August 18, 2026 · By Katherine
What Do AI Agents Actually Cost? Reading the Bill
AI agents are billed by how much work they do, not by how many of them you have. Every run costs, and every run is priced by how much text goes in and comes back. So a cheap job running constantly beats an expensive job running weekly, and the bill rarely looks like what you expected.
The instinct that gets people is thinking of an agent like a subscription. It's closer to a utility meter.
The light left on in a room nobody uses
Your electricity bill isn't driven by the thing you think about. It's the appliance you never think about, running quietly, all the time.
Nobody budgets for a hallway light. It costs almost nothing per hour, which is exactly why nobody turns it off, and why it can still be a real line on the bill by the end of the quarter.
Per run, not per agent
Two numbers matter, and neither is the number of agents.
How often it runs. A job triggered by every incoming email runs as often as you get email. On a busy inbox that's hundreds of times a day, each one small.
How much text moves. Cost is per token, which is roughly per word in and out. An agent handed a whole document to be safe is paying for the whole document, every run.
Multiply those and you get the bill. Neither is visible when you set it up, because at design time you run it twice and it costs pennies.
Where the money actually goes
Almost always the same shape: something small, automatic, and forgotten.
- Summarize every incoming email. Genuinely useful. Also runs on every newsletter, receipt and notification.
- Check for changes every five minutes. Nothing changed 287 times, and you paid to find out each time.
- Stuff in all the context to be safe. Every run carries a document that mattered once.
The video you produce monthly and worry about isn't the problem. The trigger you set once and never looked at again is.
How to read your own bill
Sort by total spend, not by what feels important. Then ask two questions of the top line: how many times did this run, and how much did each run carry.
One of those two is almost always wrong in an obvious way. Either it's running on things it shouldn't (every notification), or every run is carrying more than it needs.
Both are cheap to fix once you can see them. Neither is visible until you look.
That's the cost of running agents day to day. What it costs to actually make one, mistakes included, is a different bill entirely.
The question
Open your bill and find the largest line. Then say out loud how many times that ran this month.
If you don't know, that's the answer.
How are AI agents priced?
By work done rather than by agent. Every run costs, and each run is priced by how much text goes in and comes back, so frequency and payload size drive the bill.
Why is my AI bill higher than expected?
Usually a small job running constantly. Something set up once, triggered often, and never looked at again, such as summarizing every incoming email including newsletters and receipts.
Does adding more agents cost more?
Not by itself. An agent that rarely runs costs almost nothing. Cost follows how often things run and how much text each run carries.
What is the most common waste?
Passing more context than the job needs. If every run carries a whole document to be safe, you pay for that document every single time.
How do I find the problem?
Sort your spend by total cost rather than by what feels important, then ask two questions of the biggest line: how many times it ran, and how much each run carried.