Last August AI was eating budgets. Not cloud budgets. Business budgets.
FinOps was already in the room, solving an infrastructure problem. But AI was creating a different problem entirely. One that lived in business outcomes, not compute bills.
The real question was never how much AI cost. It was whether AI was earning anything back.
That question has changed shape four times since.
Naming the problem (last summer)
AI spend was hiding inside other budgets. Engineering bills. Product budgets. Random SaaS line items. It wasn't visible, it wasn't allocated, and nobody owned it.
The numbers were the tell. Global AI spend was on pace to reach $2.5 trillion in 2026, a 44% jump in a year, according to Gartner. Most CFOs say how much of that was running through their own books.
The market's first move was simply to say the thing out loud. This was a real cost, it was growing fast, and you can’t optimize what you can’t see.
Visibility isn't enough (late 2025)
Once people could see the spend, they had to decide what to do with it. Visibility without accountability is just a more detailed problem. A number you stare at without any team accountable for it will just keep growing.
So the conversation moved past dashboards and toward governance. That raised real questions about who’s accountable when the spend overruns and who actually has the authority to stop it. Engineering has been running AI on its own, and bill was about to land in finance.
From cost to economics (this winter)
Here the frame shifted hard. The new question was what AI earned or lost.
Some customers turn into loss leaders. Certain features never turn a profit, and a few workflows lose money on every call.
A support agent that resolves tickets at $0.40 each pays for itself, while a research agent burning $40 a query for answers nobody acts on does the opposite.
Outcomes are financial events as much as engineering events, which widens the set of people who care about AI economics. The CIO had been feeling this for a while. The CFO was starting to, and the board would not be far behind.
Agents change the question (this spring)
By spring, the conversation narrowed. AI in general had stopped being the story, and agents specifically took over. Autonomous. Continuous. Making spending decisions without anyone in the loop. They stopped being a feature and started acting like employees.
The unlimited agents era was ending. Anthropic moved third-party agent tools off flat-rate subscriptions onto metered billing in April. Translation: subscriptions were built for humans, not for agents running all day, all night, all weekend. The math no longer worked.
Not because the technology stopped scaling, but because the business model couldn't. You can't have unlimited workers and finite margins. Something had to give.
Which forced a different question. What exactly are these things?
The Era of the AI Workforce
They're workers. Every enterprise is deploying them. Agents that research, decide, execute, and spend around the clock and at scale with no one watching.
And here's the problem: we have decades of infrastructure for managing human workers. Payroll systems. Performance reviews. Headcount controls. Cost centers. We know what every employee costs and what we expect in return.
We have none of that for AI agents. Not yet.
When an agent runs unchecked over a weekend and hands you a $47,000 bill on Monday morning, that's a management failure. Nothing caught the spend because nothing has been built to watch the agent in the first place.
Every enterprise will end up with an economic control system for its AI workforce. The same way every enterprise has one for its human workforce.
That system will track what every agent costs, what it produces, and it is authorized to spend. It will sit between the AI and the business, enforcing spend limits, attributing cost to outcomes, and controlling authorization.
The CIO mostly wants visibility into what the workforce is doing. The CFO needs control over what it costs. The board needs both.
That's the next chapter of this conversation. Not "can you see your AI spend." Not "what's the ROI on your AI features."
It’s whether your AI workforce is being managed at all.
You already own this problem. Most enterprises just don't have the system to manage it yet.
