A healthcare scheduling agent needs a state machine, not a promise.

Availability is not a confirmed booking

FHIR R4 describes an Appointment as a scheduled healthcare event and separates it from the clinical Encounter. Its workflow includes proposed appointments and participant responses. The documentation explicitly notes that an available Slot does not guarantee a booking: additional resources, permissions or qualifying decisions may still be required. R4 is a specific published version; integrations must match the version and profiles used by the organisation.[1]

Keep the first workflow administrative

Consider an outpatient scheduling team receiving incomplete appointment requests. A bounded assistant could check administrative fields, retrieve permitted scheduling records, identify missing information and suggest available times for staff review. It should not infer urgency, diagnose symptoms or change the clinical priority assigned by qualified staff.

Begin with synthetic records and an agreed list of allowed fields. Make patient identity matching, accessibility needs, time zones, required resources and the authority to contact a patient explicit design questions. A production pilot needs the organisation’s privacy, security and clinical governance review, not just an API connection.

Model transitions explicitly

Represent a proposed appointment, participant responses and confirmed booking as different states. Read the authoritative scheduling system immediately before attempting an approved change. If the slot has gone, show alternatives; do not describe the old proposal as booked.

Use narrow tools for retrieving schedules and preparing a proposal. Keep booking writes behind the existing permissions and staff approval. Record the requested change and the scheduling response so that a failed or duplicated operation can be reconciled. Patient-facing confirmations should be produced only from a confirmed authoritative result.

Test the exceptions that create extra work

Test unavailable practitioners, room conflicts, ambiguous names, changed slots, duplicate requests and cancellations. Measure proposal correctness, identity mismatches, accidental writes and staff rework. Compare administrative turnaround with the existing process, while recording any new work the assistant creates.

An appealing conversation is not the success criterion. The first useful milestone is a staff-reviewed proposal with correct resource references and no false booking claims. This blueprint demonstrates an administrative design direction; it establishes neither medical effectiveness nor deployment compliance.