How is an AI agent priced, and what will you pay in practice?
AI agents charge in one of four ways, or a mix of them: per seat, per credit or unit of usage, per task, or per outcome. The plan name on the pricing page (free, freemium, paid or "contact sales") tells you where you start. The unit behind the price tells you what you will pay once you use the agent every day.
I have been caught by this myself. A plan looked cheap, then every action the agent took burned credits, and the second month cost three times the first. So this article explains each pricing model, what to look for on the page, and a simple way to estimate your bill before you commit.
The four labels you see first
Most listings and pricing pages start with one of these labels.
Free means you can use the product at no charge. Open source projects often sit here, and the cost moves to hosting and to the model you connect.
Freemium means a free plan with limits, and paid plans above it. The limits are the part to read: number of runs, number of agents, storage, or which models you can use.
Paid means a subscription or a usage bill from the first day, often with a trial.
Ask for pricing means the price depends on your size and needs, and a salesperson will quote it. This is common for agents sold to companies.
The four units behind the price
Per seat
You pay for each person who uses the agent, every month. This is easy to budget and familiar from other software. Check whether occasional users count as full seats.
Per credit or unit of usage
You buy credits, and each action spends some. An action might be a message, a pipeline run, a minute of voice or a page the agent reads. This is the model that surprises people, because one task can contain many actions. Look for a table that says what each action costs.
Per task
You pay a fixed price for each finished task, such as a support ticket resolved or a research report delivered. This is easy to compare with what a person costs for the same task. Ask what happens to the charge when a task fails or needs a retry.
Per outcome
You pay when a business result happens, such as a meeting booked or a lead qualified. It lines up the vendor's interest with yours. Agree in writing how the outcome is counted.
Costs that sit outside the price tag
Some agents ask you to bring your own model key. In that case you pay the model provider separately for every call, and that bill can be larger than the subscription.
Self-hosted and open source agents cost server time and someone's attention to keep them updated. That is often still the cheapest option for a technical team, as long as you count the hours.
Integrations can carry their own cost too. If the agent needs a paid tier of your CRM or helpdesk to connect, add it to the total.
Estimate your monthly bill in four steps
- Write down one real task you want the agent to do, and how many times a month it happens.
- Run that task once on the free plan or trial, and note how many credits, runs or minutes it used.
- Multiply by your monthly volume, then add a third for retries and growth.
- Compare the result with the plan limits, and with the cost of a person doing the same work.
This takes an hour and turns a pricing page into a number you can defend in a budget meeting.
Questions to send the vendor before you buy
- What counts as one unit of usage, and can I see my usage while the month runs?
- What happens when I reach the limit: does the agent stop, or does it bill overage?
- Are model costs included, or do I bring my own key?
- Can I change plans or cancel monthly?
- Is there a discount for a yearly plan, and is it refundable?
Questions people ask
Why do AI agents use credits instead of a flat price?
Each action an agent takes calls a language model, and those calls cost the vendor money. Credits pass that cost on in proportion to how much you use. Flat prices exist too, usually with a usage cap.
Is a free AI agent free to run?
The software can be free while the model calls, hosting or integrations cost money. Open source agents are free to use and you pay for where they run and the model they call.
How do I compare prices between two AI agents?
Run the same task on both, count the units each one used, and work out the cost per finished task. That number compares fairly across seats, credits and per-task pricing.
