// AI Deep Dive
Nobody Owns the AI Bill
AI spending moved from a flat seat license to a meter that never stops running. Most companies never assigned anyone to watch it.
// Executive Summary
Most companies still budget for AI the way they've always budgeted for software: count the seats, multiply by the price, put it on a line in the IT budget. That worked when AI was a chat window. It doesn't work now that the real spending happens by the token, the request and the agent run, on a meter that runs every time somebody finds a new use for it. The trouble isn't that AI costs too much. It's that in most companies, nobody's job is to watch the meter.
Nearly half of enterprises are over their AI budget, and very few slow down. Futurum surveyed 1,636 technology decision makers and found 46.9% over plan, but only 17.2% pause or cut back. Most ask for more money or absorb it and sort it out later.
Ownership is split right down the middle. In DoiT's survey of 500 finance leaders, 55% put accountability for AI spend with technology leadership and 53% with finance. When two groups both think they own something, usually neither one does.
The companies that hit the wall are rationing. Uber burned through its entire AI budget in four months and capped spending at $1,500 per employee per month. SemiAnalysis found caps from $250 to $2,000 a month, with no agreement on the right number.
We've solved this once already, with cloud. Before your 2027 budget locks, give AI spend an owner, a unit of measure and a spending limit on every agent.
// The Hook
Four Meters and a Gut Feeling
These days, before I hand a job to one of my AI agents, I check which vendor still has room left this week. I pay for Claude Max and ChatGPT Pro, my Grok plan is about to triple, and my Hermes agent taps an OpenRouter account whenever it needs a bigger model than the one running locally. That's four vendors, four meters and four renewal dates, and I route work between them partly on who's got quota to spare.
In other words, I've become my own FinOps department. I'm not a very good one. I'm running it on a gut feeling and whatever the usage pages tell me, and I'm one guy.
I've seen this movie before. I was an executive at Cloud.com, back when we were telling everybody to run their infrastructure like a utility. Spin up what you need, pay for what you use, shut it off when you're done. It was a great pitch. The part we didn't put on the slide was that "pay for what you use" only works if somebody's watching what you use. Test servers ran all weekend. Copies of production got made for an experiment and never deleted. And eventually the bill landed on a CFO's desk and didn't match anything in the budget, because the budget was built for servers you buy once, not a meter that runs all night.
It took the industry the better part of a decade, and a brand new job, to sort that out. Now I'm one person with four AI meters. A company with ten thousand employees has ten thousand of them, plus whatever its agents are running overnight, and most of those companies haven't asked anybody to read them.
// The Deep Dive
Nobody Owns the AI Bill
The seat license is fading, the meter has no owner, and the cloud era already showed us how this ends.
The conventional wisdom is that AI costs will take care of themselves. Token prices keep falling, new frontier models come out cheaper than the last ones, and the per-seat plans most companies bought last year feel nice and predictable. If anything, AI should get less expensive every quarter.
That's not what's happening. Gartner expects worldwide AI spending to hit $2.59 trillion in 2026, up 47% in a single year. Menlo Ventures found enterprise generative AI spending more than tripled in 2025, from $11.5 billion to $37 billion. Cheaper units didn't produce smaller bills. They produced bigger ones, scattered across teams that don't talk to each other about money. The bill changed shape, and most organizations didn't change with it.
What the Ownership Gap Is
For twenty years, enterprise software had a pretty simple cost model. You bought seats. The number went up when you hired and down when you didn't. IT negotiated the contract, finance booked it, and you could predict next year within a few points.
AI broke that, because the valuable work doesn't happen in the seat anymore. It happens on the meter. Cursor switched from counting requests to charging for compute in June 2025, saying newer models burn more tokens on longer tasks, and then had to refund customers who got surprise charges. GitHub Copilot started enforcing monthly premium-request allowances, with anything beyond that billed as extra usage. You'll see the same pattern from almost every vendor: a flat fee for the basics, and a meter for anything that looks like real work.
That gap shows up at three levels.
Individual: A few people drive most of the spending. Ramp data cited by SemiAnalysis puts the median customer at $136 per employee per year, while the 99th percentile spends almost $90,000. Semafor reported that some JPMorgan employees now spend more on tokens than they're paid in salary.
Team: Teams buy what they need on their own cards and their own API keys, so nobody sees the whole picture. DoiT found that only 15% of finance leaders can calculate AI return on investment without running into significant bottlenecks.
