# How Not to Blow Your AI Budget

By Wes Pomeroy. Published May 31, 2026. https://wespomeroy.com/writing/how-to-control-ai-spending/

We're in the part of the cycle where you're supposed to spend a little recklessly.

The playbook for this is still getting written, and it's uneven - some people have a lot of it worked out, plenty of us are further behind, and the only way to find out what AI is actually good for in your business is to let people loose and see what they build. That's not a mistake. That's the right instinct. If you've been pouring budget into experiments and your finance team is starting to twitch, you're probably doing something right.

So I want to say the careful part first: I'm not here to scold anyone who's overspent. Loosening the reins was correct. The exploration is how you learn.

## We've seen this before

Every time a new, powerful, metered thing arrives, the same pattern plays out. We adopt fast, we govern late, and somewhere around month four the bill shows up and it's an issue. It happened with cloud. Long before that, it happened with the long-distance phone call. Anyone over forty remembers it - calling another state cost real money by the minute, and you watched the clock the whole time. The meter ran in the background while you talked, and plenty of people only understood their own usage when the bill arrived. AI is the same story, just faster and bigger. It's not that anyone's being careless. The spend is genuinely hard to see in the moment.

Uber's COO, [Andrew Macdonald, was candid about this recently](https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/): their engineers have adopted AI tools enthusiastically, and he still can't draw a clean line from that spend to measurable product gains. I don't read that as a confession of failure. I read it as a serious company being straight about a rough edge a lot of us are circling in our own way. So rather than pile on, treat what they learned as a gift - free signal we can use to tighten our own setups before our own bill lands.

## Your hand is on a dial

You're managing a dial between cost and innovation. Crank it all the way toward cost and you strangle the experimentation that justifies the whole exercise. Crank it all the way toward innovation and you've got a faucet you forgot to turn off. Your hand is always on that dial. The work is just setting it on purpose instead of letting it drift.

The good news is you already know how to do this. Not the AI part - the managing part. You've watched budgets that could run away from you before. You've asked teams what a project was actually for. The only strange thing about AI is that it somehow got a pass on all of it, as if the normal questions didn't apply. They do. Here's what they look like pointed at this.

## Start with a real goal, not a gesture

What goal is this spend in service of? It's not enough to say you're "using AI in marketing." You want the actual outcome you're chasing and the specific thing you're building to get there - lifting qualified leads by automating the manual research step, say, rather than just doing lead gen with AI. The vague version feels fine right up until you try to measure it, estimate it, or cap it, and find there's nothing solid to grab onto. The specific version gives you all three.

It's the part people skip. A tool pointed at a fuzzy intention will happily spend money forever, because nothing ever tells it it's done. A tool pointed at a defined outcome has a finish line. Most runaway spend I've seen isn't reckless - it's just aimed at "explore AI" instead of at a thing.

## See where the money actually goes

You can't manage a number you can't break apart. "We spent a lot on AI last month" is not information. Which team, which project, which task - that's information.

Think about how you'd never accept "marketing spend was up" as an answer. You'd want it split by campaign, by channel, by what it returned. AI spend deserves the same treatment, and the only way to get it is to insist your spend can be traced to something before the spike happens, not explained away after. If a usage jump shows up next month and nobody can say what caused it, that's the gap to close first.

## Treat the cap as a conversation, not a wall

Everyone with access to the spend should be working against some kind of ceiling - not to stop them, but to force a check-in. When you approach the line, the conversation is simple: what did we spend, what came of it, and is it paying off? Did we learn something we wrote down, ship something real, or avoid a future cost? Or did we just stay busy?

And the bigger the ask, the more that conversation should expect. A small experiment earns a light touch. A large or growing line of spend should come with a sharper account of what it's for and where it goes next. That's not bureaucracy. That's the same instinct you'd apply to any rising cost you actually cared about.

Asking "how many tokens will this take?" is fair, by the way. It's hard to answer early - genuinely, the variables are real - so give grace at the start. But like any estimate made under uncertainty, it should get tighter as people gain reps. "We don't know yet" is a fair answer in month one. It shouldn't still be the answer in month six.

## A quick aside for my technical folks

If you're the one actually building on this, there's also a lot of cost management in how the work is executed, and that's a discipline of its own. A few of the levers, at treetop height:

- Be deliberate about context. Feed the model what's high-value up front instead of everything you've got, and watch for the point where more context stops earning its keep.
- Reuse instead of regenerate. Cache what you can, and stop paying to recompute the same thing or re-read the same giant file on every pass.
- Match the model to the job. The expensive one isn't always the right call, and plenty of work runs fine on something smaller and faster.
- Close the loop. When you learn a certain step is a token hog, write it down, so the team isn't paying tuition on the same lesson twice.

None of that is exhaustive, and none of it is news to anyone who's been in the weeds. Either way, efficiency is a skill, not an accident. You already know not to waste disk, bandwidth, CPU, or memory - point the same tradecraft at tokens.

## Both hands on the dial

So that's the job. Decide what a dollar is actually for before you spend it. Insist you can see where it went. Make the ceiling a conversation instead of a wall. And build the work to be lean while you're at it.

None of that requires you to become technical. It asks you to point the discipline you already have - the way you run every other budget that could get away from you - at this new thing that's been getting a strange exemption. A CEO doesn't have to read the code to ask the CTO whether the basics are handled, the same way they don't have to be an accountant to expect clean books. Both sides have a hand on the dial.

So loosen the reins. Spend on the experiments. Let people chase the weird, promising stuff. Then do the unglamorous work that turns all that motion into something you can stand behind when someone asks what it bought.

Not because the spending was wrong. Because the awesome is real - and the only way to keep chasing it is to be able to show what it's giving back.