What running AI agents actually costs
Seats, usage, tokens and the setup nobody prices in. How to read an AI agent bill before it arrives.
· 2 min read · AmplifyIQ
The price on an AI tool's pricing page is rarely what you end up paying. That's not always sharp practice. Agents are billed differently from ordinary software, and the bill depends on how much work they do. Here is how to understand it before it lands.
Two kinds of cost
The platform. A monthly fee for the software that runs your agents, connects them to your tools and keeps their history. Usually per seat, per agent or per workspace.
The thinking. Every time an agent reads an email, a document or a web page and writes something back, it uses a language model, and language models are billed by the token: roughly three-quarters of a word. Input (what the agent reads) and output (what it writes) are usually priced separately, and output costs more.
Some products bundle the thinking into the platform price with a usage cap. Others pass it through at cost. Either is fair, as long as you know which one you're getting.
What makes an agent expensive
- Reading a lot to write a little. An agent that reads a 40-page contract to answer one question uses far more tokens on input than output.
- Long memory. Agents that carry the whole history of a conversation re-read it on every turn. A long-running task gets more expensive per step as it grows.
- Thinking out loud. Some models reason at length before answering. That's great for hard problems, and wasteful for simple, fixed-shape jobs like "is this email a complaint? yes or no".
- Loops. An agent that retries a failing step, or checks for new email every minute when every hour would do, bills for each attempt.
What keeps it cheap
- The right model for the job. Small, fast models for sorting and extracting. Larger ones only where judgement is needed.
- Tight instructions. "Reply in under 100 words" costs less than an open question.
- Batching. Summarise the day's enquiries once, not each one as it arrives, if nobody needs them in real time.
- Limits. A spending cap per agent, and an alert before it's reached.
The cost nobody prices in: setup
The largest cost of most agent projects isn't software or tokens. It's the time spent deciding what the agent should do, connecting it to your tools, and checking its work in the first few weeks. Budget for that honestly, whether it's your own time or someone you pay.
A sensible way to budget
- Pick one job and estimate how many times a month it happens.
- Run the agent on a week's worth and look at the actual usage.
- Multiply up, add the platform fee, and compare with the hours saved at what those hours cost you.
If the saving isn't clear at that point, it won't become clear at scale. If it is, you'll know exactly what the next agent will cost before you build it.