Run a controlled acceptance test and define evidence-based launch gates.
A polished demo is one path under ideal conditions. Launch evidence comes from expected calls, edge cases, real integrations, degraded conditions, and a team that can see and repair failures.
Prepare the test environment
Use test phone numbers, non-sensitive records, sandbox integrations, and named testers. Freeze the configuration under test. Record platform version, prompt or behavior version, knowledge snapshot, voice, language, routing, integrations, hours, and test date so failures can be reproduced.
Knowledge checks
- Answers current opening hours by location.
- Handles a one-time closure.
- Distinguishes offered from unsupported services.
- States price only where an approved source permits it.
- Does not invent availability.
- Explains uncertainty and offers the right next step.
- Uses the correct source after two pages conflict.
- Reflects an approved knowledge update.
Conversation checks
- Handles a caller who gives the reason before the greeting ends.
- Recovers from interruption without restarting.
- Asks one clarifying question for an ambiguous request.
- Understands spelled names and confirms them.
- Confirms numbers, email, address, and date separately.
- Handles silence and background noise.
- Lets a caller correct an earlier answer.
- Ends cleanly when the caller chooses not to continue.
Action and integration checks
- Creates one lead—not duplicates—after a retry.
- Maps fields to the correct record.
- Applies time zone and business hours correctly.
- Requests, rather than promises, an appointment when confirmation is unavailable.
- Handles an integration timeout.
- Shows the failed action to staff.
- Queues follow-up with an owner and due state.
- Preserves a usable correlation or call identifier.
Transfer and escalation checks
- Transfers during staffed hours.
- Explains what happens before connecting.
- Recovers when the destination does not answer.
- Uses the approved after-hours path.
- Recognizes defined safety or urgent language.
- Does not exaggerate response time.
- Routes sensitive or disputed matters away from automation.
- Provides an accessible non-voice alternative where configured.
Privacy, security, and operations checks
- Gives the approved recording or automation notice.
- Avoids prohibited sensitive fields.
- Restricts records by staff role.
- Applies retention and deletion controls.
- Redacts protected values in logs and exports.
- Shows usage, provider, and delivery failures.
- Supports rapid disablement or fallback routing.
- Produces a review packet without exposing unrelated calls.
Score outcomes with severity
| Severity | Example | Launch response |
|---|---|---|
| Stop | Unsafe advice, privacy breach, wrong high-impact action, no fallback | Do not launch; correct and rerun affected suites |
| High | Incorrect core fact, lost lead, false booking, failed urgent route | Fix before live traffic |
| Medium | Recoverable misunderstanding or staff-visible integration failure | Correct or accept with documented mitigation and owner |
| Low | Minor wording or pacing issue with correct outcome | Backlog with review date |
Set thresholds before testing. Averages can hide a single unsafe failure, so stop-category cases must pass individually.
Pilot and monitor
Route a small, reversible segment. Review outcome samples daily at first, track error and escalation patterns, and compare records with actual staff follow-up. Platforms such as Receptionist Max can be assessed against these evidence needs; confirm the live capability and configuration you are actually buying.
This test works best after the call outcomes and boundaries are documented. Use the coverage estimator to size a pilot from your own call assumptions.
Sources and further reading
Primary and contextual sources used to verify definitions or give readers a relevant next resource.
- Receptionist Max A current AI-receptionist platform example readers can use to compare visible knowledge review, call handling, escalation, and operational controls against this independent test plan.