*Part 1 of 4: what it actually costs to put agents to work in your business, and where the waste hides.*
Here's a Monday morning that's becoming common. A business owner opens the bill for the AI tools their team started using, and the number has doubled. Nothing broke. Nobody misused anything. Somebody just ran a big automated job on Friday, and the meter ran all weekend.
The two instinctive fixes are the two wrong ones.
The first is switching everything to a cheaper model. That sounds like savings, but a weaker model doesn't fail cheaply. It fails slowly. It takes more tries, reads more files, hits more errors, and hands you work you have to redo. You didn't cut the bill. You spread it across more attempts.
The second is putting a hard cap on spend. Caps don't reduce waste. They reduce use. The people getting the most value out of the tools hit the ceiling first, start rationing, and stop using them where they pay off. The waste stays exactly where it was. You just get less done at the same rate.
The real fix starts with understanding what you're actually buying.
## The bill is six numbers multiplied together
Every AI agent bill, at any scale, breaks down the same way:
```
Spend = People using it
× Sessions per person
× Turns per session
× Requests per turn
× Tokens per request
× Price per token
```
Plain English for the unfamiliar terms: a *session* is one working conversation with the tool. A *turn* is one back-and-forth inside it. A *request* is each time the tool has to go ask the model something behind the scenes. *Tokens* are the units the model reads and writes, and they're what you're billed for.
Written this way, the six numbers sort into three groups, and each group has a different owner.
**The first two are adoption.** People using it, and how often. You want those going up, because that's the return side. Any cost fix that pushes adoption down, like a cap, made your business slower and called it savings.
**The last one is the price sheet.** The model vendors set the price per token. You don't negotiate it. What you control is which model handles which job, and that's a routing decision, not a pricing one.
**The middle three are yours.** Turns, requests, and tokens all measure the same thing from different angles: the work the tool does on its own behalf, on top of the work you asked for. Time spent hunting for information it could have been handed. Instructions and tool definitions it re-reads on every single turn whether it needs them or not. Long wandering sessions that should have been three short, clean ones.
That middle work is the only pure waste in the equation. A token spent on your task is money well spent whatever it costs. A token spent re-reading the same setup for the fortieth time bought you nothing.
## The proof, from a company that measured it
Uber published their numbers in August 2026, and they're worth knowing even if your business is a thousandth of their size, because they isolated the thing that matters.
Their usage exploded. Weekly active users of their agent tooling grew 7x in six months, and weekly requests grew 9.5x. Their total spend stayed roughly flat.
To prove that was engineering and not vendors quietly cutting prices, they held one model constant and measured only their own changes. On that basis, cost per session fell 52 percent from its peak. Same model. Same price sheet. Half the cost.
None of it came from buying cheaper AI. It came from working the middle of the equation: handing agents the context they'd otherwise burn time hunting for, cutting the setup material re-sent on every turn, and moving repeated multi-step chores into scripts so the in-between steps stop hitting the meter at all. In one of their tests, an agent that was given the right context answered a question in 38 seconds. The same agent on the same model, without that context, spent twenty minutes searching and got the answer wrong.
The gap wasn't intelligence. It was knowing where to look.
## How we run this
This is how we build at Apex, because we run the same workflows across many businesses, and every wasted turn multiplies by every client we run it for.
A few rules we hold ourselves to:
**Measure per task, not per month.** A monthly total tells you the bill went up. What you want to know is what one nightly report, one review reply, one reorder check actually costs to run. When you know the number per task, you can decide if the task is worth it, and you notice immediately when it drifts.
**The tool carries only what the job needs.** Most AI setups quietly load everything they *might* need into every conversation, and you pay for that baggage on every turn. Connecting a business system shouldn't mean carrying its whole manual around. It should mean reaching for the one drawer the job requires.
**Repeated work becomes a script, not a conversation.** If the same six steps run every night, those steps shouldn't be re-reasoned from scratch at model prices every night. They get written down once, run as a routine, and the model only gets involved where judgment is actually needed.
**Prove the model choice before making it.** Cheaper and worse look identical on an invoice. The only way to tell them apart is to test candidate models against your real tasks with known-good outcomes, then pick on results per dollar. Uber did this before switching models on their code review system and got better quality *and* lower cost. That's not luck. That's what the test is for.
And underneath all of it: nothing changes without your OK, and every change can be undone. An agent that's cheap to run but free to act without you isn't efficient. It's a liability with good margins.
## The takeaway
You don't control the price of AI. You control what the tools carry, what they already know, and how the repeated work is shaped. That's where the waste lives, that's where the savings compound, and none of it requires buying worse tools or using them less.
Next in this series: why handing an agent the right context beats paying for a smarter one.
---
*Apex Studio puts your whole business in one place, so the tools working on it start with the context instead of hunting for it. Start yourself, or have our team bring your systems in and run it for you.*