A receptionist that sounds convincing can still lose appointment requests. If you are hiring someone to repair an existing system, the first deliverable should be an explanation of where the workflow breaks, with evidence you can inspect.
I would scope a repair around one failing customer journey before discussing a replacement platform. Rebuilding everything can hide the original problem and leave you paying to rediscover your own business rules.
What should an AI receptionist repair include?
A repair should connect a symptom to a technical cause, change the responsible part, and demonstrate the resulting business record. “We improved the prompt” describes an edit. It does not establish that the receptionist now books correctly.
Give the engineer a redacted example: what the caller requested, what the receptionist said, and what the office actually received. Include the approximate time and any call or appointment reference. Do not share an entire customer database when one fictional reproduction will do.
For example, imagine a caller asks for Tuesday, corrects it to Thursday, and receives a Tuesday confirmation. The engineer needs to trace when the correction entered the conversation, which date reached the booking tool, and which event the calendar saved.
Ask for a small diagnosis before a broad rewrite
An initial review should produce a short failure map:
| Evidence | Question it answers |
|---|---|
| Conversation timeline | Did the receptionist capture the correction? |
| Validated tool arguments | Which date did the integration receive? |
| Calendar record | What was actually saved? |
| Retry history | Was the request submitted more than once? |
Ask the developer to distinguish a confirmed cause from a hypothesis. A plausible explanation still needs a reproduction. If access is incomplete, the report should say exactly which evidence is missing.
The repair proposal should name the affected workflow, required account access, test cases, rollback method, and exclusions. A broken appointment workflow does not automatically justify changing the voice, website, or CRM.
Evaluate the repair against saved outcomes
Use fictional calls for an ordinary booking, a corrected date, an unavailable slot, and an integration timeout. Inspect the resulting records yourself. For the timeout case, ask what prevents a second booking when the first request might already have succeeded.
The handover should include enough information for another engineer to maintain the change: where the configuration lives, who owns the accounts, which tests to repeat, and how staff find unresolved requests.
My voice AI engineering work covers integrations, debugging, and handover. The Voho case study describes my contribution to an existing platform rather than claiming that I built its entire business. Apply the same standard when reading any developer's portfolio.
Questions to ask an AI receptionist developer
Do we need a complete rebuild?
Not necessarily. Request evidence that the current architecture cannot support the required fix before approving one.
Should the developer guarantee zero failures?
No. Ask for defined acceptance cases, known limitations, monitoring, and an owner for incidents.
What should I prepare before the first call?
Bring one redacted failing journey and the record you expected. The workflow-testing guide provides a useful review format.
Start by scoping that one repair, with diagnosis and implementation listed as separate deliverables.