Clinician letter summarisation with human sign-off

This is a pattern we would propose, not an engagement we have delivered. It is set out in full because in clinical settings the governance design is the project.

The constraints that shape this pattern

Two constraints drive every decision below: patient data does not leave the trust tenant, and nothing generated reaches the patient record without a clinician actively approving it. Both are correct, and both are design inputs rather than obstacles to be negotiated down.

How we would build it

Azure OpenAI deployed with private endpoints inside the trust tenancy, in UK South, generating structured summaries of clinical correspondence. Every summary lands in a review queue. Nothing is written back to the record without an explicit clinician action, and the audit trail records who approved what and when.

What information governance will ask

We would expect to answer all of the following in writing before build starts, because an IG panel will ask them and a project that cannot answer them does not proceed.

• Where the data sits at rest and in transit, and who can reach it
• Whether abuse monitoring and human review are opted out at the Azure OpenAI service level
• What the DPIA says, and who signed it
• What the clinician sees when the model is uncertain
• How an incorrect summary is detected, reported and corrected

Where this gets hard

Adoption, not accuracy. A summarisation tool clinicians do not trust simply goes unused, and trust is earned by watching how the work is actually done rather than by improving a benchmark. We would expect to spend the first fortnight on the ward before writing code, and we would treat that as delivery time, not discovery overhead.