An ambitious demonstration, not a deployable medical receptionist.
The snapshot demonstrates triage, tags, pipeline movement, intake, calendar booking and post-call generation on one canvas. However, both calendars are unbound, the Call Ended flow is manual, one emergency stage does not exist, patient data is loosely mapped and a transcript-derived summary may be emailed to the patient.
Platform breadth
Classification, Ask & Save, pipeline state, tags, calendar nodes, LLM summaries and email.
Booking promise
Two Smart Calendar nodes have no calendar and their outcomes are not handled.
Medical data
Summary delivery and activity-only fields need purpose, consent, retention and access review.
One agent, two workflows and a five-stage healthcare pipeline.
GPT-5.4 Mini, Janet by ElevenLabs, 304-word prompt and zero active channels.
Triage, emergency routing, symptom classification, intake and two booking nodes.
Manual trigger → GPT-4.1 Mini summary → email to contact.
New Request, Provider Review, In Progress, Awaiting Patient and Completed. Emergency is referenced but absent.
The graph references six unique values although the library manifest reports five.
Two unbound calendars and a literal “Add Phone Number” transfer placeholder.
Two emergency implementations and no complete appointment terminal.
Workflows
CONVERSATIONCALL ENDEDInstalled CRM
Health Care pipelineTagsLife-threatening route
Chest-pain route
Important values are captured without durable structure.
The entire response is written into one name field.
No defined field, access purpose or retention rule.
Not mapped into a governed callback field.
Clinical context lacks structured destinations and governance.
Transcript-derived medical content may be sent through an unapproved channel.
Nine findings across automation, safety and data integrity.
Post-call automation will not run automatically.
The primary appointment promise cannot complete.
An emergency pipeline write targets a state that is not installed.
Urgent callers cannot reach staff as promised.
Transcript-derived health information may be emailed to an unintended or insecure destination.
Name and intake values are not stored in reliable structured fields.
The manifest says five tags while the graph references six values.
Similar symptoms produce different states and terminal behavior.
Inconsistent labels reduce reporting quality.
Resolve patient safety and data governance first.
- 1Consolidate emergency routingSafety
Use one clinically reviewed rule, one valid pipeline state and an explicit staff/911 fallback.
- 2Bind and test calendarsRequired
Configure one approved timezone and close success, conflict and error paths.
- 3Install real post-call automationRequired
Replace the Manual trigger with the correct event and validate its payload.
- 4Govern patient dataRequired
Minimize collection, map approved fields and define access, delivery, retention and consent.
- 5Test synthetic pathsTest
Exercise emergency, urgent, routine, decline, success and calendar failure without real patient data.
E-Health audit complete.
The workflow graph, patient-data contract, dependencies and defects are documented.
Return to the Snapshot Library →