An AI phone call can use minutes, trigger retries and create staff work without resolving the caller's request. If the business counts only completed calls, those unresolved attempts disappear from the cost discussion.
For owners evaluating a pilot, measure the work required to reach a useful final outcome. Keep software charges and staff recovery effort visible without assigning fictional revenue to every conversation.
What is an unresolved AI phone call?
An unresolved call ends without the agreed business outcome or a properly owned next step. A callback request can be a valid outcome when that is the intended service; it is unresolved when the caller was promised a confirmed booking and nobody owns the missing work.
Write the outcome definition before measuring. Otherwise the team may reclassify failed automation as successful intake merely to improve the dashboard.
Calculate recovery cost with explicit inputs
Consider a fictional pilot with 20 calls requiring staff correction. Staff spend a measured total of 100 minutes reviewing records and contacting callers. At an illustrative internal labour allocation of 30 currency units per hour, that represents 50 units of recovery effort.
| Cost component | Evidence to use |
|---|---|
| AI usage | Billable records for relevant attempts |
| Telephone usage | Applicable provider charges |
| Staff recovery | Recorded handling time and chosen allocation |
| Duplicate correction | Time and direct reversal costs where applicable |
| Customer follow-up | Measured additional contact work |
The example is arithmetic, not a market labour rate or a claimed client result. Use the business's own accounting approach before making a purchasing decision.
Avoid counting the same work twice
If staff recovery already includes the callback, do not add that same time under another category. Link costs to a defined request where possible, since several calls may relate to one enquiry.
Separate direct measured costs from consequences that remain uncertain. A caller leaving without booking may represent lost opportunity, but the business cannot assume every caller would have purchased. Record the evidence gap rather than fill it with a convenient sales value.
The call-log cost model addresses workload charges. Combining it with outcome and recovery records gives a more complete pilot view.
Use the result to choose a repair
Group unresolved cases by cause: missing details, unavailable integration, wrong routing or unsupported requests. The largest call category is not necessarily the most expensive if another category requires much more staff effort per case.
Choose a contained repair and remeasure the affected workflow. If the service still requires too much supervision, narrow coverage or retain human-first reception while the underlying boundary is addressed. A smaller useful service can be better than a broader one with hidden recovery work.
Is every staff-handled callback a failure?
No. Judge it against the intended service and whether a responsible person received usable information.
Should we include assumed lost revenue?
Keep it separate from measured operating costs and state the evidence and assumptions clearly.
What should we improve first?
Choose the failure category with meaningful business impact and review its evidence through the voice AI project contact.