Seventy-three percent of enterprises went over their AI budget this year. That number isn't a forecast. It's a post-mortem. The FinOps Foundation found it in their 2026 State of FinOps survey, and if you run anything agentic, you're probably in that group and don't know it yet.
Seventy-three percent. That's the share of enterprises in the FinOps Foundation's 2026 State of FinOps report whose AI costs ran past what they planned. Not 73% who saw a surprise spike. Seventy-three percent who blew the budget outright.
Why now? Two years ago most enterprise AI was a chat box on an intranet. A question went in, an answer came out, the meter barely moved. 2026 is different. The same companies now run agents that plan, call tools, read documents, and decide what to do next. The bill follows the agent, not the chat box.
Three things hit at once, and almost no budget accounts for any of them.
First, token count per task. A plain chatbot query moves maybe a few hundred tokens. An agentic workflow runs 5 to 30 times that per task, according to Stanford's Agentic AI Cost Attribution study from 2025. The model isn't answering one question. It's holding a working session, with notes, retries, and self-talk.
Second, re-sent context. This is the quiet killer. Stanford's lab found that 62% of an agent's inference bill is the model re-reading information it already has. The agent pastes its whole memory back into the prompt on every step, so the model pays to read the same facts twenty times. That isn't a model problem. It's a plumbing problem.
Third, runaway loops. An agent that can't tell it's done will retry until something stops it. I've watched a single stuck task burn a day's token budget in an hour because nothing killed the loop. No ceiling, no kill-switch, no alarm. The meter just ran.
Here's the part that throws people. Token prices dropped. Ramp's data shows the average cost per million tokens fell from about $10 to $2.50 in a single year. So you'd think bills would fall too.
They didn't. The median enterprise LLM bill grew 7.2x year over year, per Gartner. Cheaper per token, but you're buying a thousand times more of them. The unit economics improved and the total bill exploded anyway. Falling prices are not a cost strategy.
A few stories made the rounds, and they aren't edge cases.
Uber's CTO said in April 2026 that the budget he thought he'd need for the whole year was already gone. By April. The annual AI number, spent in four months.
Axios reported a client that spent $500M in one month after handing out Claude licenses with no usage limits. A CTO there said employees were using the models to check the weather. Not to write code. The weather.
Sam Altman said the AI cost ROI question is "the most fair criticism right now of AI." He noted customers had already burned through their entire 2026 AI budget, and that cost went from a topic that never came up to the second most common concern in a short stretch. When the person selling the compute says the bill is the problem, the bill is the problem.
Defaulting every request to one model is expensive in a way most teams never measure. Routing the right task to the right model cuts cost hard. Teams that route instead of defaulting report an 87% premium from staying single-model. That's money left on the table by habit, not by need.
And the people who own this number aren't ready. FinOps data shows 98% of cloud teams now carry AI cost responsibility. Almost none of them had a playbook for inference before this year. You handed the biggest new line item to the team that was budgeting storage and compute, and gave them no tools built for it.
OK, here's what stops it. None of this is exotic.
I run a multi-agent production fleet. Not a demo, not a blog diagram. Real agents doing real work, and I've watched the bill do every stupid thing in this post.
The fix isn't a better model. It's discipline. Ceilings, kill-switches, less context, routing, and a dashboard someone actually looks at. I treat agent cost like I treated uptime at Wells Fargo: a thing you measure every day, or it breaks you.
So we built Agentic Ops around exactly this. Same instinct that ran security scans on AI systems, now pointed at the spend. If your agents are costing more than they should, that's the service.
Related: Prompt Injection: LLM's #1 Risk (2026) · Agentic Ops