A voice AI quote can include the build and leave out almost everything that happens after launch. Another quote can bundle usage and support into a monthly price. Comparing the headline numbers tells you very little.
I would split the estimate into three parts: the initial engineering work, the cost of running calls, and the cost of keeping the workflow working.
What belongs in a voice AI budget?
Development covers the work needed to connect the agent to a specific business process. That might include a call flow, a phone connection, a customer-system integration, tests and handover documentation. The scope matters more than the number of prompts.
Operating costs can include telephony, the agent runtime, speech recognition, language-model processing, speech output, recordings and hosting. Some providers bundle several of these. Others charge separately or pass through another provider's bill.
Support covers investigation, changes and maintenance. A new CRM field, a changed business process or an expired credential can create work even when the agent's conversation has not changed.
Start with call minutes, then price the components
For a first estimate, multiply monthly calls by average connected duration. Include abandoned calls and transfers according to how the relevant services meter them. A short enquiry and a long troubleshooting call should not share an assumed duration without evidence.
Here is an illustrative worksheet, not a vendor quote or a forecast:
| Input | Example assumption |
|---|---|
| Monthly connected calls | 600 |
| Average connected duration | 3 minutes |
| Monthly connected minutes | 1,800 |
| Combined variable rate assumed for the exercise | $0.08 per minute |
| Variable usage under those assumptions | $144 |
| Hypothetical fixed hosting and number costs | $40 |
| Illustrative operating subtotal | $184 |
The arithmetic is 600 × 3 × $0.08 + $40. It excludes engineering, support, taxes and any usage missing from the assumed rate. The purpose is to expose the inputs, not suggest that $184 is the price of your deployment.
In a real worksheet, replace the combined rate with line items from your providers. Record the currency, region, plan, billing unit and date checked. Model tokens, generated speech and connected phone minutes are different units; convert them using explicit workload assumptions.
A cheaper minute can still produce a more expensive outcome
Suppose one configuration spends less on speech but requires callers to repeat their details. Longer calls can erase the saving. A failed transfer can also create staff work that never appears on the voice provider's invoice.
Track an outcome alongside usage: completed service requests, correctly routed calls or another task your team can verify. Divide the relevant operating cost by that outcome count. Keep the definition stable when comparing configurations.
I would also price a busy period separately. Concurrency limits, higher call volume and support coverage may affect the architecture or plan even if the average month looks small.
How my published service prices fit into the budget
My service page currently lists a $150 diagnosis, a $500 scoped repair that includes diagnosis, a $250 handover review, and a pilot from $1,500. These are my listed offers, not market averages. Scope and current pricing must be confirmed before work starts.
A custom agency build is quoted after reviewing the workflow. Phone, model and hosting charges are separate. A pilot starting price should not be read as the price of an unrestricted multi-client platform.
For provider inputs, check the current LiveKit pricing page and your carrier's own rate sheet. This article deliberately does not freeze their changing rates into a supposedly universal per-minute price.
Questions about AI phone-agent pricing
Can I use the free tier to estimate production costs?
It can help with an initial experiment. Build a second estimate using the plan and allowances you expect after launch, including what happens when an allowance runs out.
Should maintenance be included in the build price?
Only if the agreement defines it. Specify the support period, included changes, response expectations and usage charges rather than relying on the word maintenance.
How do I compare two proposals?
Give both providers the same workflow, expected call volume, integration requirements and acceptance checks. Ask each to separate one-time work from recurring charges.
Bring those inputs to a scoping conversation. They are enough to make a cost discussion more useful than asking for a price for an unspecified AI agent.