An integration can remain configured in the dashboard while its credentials no longer work. The next caller becomes the first person to discover the problem, often halfway through booking.
Detect credential failures with a small operational check before relying on the integration in a live conversation.
Monitor usable access, not the presence of a secret
A stored token or key does not prove the application can perform the required operation. Access may expire, be revoked, lose scope, or point to the wrong environment.
Where supported, inspect expiry information and refresh behaviour using the provider's documented flow. Do not assume every token is self-describing or try to infer validity by decoding an opaque value.
ServiceTitan's first-API-call guide, for example, documents its authentication sequence. Other providers have different lifecycles and permission models.
Use a safe capability check
Choose a low-impact, provider-supported check that verifies the access needed by the workflow. A read-only account or resource request may help, but it does not necessarily prove write permission.
| Check result | Operational response |
|---|---|
| Access valid for the intended capability | Keep the dependency available |
| Access revoked or scope missing | Disable the affected action and notify its owner |
| Provider temporarily unavailable | Use bounded retries and the approved fallback |
| Check result uncertain | Keep the distinction visible in monitoring |
Do not create real bookings or paid calls as an unnoticed heartbeat. If you need a write-path test, use a controlled test environment or an explicitly managed synthetic workflow with cleanup.
Route alerts to the person who can restore access
Record the integration, tenant, environment, last successful check, observed failure category, and responsible owner. Keep the secret itself out of alerts and logs.
Differentiate a permission failure from a provider outage. Asking the client to reconnect an account during a temporary network problem creates unnecessary work and may obscure the original evidence.
The receptionist also needs a usable response. Disable only the affected capability where practical and explain that the action cannot currently be confirmed. Preserve any collected request under the agreed fallback process.
Test revocation before it happens unexpectedly
In a safe environment, revoke access, remove a required permission, expire a credential where supported, and interrupt the provider connection. Verify alert routing and the caller-facing fallback.
Then restore access and check that normal operation resumes deliberately. A dashboard turning green should correspond to a successful capability check, not merely to someone saving a new credential.
The agency onboarding checklist should assign credential ownership. The monitoring and outcome checks help verify the business workflow after access is restored.
Credential-monitoring questions
Is an expiry date enough?
No. Revocation and changed permissions can invalidate access earlier.
Should the AI see provider credentials?
Keep them in the authorised integration layer, outside conversational context.
What should a client handover include?
Name the credential owner, renewal process, safe checks, and fallback. Include those in your operating review.