Every search for AI agent pricing lands in the same place: “it depends”. It really does, but not mysteriously. It depends on four nameable things, and anyone who understands the four can read any quote in ten minutes.
What moves the cost of an AI agent is not the number of conversations: it is the ratio of voice to text, the number of languages, the depth of the integrations and who curates the knowledge base. The platform licence is usually the smallest of the four lines, and the only one that appears on a pricing page.
The four lines
Agent cost is not a number, it is a sum. Separating the parts avoids the most common surprise: signing something cheap and discovering the real cost three months later, once it has already changed owners inside the company.
1. Platform licence
The part on the pricing page, and the easiest to compare. Look at what the plan includes and what it meters: conversations, agents, channels, workspaces. A low price with a tight conversation cap becomes a high price in the month the campaign works.
2. Model consumption
Text and voice live an order of magnitude apart. A text conversation consumes tokens; a real-time voice conversation consumes audio in both directions, continuously, for the whole length of the call. If your case is mostly telephone, this line dominates, and the product design changes with it: latency, average duration and transfer rate become financial metrics, not only experience ones.
3. Deployment
Somebody’s time to point the agent at the right content, write the instructions, define the hand-off rules and test. On a site with a tidy FAQ and catalogue, that is days. In an operation whose knowledge lives in three people’s heads, it is the most expensive of the four lines, and nobody puts it in the spreadsheet.
4. Maintenance
Prices, exchange policies, delivery times and catalogues change. Someone has to keep the knowledge base reflecting that. It is a small amount of time per week, but it is recurring, and it is what separates an agent that ages well from one that starts getting things wrong on its own in month four.
The mistake of comparing list prices
Two quotes with the same monthly figure can carry total costs that differ fivefold, because one assumes your content is ready and the other includes the curation. Ask for the whole picture: licence, estimated consumption at your volume, deployment hours and who maintains it afterwards.
A rough example: 2,000 text conversations a month cost a fraction, in model consumption, of what 200 ten-minute calls cost. Before comparing vendors, decide which of the two you are.
What about building it?
It is the alternative that looks cheapest and almost never is, because the cost migrates from a visible line (a subscription) to an invisible one: engineering salary, indefinitely. The full comparison is in AI agent vs. building your own, including the cases where building really is the right call.
How to estimate before talking to anyone
It helps to have your own number before requesting a quote, if only to know whether the conversation is worth having. The ROI calculator runs that estimate on your traffic, your conversion rate and your average order value, with both assumptions exposed for you to adjust. If the result does not clear the cheapest licence on the market comfortably, the problem is not the price of the agent: it is the volume of the site.
