Compare complete operating models, not an AI minute rate with an employee’s hourly wage. Request like-for-like scenarios for better routing, human answering, an existing platform’s native capability and any proposed specialist deployment.
For ServiceTitan users, check the existing account before commissioning a separate integration. Its FAQ, updated on 22 October 2026, describes inbound booking, confirmation, rescheduling and configurable escalation, alongside package requirements and limitations. That is a reason to investigate the native option—not evidence that it fits every account or passes your booking tests. (help.servicetitan.com)
What belongs in the AI voice agent cost model?
- Implementation, integration, testing and internal project time.
- Platform, model, speech and telephone charges, without double-counting bundled items.
- Minimum spend, numbers, concurrency and support commitments.
- Transfer legs, waiting time and human handling after transfer.
- Monitoring, incident investigation and corrections.
- Repeat contacts and unsuccessful attempts.
- Ongoing configuration changes and exit costs.
Billing boundaries deserve explicit questions. For example, Retell’s pricing FAQ says silence and hold time are billable, and that telephony charges continue after transfer even when its AI-agent fee stops. These are supplier-specific terms, not universal market rules. (retellai.com)
Use this desk’s governing metric:
Fully loaded workflow cost ÷ unique, correctly completed, policy-compliant tasks.
Report autonomous completions and human-rescued completions separately. Include both in the overall workflow result only when each meets the same completion definition; do not call the combined rate “autonomous resolution”.
What does the calculation look like?
The following is an illustrative monthly scenario, not a quote, forecast or benchmark. Assume 1,000 calls representing 1,000 distinct task attempts.
| Assumed cost | Amount |
|---|---|
| AI usage: 1,000 calls × four minutes × $0.18 | $720 |
| Human rescue and correction labour | $800 |
| Monitoring and quality review | $400 |
| Other recurring charges | $180 |
| Allocated implementation cost | $300 |
| Total | $2,400 |
Assume 600 tasks complete autonomously, 200 complete with human help and 200 remain unsuccessful. Fully loaded cost is $2,400 ÷ 800 = $3.00 per successful task. Direct AI usage alone is $0.72 per attempted call; it is not the outcome cost.
Show implementation cash outlay separately from its monthly allocation. Compare alternatives over the same workload and time horizon.
How should you recognise benefits?
For this decision model, keep four benefit categories separate:
- Cash savings: an identifiable payment actually falls.
- Avoided hiring: an approved future staffing requirement changes.
- Released capacity: staff time becomes available, without assuming payroll falls.
- Recovered contribution: additional completed work generates contribution after fulfilment costs.
For missed-call recovery, use a waterfall: missed calls → unique enquiries → eligible demand → incremental bookings → completed jobs. Deduct demand already recovered through callbacks and constrain the estimate by actual capacity.
Decision: approve a capped experiment only if conservative assumptions justify learning more. Do not require speculative revenue recovery to disguise poor service economics.
Owners: finance and operations. Deliverable: comparable scenarios, sensitivity analysis and pilot spending limit.