The agent says the replacement card is on its way. The customer hangs up relieved. The card-order API had timed out, so no order was created.
Three days later the customer calls again. A supervisor now has a complaint, an avoidable callback, and a person searching through call logs to work out what happened.
The first call was complete. The customer request was not.
A good conversation can still create work
Voice teams tend to watch what is closest to the conversation: transcription quality, response time, interruptions, intent recognition, and whether a tool returned an error. Those signals matter. They do not tell you whether the customer got what they called for.
In the replacement-card call, the agent sounded clear and empathetic. It made the right API call. The failure sat in the gap between a request being sent and an order actually existing.
That gap is where repeat contact starts. It is also where teams can fool themselves with a healthy containment rate. The call did not reach a person, but the customer still has to come back.
Ask for proof before the agent makes a promise
For every action an agent can claim to complete, decide what proves it:
- A replacement card: an order reference returned by the card system.
- A service appointment: a slot that appears in the scheduling system with the right contact details.
- A support case: a case number a customer can use.
- A callback: a task assigned to a real queue with a time attached.
The evidence should shape the language. If the card system returns an order reference, the agent can confirm the order. If it times out, it should say what is true: the request is not confirmed, and someone will check it.
That is not cautious wording for its own sake. It prevents a customer from waiting for work that never started.
Review the calls that create doubt
You do not need to inspect every conversation. Start with calls where the outcome is uncertain:
- the agent used confirmation language but no reference number exists
- a tool retried or timed out
- a transfer or callback happened
- the caller repeated the same request or called again
- a human later corrected the work
Give an operator one record with the call, the claimed action, the tool result, and the current customer state. The operator should be able to answer one question quickly: does this customer still need something from us?
If the answer is yes, the record needs a next owner and a deadline. A transcript alone cannot do that work.
The management question
Look at last week's calls marked as resolved. How many created a case, booking, order, or callback that actually exists?
If nobody can answer without opening several systems, the operation is measuring conversation completion. It is not measuring customer resolution.