An AI receptionist should not begin with a list of features. It should begin with a list of permitted actions. A clinic needs to know which calls or messages the system may handle, which decisions remain human and what immediately ends the pilot.
Define the narrow first job
Start with one bounded intent such as office hours, location, appointment confirmation or transfer to a named queue. Do not combine booking, triage, prescription requests, clinical advice and billing questions in the first test. Every added intent creates another path for misunderstanding or failed follow-up.
Write the boundary in plain language
- What information may the system collect?
- What may it read back or confirm?
- Can it create, change or cancel an appointment?
- Which requests always require a person?
- Which urgent phrases trigger immediate emergency language and human escalation?
- What happens when confidence is low, the caller changes language or the integration fails?
Map the information before the demonstration
List recordings, transcripts, summaries, logs, prompts, identifiers and analytics. Identify every supplier and subprocessor, where each item is stored, how long it remains and how it is deleted. Article 106 of Quebec’s health-information law addresses privacy impact assessments for certain technology acquisitions, developments and redesigns involving held health information. The clinic should involve its privacy lead early and obtain qualified advice for its specific project.
Test the handoff, not only the voice
Run French and English scenarios with noise, accents, incomplete names, repeated callers, accessibility needs and unavailable staff. Confirm that the receiving person can see why the call was escalated and what the system already told the caller. A successful transfer that loses the request is still a failed workflow.
Set stop rules before launch
Pause the pilot for a privacy incident, an unsafe response, a repeated booking error, an untraceable action, a failed urgent escalation or a material difference between the approved configuration and the live system. Assign who can stop it, who investigates and what evidence is required to restart.
Use the AI receptionist evaluation guide to structure the comparison. For a workflow-first review, see the Clinical Operations Assessment.
Official sources reviewed
Evidence reviewed: August 28, 2026. This framework is practical information and does not replace privacy, security, legal or professional advice.
