The AI receptionist has collected a date, service, and callback number when calendar access fails. Asking the caller to start over wastes work; pretending the appointment was saved creates a different problem.
Treat access failure as a change in what the workflow can currently do, while preserving the information the caller already supplied.
Classify the failure before asking the caller for more
Missing permission, revoked access, expired authentication, provider outage, and invalid booking data require different responses. Repeating the customer's email will not repair a revoked integration.
| Failure class | Next owner |
|---|---|
| Caller-correctable field | Receptionist asks a specific question |
| Integration permission or access | Authorised account owner |
| Temporary provider failure | Integration recovery process |
| Unknown write outcome | Reconciliation owner |
Use the provider's documented error information and avoid exposing raw credentials or internal stack traces in the conversation.
Preserve intake without claiming a booking
Store the proposed appointment details as a pending request with an operation reference and named follow-up route. Keep it distinct from confirmed calendar events.
For a fictional service visit, the record should show the requested time, required service, usable contact information, last completed step, and the reason confirmation remains unavailable.
Tell the caller the same thing. If the business offers a callback, use an agreed promise that staff can meet rather than inventing a time to make the interaction feel complete.
Do not reconnect accounts through conversational guesswork
Restoring access belongs with an authorised administrator and the provider's approved process. The voice agent should not ask callers for account passwords or privileged tokens.
Monitoring can reduce how often callers discover the problem first. The credential-monitoring guide covers safe capability checks and ownership.
The trade-off is that a fallback request may require manual scheduling. That is preferable to a false confirmation, provided the office queue is visible and actively owned.
Test recovery after access returns
In a controlled environment, interrupt access during a fictional booking. Verify that the caller's data survives, no confirmation is spoken, and staff can see the pending request.
Restore access and reconcile whether a write occurred before retrying. A credential error reported after a multi-step flow does not establish that every earlier step failed.
The tool-timeout article explains why the outcome can remain uncertain even after connectivity is restored.
Calendar-access questions
Should the receptionist retry indefinitely?
Use bounded, failure-specific recovery with a clear caller fallback.
Can access restoration automatically confirm every pending request?
Recheck current availability and the customer's agreed request first.
What should staff receive?
Provide the pending details, failure category, and next owner when reviewing the workflow.