An assistant should confirm a business action only when the tool returns evidence of completion. A timeout, an accepted job, and a rejected write require different wording even if all three interrupt the same conversation.
Tool errors need outcome semantics
Handling a tool error means deciding what is known about execution, what remains uncertain, and what recovery is allowed. An exception alone does not establish that nothing happened. A remote service may commit a write before its response is lost.
A proposed response contract separates completed, rejected, pending, and unknown outcomes. Completed includes an authoritative reference. Rejected includes a safe reason and whether correction is possible. Pending includes an operation reference for later lookup. Unknown requires reconciliation before claiming success or retrying a non-idempotent write.
Keep technical diagnostics in protected logs. The agent needs a stable failure category and a truthful next step, not a database stack trace. Caller-facing wording should reflect the outcome category rather than the emotional tone of the exchange.
Compare four fictional booking responses
A caller asks for Monday at 9:00. If the tool returns a confirmed appointment reference, the assistant can read back that appointment. If the service says the slot is unavailable, it can offer alternatives. If the request is queued, it should explain that confirmation is still pending.
When the response is lost after submission, “you are booked” and “nothing was saved” are both unsupported. The next step is checking the operation reference or using the approved staff route.
| Tool result | Permitted conversation claim |
|---|---|
| Confirmed appointment reference | Appointment was created |
| Validation rejection | Specified information needs correction |
| Accepted background job | Request was submitted, not yet confirmed |
| Unknown execution status | Completion is being checked |
Include a test where the model previously promised to book before the error arrived. The final response must correct that expectation using the actual result, rather than preserve conversational consistency at the expense of accuracy.
Recovery should preserve the original operation
For transient errors, retry only under the tool’s defined duplicate-protection rules. For validation failures, ask a targeted question. For denied actions, use the approved verification or escalation route. Repeating every failure with the same broad tool call is not a recovery policy.
Strict outcome handling can sound less confident. That is a useful tradeoff when the alternative is a false confirmation. Measure unsupported success statements as a distinct release defect.
The voice platform work describes relevant integration experience; these response semantics are proposed. Read partial-success handling when several operations contribute to one caller request.
Tool failure FAQ
Does a timeout mean the write failed?
No. It may mean the outcome is unknown until reconciled.
Is an accepted request the same as completion?
No. Background work needs a later authoritative completion signal.
Should the assistant apologise and continue?
It can acknowledge the problem, but must also state the known outcome and the next available action.