AI Agent Cost Calculator

The Four Things You Actually Pay For

Most quotes show one number. There are four, and only one of them is fixed.

1. Model usage — metered, and the one that scales with success. You pay per token, so cost tracks conversation length as much as conversation count. A chat agent answering a two-turn question is trivially cheap. The same agent handed a long support history in every request is not, because the whole history is re-sent and re-charged on every turn. Prices per million tokens have fallen steadily and vary by an order of magnitude between the small and large models in the same family, so check the provider's current pricing page — OpenAI's API pricing is the one most builds start from — rather than trusting any figure you read in an article, this one included.

2. Platform fees — mostly fixed, with a metered tail. Automation platforms such as Make.com and n8n sit in the low tens per month at small scale. Voice platforms such as Retell and Vapi charge per minute of call time, and that per-minute rate usually sits on top of the model cost and the telephony cost rather than including them — Retell's own documentation is explicit about which components are billed separately. CRM subscriptions are separate again.

3. Build time — one-off, and larger than the demo suggests. Getting an agent to work in a demo is quick. Getting it to handle the caller who interrupts, the record that does not exist, the API that times out and the question nobody anticipated is the actual work. A production-ready agent is normally a matter of days rather than hours.

4. Maintenance — small, recurring, and skipped in every estimate. Prompts need tuning once real conversations arrive, integrations change, and edge cases surface for months. Budget a couple of hours a month and you will be roughly right.

Figures on this page describe the shape of the costs rather than current prices. Model pricing, per-minute voice rates and platform tiers all move; check the vendor's page for today's number.

A Worked Example

A dental practice wants an agent to answer overflow calls and book appointments. Around 400 calls a month, averaging three minutes each — 1,200 minutes.

• Voice platform, per-minute rate × 1,200 minutes. This is the dominant line and it scales linearly with call volume.
• Model usage, one conversation's worth of tokens per call. Small next to the per-minute cost at this volume.
• Telephony, the actual phone number and inbound minutes, billed separately by Twilio or the platform's bundled equivalent.
• Automation platform, a fixed monthly subscription to move the booking into the practice management system.
• Build, a one-off covering conversation design, calendar integration and testing.
• Maintenance, two hours a month.

Two things fall out of this that a single headline number hides. First, the per-minute costs dominate, so the lever that changes your bill is average call length, not the model you picked. Shortening the agent's turns and getting to the booking faster is worth more than any model swap. Second, doubling call volume roughly doubles the running cost, which is exactly the opposite of how software usually behaves and the reason the ROI case has to be built on cost per call rather than cost per month.

AI Agent vs a Human

The honest comparison is narrower than the one usually presented.

The agent wins decisively on the calls that currently go to voicemail, because the alternative there is nothing at all rather than a person. It wins much less clearly against a good receptionist during working hours. Most sensible builds treat it as overflow and out-of-hours cover rather than a replacement, and the business case is stronger for being framed that way.

Costs People Forget

Telephony is separate. The per-minute platform rate rarely includes the carrier cost. Check whether the quoted rate is all-in.

Failed and abandoned calls still cost. A caller who hangs up after twenty seconds has still consumed minutes and tokens.

Testing is billed. Every test call during the build runs through the same meters.

The second build. Nearly every agent gets a substantial revision four to six weeks in, once real conversations have shown what the design missed. It is not a failure, it is the normal shape of the work, and it belongs in the budget.

Frequently Asked Questions

Why is the per-minute rate the number that matters most?

Because at any realistic call volume it dwarfs the model cost. A voice agent's bill is essentially minutes × rate, so the two things that move it are how many calls you take and how long each one lasts. Trimming thirty seconds off an average call cuts the bill by a sixth. Switching to a cheaper model saves a fraction of a much smaller line.

Does a cheaper model make the agent noticeably worse?

For narrow, well-defined tasks such as booking, qualifying and routing, often not. The smaller models in a family handle structured conversation with a tight prompt perfectly well. Where the gap shows is open-ended reasoning and unusual phrasing. The practical approach is to build on the stronger model, get the conversation design right, then test whether the cheaper one holds up on your actual transcripts.

How do I stop costs running away if volume spikes?

Set a spend cap at the provider and a separate one at the platform, and make sure a breach alerts you rather than silently degrading. Then cap conversation length in the agent's design: a hard limit on turns, with a transfer or callback offer once it is reached. An agent with no exit condition will happily talk to a confused caller for fifteen minutes.

Is a chat agent cheaper than a voice agent?

Substantially, because there is no per-minute charge and no telephony. A chatbot's running cost is essentially model usage plus a platform fee. If the interaction does not need to happen on the phone, it is the cheaper route by a wide margin.

Related Tools and Reading

The AI Voice Agent Cost Simulator models the call-volume side specifically, and the Automation ROI Calculator turns the cost figure into a payback period. For what actually goes into the build, How to Use Retell AI walks through it step by step, and AI Voice Agent for Small Business covers the ROI case in more depth.

Build Your AI Agent

I build these systems — voice agents, chatbots and the workflow automation behind them — for clients across property, healthcare, home services and more. Book a free discovery call for a fixed-price quote based on your actual call or conversation volume, or take on the work through Upwork.