An unavailable field-service API must not leave an urgent caller hearing that help is on the way when nobody has accepted the work. Design the outage path before the first live call.
Separate emergency instructions, business urgency, and scheduling availability. The business should approve the safety language and escalation policy. The voice workflow should not diagnose danger, promise emergency response, or improvise technical repair instructions to compensate for a failed integration.
Preserve the request outside the failed dependency
If intake storage depends on the same unavailable system, the fallback is fragile. Use a separate durable request record containing the service property, callback route, reported concern, timestamps, and attempted actions. Limit collection to facts needed for the next operator.
Show the agent a clear tool result: unavailable, rejected, uncertain, or confirmed. These are operationally different. A timed-out creation may require reconciliation; an explicit failure before submission may permit a later retry.
Rehearse an after-hours outage
A fictional heating company receives a no-heat call after hours while its scheduling API is unavailable. The agent captures the address and callback number, follows the approved urgent-call questions, and attempts the on-call transfer.
If the on-call person accepts, record that acceptance separately from a job booking. If nobody answers, retain the request and explain the actual fallback. A notification sent to a phone does not prove the technician received or accepted the request.
| Outage scenario | Required evidence |
|---|---|
| API unavailable before write | Durable pending request |
| Write submitted, response missing | Uncertain action requiring lookup |
| On-call person answers | Handoff acceptance recorded |
| Transfer destination unavailable | Approved secondary route followed |
| System recovers | Queue reconciled before replay |
Choose a realistic degraded service
Some businesses can continue intake and manual dispatch during an outage. Others must stop offering times entirely. Decide that boundary explicitly and keep the fallback script short enough for a stressed caller.
A secondary notification channel improves resilience only if it does not share every dependency with the primary path. Test it from the actual call route. Staff also need a way to see which requests remain unacknowledged and who owns them.
The tradeoff is extra operational work during an incident. Automatic replay can reduce that burden for clearly failed writes, but it must not duplicate jobs or book expired preferences. Review each urgent request against its current state before normal service resumes.
The human call transfer guide provides related questions for testing the handoff boundary.
Questions about urgent outage handling
Should the agent say a technician has been notified?
Only when the notification action is verified, and clarify that notification differs from acceptance or dispatch.
Can intake continue when scheduling is down?
Yes if the request can be stored safely, staff can act on it, and the caller receives accurate expectations.
What is the most important recovery metric?
Track unacknowledged urgent requests and their age. A recovered API does not mean affected customers have received a useful response.