"Agentic" is the word of the year in quality software, and most of the time it means a chat window attached to an old system. That is a pity, because there is a useful idea underneath it, and quality teams are the people most likely to benefit from it.
An agentic QMS is one where records move the work forward instead of waiting for a person to remember the next step. The AI helps with the writing. The system handles the hand-offs. Qualified people still make every decision that matters.
Where quality time actually goes
Ask a QA manager where the week went and the answer is rarely "making decisions". It is writing up events that are already understood, chasing actions that were agreed in a meeting, and checking that the SOP change promised in a CAPA was actually started.
Each record lives in the system. The links between them live in people's heads, in email and in spreadsheets. That is where things fall through: a CAPA closed on a promise, an SOP still effective months after an investigation said it was wrong, a change approved without anyone asking what it did to the validated state.
What agentic should mean
Three things, and none of them is a chatbot.
- Records that open the next record. A committed deviation opens its investigation. A concluded investigation raises the CAPA. A committed CAPA raises the change control and the SOP revision. Each step still has its own review and signature.
- Drafting from the facts already captured. The description of an event, the summary of an investigation, the CAPA plan, the redlined SOP. These are writing tasks, and the facts are already in the record.
- Controls the system enforces. A major deviation cannot close without a concluded investigation. Someone whose training has lapsed cannot sign for the work. A failing result cannot be averaged into specification.
What it must not do
It must not classify severity, decide root cause, approve a CAPA or release a batch. Those are judgements a qualified person makes and signs, and an inspector will ask who made them.
The draft EU GMP Annex 22 points the same way: adaptive and generative models should stay out of GMP-critical decisions. Annex 22 is still a consultation draft, but the direction is clear enough to design around now rather than retrofit later.
How to tell the difference in a demo
Ask the vendor to take one deviation all the way through: intake, investigation, CAPA, change, SOP revision and retraining. Watch how many times someone has to re-enter information or open a new record by hand. Then ask what happens to the SOP revision if the change is rejected. If the answer involves a person remembering, the system is not doing the connecting.
Finally, switch the AI off and run the same deviation again. Every step should still work. AI that the workflow depends on is a risk; AI that saves time on a workflow that stands on its own is a benefit.
This is how we built Cybroscape eQMS: the chain from deviations and investigations to CAPA and change control is one connected flow, with Quality Copilot drafting and people deciding.
Where to go next
Explore GxP Copilot for AI-native validation, TraceDraft for source-traceable clinical documentation, or book a demo to see either on your own data.