A New Way to Manage AI Budgets

A client of mine once asked for $50 a month per person for AI tools. Three weeks later, that number was $100. Then it was super users at $250 a month. Then it was super users with unlimited access. The whole thing happened in about two months, and nobody was being reckless. Once people learn how to use AI well, they use it more, and the budget you set at the start of the quarter is already wrong by the time anyone reads the memo.

That’s the problem with AI budgeting right now, and it’s not just my client. Seventy-three percent of enterprises have no rule for who actually owns AI costs. Seventy-nine percent keep increasing spend without proven ROI. Agentic AI budgets are growing about 32% a year, while the traditional IT budgets meant to fund them grow less than 4%. Nobody has figured out how to put a number on this that holds for more than a few weeks.

The Two Approaches Everyone Tries First

Most companies land on one of two options, and I’ve watched both fail.

The first is top-down. Management sets a number, usually a guess dressed up as an estimate, and tells every team what they have to work with. It falls apart because nobody at the top has visibility into what each team actually needs, and the number is stale within weeks.

The second is bottoms-up. You ask every team to estimate its own usage. This fails for the opposite reason: most teams don’t know what they need either, especially early in AI adoption, when usage tends to spike, dip, then climb again as people get more comfortable with the tools.

When companies get burned, they usually try one of two fixes: unrestricted access, which is how Uber burned an entire year’s AI budget in a single quarter, or a flat cap, which stops the bleeding but kills the reason anyone wanted the tool, and pushes people toward shadow AI: personal accounts run on company work.

Borrowing a Market Mechanism That Already Works

My pitch is a version of a market mechanism that’s been running for over thirty years: cap-and-trade. The EPA used it for factory emissions. Regulators set one hard cap on total emissions allowed, handed out allowances, and let plants trade them. A plant that could cut emissions cheaply did, and sold its surplus to a plant where cutting was expensive. The whole system cost far less than anyone projected, because a price mechanism decided who acted, instead of a uniform mandate telling every plant to hit the same number.

How It Works for AI Budgets

Apply that same structure to AI tokens instead of emissions, and here’s what it looks like.

First, every business unit gets an automatic floor, based on headcount. No bidding, no meeting. If your team is fine on its floor, you do nothing, and that’s most teams, most months.

Second, finance sets aside a float: a smaller pool auctioned off before the period starts. Only teams that want more than their floor bid on it, and it’s a uniform-price auction, meaning every winner pays the same price, the lowest winning bid. Nobody has to guess what everyone else will pay.

Third, trading during the period. A team that won float it doesn’t need can resell the unused part to a team that’s now short. This only works if you close the loophole: resale is capped at cost, and a team has to actually use a real share of what it won before reselling the rest. Otherwise the whole thing turns into a trading game instead of a capacity tool.

The clearing price becomes your forecasting tool. Watch what the float sells for across a few cycles, and use that, not a guess, to size the next period’s float and check whether the floor itself needs to move.

One Cap, Two Decisions

Here’s the part that actually sold me on this. Top-down fails because leadership sets a number and also tries to dictate how every team uses it. Bottoms-up fails because there’s no honesty mechanism forcing team estimates to mean anything.

Floor plus float doesn’t force you to choose between those two. It splits the decision and hands each half to whoever is actually positioned to make it. Leadership sets exactly one number, the total cap, because only leadership sees the whole budget. Everything below that line, how much float to bid for, when to resell, how to get the most out of what you’ve got, is a team decision, because only the team knows what its own AI use is worth. The auction keeps that second decision honest.

Key Takeaways

• One cap, two decisions, split correctly. Leadership sets the enterprise total. Teams decide how to use their share of it.
• Floor plus float beats guess-and-hope. Every team gets a guaranteed, headcount-based floor. Only the marginal need goes through an auction.
• The uniform-price auction rewards honest bidding, since every winner pays the same clearing price.
• Two rules keep trading from becoming a strategy: cap resale at cost, and require real drawdown before resale.
• The float’s clearing price is a real forecasting signal for next period’s float, and for the floor rate itself.

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: A New Way to Manage AI Budgets