AI in medical devices · Expert · 13 min lesson

Using AI tools safely in QA/RA work: validation of tools, confidentiality and records

ISO 13485:2016 clauses 4.1.6, 4.2.4, 4.2.5, 6.2 and 7.4, FDA QMSR (21 CFR Part 820) and CSA guidance, EU AI Act Article 4, GDPR Articles 5 and 28

What you will learn

Decide when an AI tool used in quality work falls under clause 4.1.6, and plan a risk-proportionate validation with a defined intended use, acceptance tests and review control.
Apply confidentiality rules to what may and may not be entered into an AI tool, including personal data, client information and unpublished regulatory material.
Keep records that show what the AI did, who reviewed it and what was approved, and answer the questions an auditor asks about AI use.

Key takeaways

Clause 4.1.6 applies to AI tools used in the quality system: validate before use and after changes, in proportion to risk, with records.
Write the intended use, list failure modes such as invention, omission, leak and silent vendor change, and test with representative inputs and criteria fixed in advance.
Human review against the source is the main control for drafting uses, and a named person, never the tool, approves the record.
Personal data, client information and unpublished data are red and stay out of general tools (GDPR Articles 5 and 28).
Keep a short record of tool, version, reviewer, corrections and approver, and define revalidation triggers for vendor changes.

Free practice question

Calder's regulatory team starts using a hosted AI assistant to draft sections of CAPA records. The IT manager says the assistant is a well-known product, so no validation is needed.

What does clause 4.1.6 of ISO 13485 imply?

No validation is needed because the product is commercially available.

Not quite. Clause 4.1.6 makes no exception for commercial or well known software used in the quality management system.

Only the vendor must validate it, and Calder only files the vendor brochure.

Not quite. Clause 4.1.6 puts the validation of the application on the organisation using the software, not only on the vendor.

Validation is needed only if the tool produces a final signed record.

Not quite. Clause 4.1.6 depends on use within the quality system and risk, not only on whether a signature is produced.

The software is used in the quality system, so its application must be validated in proportion to risk, with records.

Correct. Clause 4.1.6 applies to software used in the quality management system regardless of how popular or well known the product is.

Key terms

intended use: A written statement of the task, inputs, outputs and users for which a tool is validated; changes to it can require revalidation under clause 4.1.6.
computer software assurance (CSA): The FDA's risk-based approach to establishing confidence in production and quality system software, set out in final guidance of September 2025.
hallucination (confabulation): A fluent output from a generative model that contains invented facts, sources or clauses.
red data: Information that must not be entered into a general AI tool, such as personal data, client-confidential and unpublished data.
revalidation trigger: A predefined event, such as a vendor model change or a failed periodic check, that requires the validation to be repeated in part or in full.
AI literacy: The skills and knowledge staff need to use AI systems with awareness of their opportunities and risks, required to be built by providers and deployers under AI Act Article 4.

Official sources

Go further

The full lesson has narrated slides, more scenario questions with explained answers and a simulation.

Open the Academy Watch on YouTube

Illustrative example. Educational only, not legal advice.