Navata
← All Library

Regulated AI

How to Define the Intended Use of AI in a GxP Workflow

AI-assisted research and drafting · Practitioner reviewed by Rohith Karanam Sreedhar · 4 September 2026

ⓘ Library content is researched and drafted with AI assistance and reviewed by a Navata practitioner before publication. For original analysis and long-form practitioner perspectives, visit Navata Insights →

“AI for deviations” does not define an intended use. It does not say whether the system retrieves records, summarises them, proposes a classification, drafts a conclusion, routes work or closes the record. Those contributions carry different evidence needs and consequences.

Describe the contribution precisely

A usable statement should answer:

  • who invokes or receives the output;
  • which approved sources and record states may be used;
  • the exact task performed;
  • the form and destination of the output;
  • whether it is informational, preparatory or decision-supporting;
  • what the human must examine and decide;
  • what the AI may write or execute in the regulated system;
  • the conditions, populations, languages and process variants covered;
  • known exclusions and prohibited uses;
  • what evidence must be retained.

For example: “Generate a first-pass summary of specified fields and attachments from an open deviation for an assigned investigator, with citations to the retrieved source content; the investigator must verify and may edit or reject the summary; the system may not classify, approve or close the record.” That statement creates testable boundaries.

Let intended use determine evidence

The FDA and EMA's January 2026 Good AI Practice principles support a clear context of use and risk-proportionate validation, mitigation and oversight. Within its medical-device production and quality-management-system scope, FDA's February 2026 Computer Software Assurance guidance likewise begins with intended use. It is not used here as general pharmaceutical GMP authority.

Validation can support claims only inside the defined conditions. Tests for grounded summarisation do not establish authority to assign root cause. Accuracy on one document population does not establish performance on another. A human approval step does not compensate for unclear retrieval scope or excessive execution permissions.

The European Commission's proposed Annex 22 is consultation draft material, not the currently effective EU GMP annex. Its scope is also narrower than generic discussion often suggests: the draft addresses specified static deterministic models in critical GMP applications and expressly treats generative AI separately. It can inform careful thinking, but claims about its status and applicability must remain qualified.

Connect the boundary to change control

Changes to the model, prompt, retrieval sources, metadata, workflow, permissions, user population or permitted reliance can move the system outside the evidence state. The intended-use record should name which changes trigger impact assessment and who owns that decision.

Sources