Enterprise AI spending tripled in one year. That alone would be news. The more important fact is why costs are outpacing expectations even as the price of running AI has collapsed.
The cost per unit of AI output has fallen roughly 99.7% since the early days of large language models. Vendors have been competing hard on price. Yet according to data from the FinOps Foundation, 73% of enterprises still report their AI costs exceeded original projections. The bills are going up because usage is growing much faster than unit prices are falling.
There is a name for this pattern. In 1865, economist William Stanley Jevons observed that as coal engines became more efficient, total coal consumption rose rather than fell. The same thing is happening with AI. Cheaper access means companies run more tasks through it, including ones that were never in the original plan.
A specific version of this catches many companies off guard. Most AI tools today do not charge a flat monthly fee. They charge based on how much you use: every word sent to the AI, every word it produces back, every step an automated process takes. An automated workflow does not just answer one question and stop. It reads files, forms a plan, executes a step, checks its own output, revises, and loops. Each of those steps is a separate charge. Enterprise data shows that roughly 73% of AI usage was being routed to the most expensive model tiers, even for basic tasks a cheaper tool could handle equally well.
Uber reportedly burned through its entire planned AI budget for 2026 within the first few months of the year. That story circulated because it is not unusual. Nearly 8 in 10 IT leaders report being hit with unexpected charges tied to how AI gets billed.
The good news is that the levers are straightforward, even if they require discipline to apply.
The first is matching the tool to the task. Premium AI models cost 20 to 50 times more per unit of use than smaller, capable alternatives. Using a top-tier model to answer a routine internal FAQ is the equivalent of hiring a specialist consultant to file basic paperwork. Companies that have started routing routine tasks to cheaper models and reserving expensive ones for genuinely complex work are seeing meaningful reductions in their bills.
The second is setting hard limits before costs accumulate, not after. This means negotiating spending caps into vendor contracts, building automatic brakes into AI workflows that trigger when a department hits its monthly limit, and treating AI spend like any other budget line with an owner accountable for it. Many vendors will negotiate token rollover clauses, meaning unused capacity from one period carries forward rather than disappearing. That reduces waste on both sides.
The third is scrutinising what you are actually paying for. Many software vendors have added AI into their existing product bundles and raised prices as a result. Some researchers are calling it an "AI tax." AI software prices across vendor categories have risen 20 to 37% in some cases, regardless of whether the underlying product improved meaningfully. A contract review that includes someone with AI knowledge, not just a procurement generalist, will catch charges that a standard renewal process misses.
The broader shift to watch is outcome-based pricing. Instead of paying per word processed or per step taken, some vendors are beginning to charge based on what the AI actually delivered: how many documents were completed, how many customer queries were resolved. Adobe has moved its enterprise AI suite to this model. It removes the incentive to overconsume and aligns the vendor's interest with the buyer's. That alignment is rare in consumption-based contracts and worth negotiating for where possible.
For any organisation now moving AI from pilot to production, the cost model changes significantly at that transition. A pilot running on a small team gives no reliable signal about what full deployment will cost. The architecture of the workflow, which model handles which task, and whether spending limits are in place before launch will determine whether AI becomes a controlled investment or an unpredictable liability.