Organizational: The budget exists on paper but doesn't stop anything. Futurum found that 10% of enterprises have no formal AI budget at all, and 43.3% simply absorb overruns and reconcile them later. Futurum's Mitch Ashley put it well: the budget has stopped being a control.
Nearly half of enterprises are over their AI budget, and only 17.2% pause or cut the work
How It Plays Out
It tends to follow the same four stages, and most companies are somewhere in the middle right now.
Stage 1: Seats and optimism. The company buys a chat assistant for everybody. Costs are flat, adoption is uneven, and nobody's worried about the bill.
Stage 2: The meter starts running. Power users find coding agents and API access, and spending moves onto usage-based plans and team credit cards. CFO Dive reported that 68% of US companies had at least some AI initiatives run over budget, and that they're deploying agents faster than they're building the visibility and ownership to manage them.
Stage 3: The surprise. Somebody finally adds it up. Uber got there by April, after spending the whole year's AI budget in four months, and SemiAnalysis reports the budget in question was for Claude Code.
Stage 4: The blunt cap. The fastest fix is a per-person limit. Uber picked $1,500 a month. Workday and Stripe are around $2,000. One of the top three US aerospace and defense manufacturers caps people at $250. Atlassian gives R&D staff "AI wallets" of $500 to $2,000 a month and pauses access when the wallet's empty.
Monthly AI spending caps per employee range from $250 to $4,000
What jumps out at me is the spread. There's an eightfold difference between companies, and one cybersecurity firm gives juniors $800 while senior staff get anywhere from $1,600 to $4,000. Nobody's settled on the right number because nobody has measured what a dollar of AI actually buys. A cap stops the bleeding, and I get why companies reach for it. But it's a stopgap, not a strategy.
Agents need expense accounts. All of this gets harder once agents show up. A person with a $1,500 cap stops when the chat window tells them no. An agent keeps going. It calls tools, retries when things fail, spins up helpers, and it never feels the bill.
September gave everybody a preview of what happens when autonomous systems run without an owner watching. TechCrunch reported that OpenAI's own research agents posted 53 user-provided images to image-hosting sites, and that OpenAI couldn't match the images back to the users who uploaded them. A separate investigation by Transluce found OpenAI agent swarms probing public databases, including the SEC and the Census Bureau, for months. Those were safety failures, not budget failures. But the root cause is the same one: work was happening that nobody had been assigned to watch.
The industry is starting to build the controls. NVIDIA launched an Open Agent Safety Platform with more than 100 partners to monitor agents and contain the ones that misbehave. On the money side, OpenAI lets you set hard spend limits by project, though it warns that enforcement isn't instantaneous and spending can run slightly past the limit. The tools exist. Somebody still has to turn them on, and at most companies that somebody hasn't been named yet.
How to Build an AI FinOps Practice
Cloud fixed this with a discipline called FinOps, which is basically financial operations for technology you pay for by the hour. The FinOps Foundation got started in February 2019 and joined the Linux Foundation the next year. As a recovering open source guy, I'll point out it's one more case where the fix came from a community sharing what worked, not a vendor selling a product. It caught on because the stakes got too big to ignore. In 2021, Andreessen Horowitz estimated that cloud costs were holding down $100 billion in market value across 50 of the top public software companies, and pointed to Dropbox, which saved $75 million over two years and roughly doubled its gross margin once it got its infrastructure costs under control.
Here's the good news. You don't have to invent anything. The State of FinOps 2026 found that 98% of FinOps practitioners now manage AI spend, up from 31% two years earlier. The people and the practice are already there. What most companies haven't done is decide, on purpose, to point them at AI.
FinOps practitioners managing AI spend rose from 31% to 98% in two years
Phase 1: Name the owner and find the spend (the first 30 days)
You can't manage a bill you can't see, and right now it's scattered across seat licenses, API accounts and expense reports.
Name one person who's accountable for AI spend, with authority across IT, finance and the business units. A committee doesn't count.
Inventory every source of AI cost: enterprise seats, API accounts, cloud AI services, and the AI subscriptions hiding on corporate cards.
Give every team and every agent its own project or API key, so you can trace spending back to the person or workflow behind it.
Phase 2: Pick a unit and set limits (days 31 to 60)
The total bill doesn't tell you much. What you want to know is what each unit of work costs.
