An average call duration hides the short wrong-number calls, unanswered transfers and long enquiries that can all affect voice AI costs differently. A spend estimate becomes more useful when it starts with the business's actual call pattern.
This guide is for owners estimating a proposed AI receptionist. It uses call logs to model workload, without treating historical human calls as an exact prediction of future AI conversations.
What information should a voice AI cost estimate use?
Start with eligible inbound calls, duration, time of day and intended coverage. You usually do not need call audio or complete customer identities to create an initial workload summary.
Document what the provider bills: connected time, rounding increments, transfer legs, inference components or bundled allowances. These definitions differ by supplier and plan, so use the current written quote rather than assume a universal per-minute rule.
Work through a fictional call-log sample
Suppose an illustrative week contains 120 calls within the proposed coverage period. Of those, 80 last two minutes, 30 last four minutes and 10 last eight minutes. The observed total is 360 minutes. These are fictional inputs, not a benchmark.
| Workload category | Count | Observed minutes |
|---|---|---|
| Short enquiries | 80 | 160 |
| Standard enquiries | 30 | 120 |
| Longer calls | 10 | 80 |
Do not multiply 360 minutes by a headline price and stop. Apply the supplier's actual billable units, then add subscription charges, numbers and other quoted components. If some calls transfer to staff, ask how each relevant leg is charged.
Model uncertainty explicitly
AI conversations may be shorter or longer than historical calls. Routing may also capture callers who previously reached voicemail and hung up. Build a baseline, a lower-workload case and a higher-workload case using stated assumptions.
Change one assumption at a time when discussing sensitivity. For example, show what happens if the AI handles more calls, if the duration grows or if transfers become more common. This tells the owner which pilot measurements will most improve the next estimate.
A call log cannot reveal whether an enquiry produced a useful outcome. Pair workload estimates with a proposed outcome definition such as a verified request, completed booking or staff-accepted transfer.
Replace assumptions with pilot evidence
During a controlled pilot, compare connected call records with the bill and the resulting business actions. Investigate differences between expected and billed usage before widening coverage.
The voice platform case study gives engineering context, not a universal cost claim. Each proposed service needs its own billing reconciliation and operating assumptions.
Do we need to upload recordings to estimate spend?
Usually a minimal call summary is sufficient for an initial workload model. Request only data needed for the decision.
Can historical averages predict the final bill exactly?
No. Routing, caller behaviour and provider billing definitions can change the result.
What should we measure during the pilot?
Track eligible calls, billable usage and verified outcomes together, then review the model through the voice AI contact section.