Something strange is happening inside enterprise AI budgets in 2026.
The price of a token has collapsed. Stanford’s AI Index puts the drop at roughly 280 times over two years. On paper, this should be a golden era of savings.
Instead, finance leaders are walking into the second quarter with the same sentence on repeat: we are already over budget.
Both things are true at once. The cost of intelligence is falling. The cost of using intelligence is rising. If your AI strategy cannot explain that gap, it cannot control it.
The price of intelligence fell. Your bill did not.
Ramp’s enterprise data shows business AI spend growing about four times year over year. The FinOps Foundation now names AI and data platforms the fastest growing new category of enterprise spend. Some large enterprises are reporting monthly inference bills in the tens of millions.
The clearest public example is Uber. Reporting on its engineering rollout described Claude Code adoption jumping from about a third of the org to more than four fifths in a single quarter. By April, the annual AI budget was gone. Per engineer, monthly API costs were running from several hundred to a couple thousand dollars.
Uber is not a story about reckless adoption. It is a story about what happens when a useful tool meets a pricing model that scales with success. The more value people found, the faster the meter ran.
In June 2026, Sam Altman told CNBC that questions about whether AI spending will produce returns are the most fair criticism of AI right now. He said cost went from a topic that never came up to one of the most common concerns he hears, in a matter of months. When the person selling the product agrees the bill deserves scrutiny, the bill deserves scrutiny.
Sending every task to the best model is the expensive mistake
There is a quiet assumption buried in most AI deployments. It says that if a frontier model is the most capable, it should handle everything.
That assumption has a name now. Analysts call it the Big Model Fallacy, and it is one of the most expensive architectural habits in enterprise AI.
Most enterprise work is not frontier work. Summarizing a ticket, classifying an email, extracting a field, formatting a response. These are routine tasks, and routine tasks do not need the most powerful reasoning model on the market. They need a model that is good enough, fast, and cheap.
When every request goes to the premium path by default, you are not buying capability. You are buying overhead. Industry write-ups on model routing describe simple queries making up around 80 percent of enterprise traffic. Paying premium rates for that 80 percent is the leak.
The problem is not unit cost. It is unit count.
Here is the part that breaks old budgeting instincts.
For a chatbot, one question meant one model call. That math was easy to forecast. Agents broke it. An agentic workflow can reason in loops, call tools, check its own output, and correct itself. One task can trigger ten or twenty model calls before it finishes.
By most 2026 estimates, inference now accounts for around 85 percent of enterprise AI budgets, and agentic workflows consume many times more tokens per task than a single chatbot query. Always-on monitoring agents add more, scanning logs and inboxes even when no human is watching.
So the unit price falls, and the unit count explodes. The savings on each token get buried under the volume of tokens. This is why roughly 80 to 85 percent of enterprises reportedly miss their AI forecasts by more than a quarter. They budgeted for a price when they should have budgeted for a pattern.
Cost control is a routing decision
You do not fix a routing problem with a discount. You fix it by deciding, on purpose, where each piece of work should go.
That is what cost-aware routing means. It is directing AI work based on the value of the task, not the habit of the tool. Route simple, high-volume work to lower-cost models. Reserve premium reasoning for the work that actually needs it. Keep some work local or private when that is the right call.
The results are not theoretical. One team, after auditing its token usage and moving simpler subtasks to cheaper models, cut monthly API costs from around 40,000 dollars to 24,000 with no product changes. Broader routing implementations have reported inference cost reductions of up to 85 percent while holding output quality steady.
None of this is anti-frontier. Frontier models are worth their price on the work that justifies it. The point is choice. A premium model should be a decision you make, not a default you inherit.
What the board actually wants to see
The 2026 board conversation has moved past token charts. Leadership does not want to see total token spend. It wants to see whether the spend produced anything.
That shift shows up as efficiency ratios. Cost per resolved ticket instead of total tokens. The compute cost of an AI agent measured against the human hours it replaces. Revenue per workflow measured against the inference it consumed.
This is a healthier way to think, and it depends on one thing you may not have yet: visibility. You cannot route by value if you cannot see which tasks, teams, and workflows are driving the cost. Token budgets, usage visibility, and per-workflow attribution are becoming standard management tools for exactly this reason.
What to control before the next budget review
You do not need a rebuild to start. You need a few honest questions.
- Which workflows send routine work to premium models by default?
- What share of your traffic is simple enough for a cheaper model?
- Can you see cost broken down by task, team, and workflow, or only as one line item?
- If a provider raised prices tomorrow, could you move work to a cheaper path without rewriting everything?
The last question matters most. Much of today’s API pricing is held down by venture funding and hyperscaler subsidies. That will not last forever. The enterprises that stay in control are the ones that can move work by cost, capability, privacy, and policy before the price changes for them.
The future of AI cost will not belong to the companies with the biggest budgets. It will belong to the companies that know which model to use, when to use it, and how to route the rest.
Use the right AI for the right work.
Explore ThinkFreely Talk to Buildtelligence
Think Freely.