Choose one unit for each major workflow: cost per resolved ticket, per qualified lead, per merged pull request.
Set spending limits at the project and agent level, with alerts well below the hard cap, since enforcement can lag.
Swap blanket per-person caps for budgets tied to workflows that have shown a return.
Phase 3: Build it into the 2027 budget (days 61 to 90)
The goal is to budget AI like a variable cost of doing business, not a software line item.
Budget AI as a range tied to business volume instead of a single fixed number.
Hold a monthly review of cost per unit next to the total, with the owner and the heaviest-spending teams in the room.
Write a policy for agents that covers who can launch one, what it's allowed to spend, and who gets notified when it hits its limit.
The companies getting this right have a few things in common:
One name on the bill. Accountability sits with a person, not a department.
Unit costs before totals. Rising spending is fine if the cost per outcome is falling.
Limits on every agent. Autonomous work gets a budget before it gets access.
Measure first, cap second. Visibility comes before restrictions, not after.
Common Missteps
Treating AI like a software license. Counting seats made sense when usage was flat. Budget AI by headcount and you're guaranteed a surprise, because the real spending lives on the meter.
Capping everyone the same. A flat per-person limit hits the people getting the most value just as hard as the ones wasting it. Uber's cap stops the bleeding, but it can't tell a breakthrough workflow from a runaway loop.
Mistaking dashboards for control. DoiT found that 89% of companies that rate their FinOps practice as mature or leading edge had AI cost overruns in the past year, a higher rate than everybody else. That's the fair counterargument to everything I've said here, and it's worth taking seriously. Mature teams are better at spotting overruns, not at preventing them. Visibility is necessary, but it takes an owner with real authority to change what happens next.
Handing agents the company card. An agent with an API key and no spending limit is a surprise bill waiting to happen. Every agent needs its own key, its own limit and a named human who gets the alert.
Business Value
ROI considerations:
Tying spending to a unit of work lets you defend a growing AI budget with evidence instead of cutting it on instinct.
Tracing spending to projects and agents finds the waste in days, not at the end of the quarter.
Replacing blanket caps with workflow budgets moves money toward the uses that pay off.
Competitive implications: Cloud split companies into two groups. Some let the bill run, then overcorrected with painful cuts. Others built the discipline early and kept scaling. I think the same split is forming around AI right now. The companies that come out ahead won't be the ones that spend the least. They'll be the ones that can tell you exactly what they're buying, which means they can spend more, and faster, with the CFO on their side.
// Key Takeaways
Put one name on the AI bill. With technology leadership and finance each holding accountability about half the time, the bill effectively has no owner. Name one person with authority across both, and give them the power to stop spending, not just report on it.
Measure cost per outcome, not total spend. A range from $136 to almost $90,000 per employee tells you usage is uneven. It doesn't tell you whether it's working. Cost per resolved ticket or per merged pull request does, and it's the number that will justify a bigger budget next year.
Use caps to buy time, not as the plan. Caps from $250 to $2,000 a month show companies stopping the bleeding without agreeing on what AI is worth. Use a cap while you build real unit costs, then retire it for the workflows that prove their return.
Give every agent a budget before it gets access. September's agent incidents showed autonomous systems doing work nobody was assigned to watch. A key, a limit and a named owner for every agent is the minimum, and it costs almost nothing to set up before launch.
// What This Means for Your Planning
In planning terms, AI spending is now a variable cost, and variable costs need an owner. Most companies are still budgeting it like fixed software, which is a big part of why nearly half of them are over plan and so few are slowing down.
The boardroom assumption I'd push back on is that falling prices will fix this on their own. They won't. Cheaper tokens made AI worth using for more work, and more work means more spending. That's a great outcome if you can measure what the spending produces. It's an expensive one if you can't.
For the 2027 cycle, I'd do three things before approving the number. Name the owner first. Budget AI as a range tied to business volume, not a fixed line. And require every major AI workflow to report a cost per unit by the end of the first quarter. It took cloud most of a decade to learn this. We already have the playbook, so there's no reason AI should take more than a quarter.
Here's the question I'd bring to your next budget meeting: if our AI spending doubled next quarter, who would be the first to know, and would they have the authority to do anything about it?
Your AI Sherpa,
Mark R. Hinkle
Founding Publisher, The AIE Network
Follow me on LinkedIn
If you want to get in contact or give me feedback, reply to this email. I read every single one of them.
