Getting maximum out of your AI budget

A few months into running AI development across multiple project teams, I started looking at the actual spend per developer per month. The variance was enormous, and not in the direction I expected. The most productive people weren't the biggest spenders. Often the opposite.

Here is what I now tell every tech lead about budget.

Spend on capability, not volume

It's tempting to "save money" by routing every request through the cheapest model available. This almost always costs more in the end, because the cheap model produces code that needs three retries and a senior review, while the strong model gets it right on the first attempt and lets the developer move on.

The real metric is cost per merged PR, not cost per token. Track that and you'll quickly stop optimising the wrong thing.

Pay for the agent, not the chat

A developer who uses AI through a chat window: paste, copy back, repeat, is doing roughly 10% of the workflow the tool is capable of. The bigger wins come from agentic loops that can read files, run tests, and iterate. That's where the budget should go.

Cache the boring parts

Prompt caching is the single biggest budget lever most teams ignore. If your repo's rules file, conventions doc, and a few canonical examples are part of every prompt, cache them. We've seen 60-70% reductions on repeated workflows just from getting the cache strategy right. The model providers have done the engineering. You just have to structure your prompts to take advantage of it.

Set a budget per developer, not per team

A team-wide budget gets eaten by whichever loud person is running the most experiments. A per-developer cap forces everyone to learn the cost of what they're doing. It also gives quiet people permission to actually use the budget instead of saving it "just in case."

Reasonable starting points, in my experience:

• Active builders: give them a real budget. Cap it high enough that they don't have to think about it during normal work.
• Light users: give them the same budget anyway. Most light users are light because they don't know what's possible, not because they don't need it.
• Reviewers / leads: separate, smaller budget for code review and triage flows. These are cheaper but high-leverage.

Audit monthly, ruthlessly

Once a month, look at the top spenders and the top zero-spenders. Talk to both. The spenders usually have one workflow that's broken and burning tokens. Fix that and spend halves. The zero-spenders usually have no idea what to do. Ten minutes of pairing and they double their throughput.

The thing nobody likes to hear

Your AI budget is downstream of the quality of your codebase. A clean repo with strong conventions costs a third of what a messy one does, for the same work.

If you're spending too much, the first instinct is to negotiate a better deal with the vendor. The second, more profitable instinct is to spend a week cleaning the parts of the codebase the AI touches most. The token bill comes down on its own.