An Irish hospital group wants Claude to turn discharge notes into summary letters for family doctors. Discovery has captured the stakeholders, the current process (a junior doctor spends about 25 minutes per letter) and a success criterion of a reviewed letter within an hour of discharge. The architect must now choose between calling a provider-hosted model endpoint and deploying through the group's existing cloud tenancy, and the information governance team has not yet been consulted. Which question must discovery answer before that choice is made?
- AWhether patient-identifiable notes may be processed outside the group's own boundary, and under which data processing agreement. Correct
- BHow many discharge letters each ward produces on a typical day, so the team can size throughput for the hosted endpoint.
- CWhich layout and tone family doctors prefer in a discharge letter, so the prompt template can match the group's house format.
- DWhether clinicians would accept a confidence score shown on each letter, so low-scoring letters can be routed for closer review.
Why A is correct: Correct. Where patient-identifiable data may be processed, and under what agreement, is a compliance constraint that rules deployment paths in or out. Until information governance answers it, choosing between a provider-hosted endpoint and the group's own tenancy is designing before a binding requirement is known.
Why B is wrong: This is tempting because volume is a standard discovery item and affects capacity planning. It is wrong because either deployment path can be sized for a hospital's daily discharge volume, so the answer does not separate the two options; it assumes the hosted path is already permitted.
Why C is wrong: This is tempting because the doctors receiving the letters are key stakeholders and their acceptance drives adoption. It is wrong because letter format is settled in the prompt and evaluation work on either path; it has no bearing on where patient data may be processed.
Why D is wrong: This is tempting because it looks like a sensible way to target clinical review. It is wrong because it does not bear on the deployment decision, and a model's self-reported confidence is not a reliable accuracy signal to route clinical review on.