Validating an AI assistant that answers from your SOPs

Is it GxP at all?

It depends on how the answers are used. If the assistant only helps people find the right document, which they then read and follow, it is a search aid, and the controlled document remains the instruction. If people act on the assistant's answer without opening the document, the answer has effectively become the instruction, and it carries GxP weight.

Be honest about which one you will have in practice, not which one the project plan says. Busy people will act on a confident answer.

The failures to design against

• Answers from superseded versions. The assistant draws on an old SOP that has since been revised.
• Answers from the wrong scope. A procedure for one site or product applied to another.
• Confident gaps. No document covers the question, and the assistant produces a plausible answer anyway.
• Lost conditions. The procedure says do this unless that, and the summary drops the unless.

Every one of these is invisible to the user, which is what makes them serious.

Controls that work

Every answer should cite the specific document, version and section it came from, with a link. Only effective versions of controlled documents should be available to it, linked to your document management system so that superseded versions drop out automatically. When the documents do not answer the question, the assistant should say so plainly rather than fill the gap.

And the user interface should make clear that the controlled document is the authority. A small habit, such as always showing the source passage next to the answer, does more than any disclaimer.

How to test it

Build a set of real questions with known correct answers, drawn from the people who will use it, including questions the documents do not answer and questions where site or product matters. Agree acceptance criteria in advance: what share must be correct, what share must cite the right source, and how gaps must be handled. Our pieces on building an AI test set and AI acceptance criteria cover the method.

Keep the scope narrow at first

It is tempting to load every controlled document into the assistant on day one. Starting with one area, such as warehouse procedures or one laboratory's methods, makes testing manageable and lets you learn how people actually use it.

It also makes failures easier to see. If the assistant gives a wrong answer about one area, the people who know that area will notice. Spread across the whole company from the start, the same mistake can sit unnoticed for months.

After go-live

Document changes are constant, so re-run the test set on a schedule and after significant document updates. Give users a simple way to flag a wrong answer, and review those flags. A sudden fall in flags is as interesting as a rise. Our GxP AI validation page covers the wider lifecycle.