your team, we would like to talk.

Read next: how we approach regulated work, and how it works.

Sources

  1. FDA, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (draft guidance, January 2025)
  2. FDA, Computer Software Assurance for Production and Quality Management System Software (February 2026)
  3. ISPE, GAMP Guide: Artificial Intelligence (2025)
  4. Anthropic, Building effective agents

Tell us which work you want to hand off first.

Book a call