Who is watching what your AI agents spend?
Who is watching what your AI agents spend?
Say you run a small team and set up two AI agents to produce a weekly research brief. One gathers sources and writes the draft. The other checks the draft and sends it back when something isn't clear. The first week goes well. The brief arrives on Monday morning, it's useful, and nobody thinks about it again.
The next week, the checking agent isn't satisfied with one of the claims and asks for clarification. The writing agent answers and asks it to confirm. The checker replies with more questions, and the writer answers those too. Each exchange is small and cheap, and each one looks like the system doing its job. Nobody is watching, because it worked last time.
By Monday morning the brief hasn't arrived, and the two agents are still going back and forth over the same claim. Each one is doing what it was built to do. What's missing is anyone whose job is to notice that the work has stopped being worth what it costs.
Cheaper tokens, bigger bills
The price of running a model has fallen fast. Stanford's 2025 AI Index found that the price of using a model that scores as well as GPT-3.5 on a standard benchmark went from $20 per million tokens in November 2022 to $0.07 by October 2024, a more than 280-fold drop.
You'd expect bills to fall with it. For a lot of companies, they haven't. In an IDC survey sponsored by DataRobot, 96% of organizations deploying generative AI and 92% of those implementing agentic AI said their costs were higher or much higher than expected. Seventy-one percent said they had little to no control over where those costs were coming from. For agents in particular, token consumption was named among the top unexpected costs.
Part of the reason is ordinary. When something gets cheaper, people tend to use more of it. With agents, the thing spending the money also keeps going on its own, for as long as it thinks it hasn't finished, and it has no idea what your budget looks like.
That's what happens to the research brief. Each exchange between the two agents costs so little that nobody looks at it, and the brief can make hundreds of them before anyone checks.
The cost of nobody watching
Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, and among the reasons it gives are escalating costs and unclear business value, as Reuters reported.
Those two reasons are connected. If you can't see what a job costs while it runs, you can't judge whether it's worth doing. The research brief might be worth paying for every week. But if the only time anyone looks at its cost is when the invoice arrives, the decision about its value has already been made for you.
You wouldn't hire someone to buy things for the company, hand them a company card and tell them there's no spending policy, nobody to approve a large purchase and no manager checking in. With agents, that's often how it goes. They get a model, a task and access to the tools, and the budget is whatever shows up on the bill.
Alerts need someone reading them
Most teams that think about this start with alerts: send a message when spending passes a certain amount. That helps, but an alert only works if someone reads it in time. The research brief runs overnight. The alert lands in an inbox at 2 a.m., and the two agents keep going.
What changes the outcome is a limit that stops the work, whether or not anyone is awake. If the brief has a limit, the back-and-forth stops when it reaches it. In the morning someone looks at what happened and decides whether to raise the limit, rewrite the instructions that started the back-and-forth, or end the job.
Questions to answer before agents run on their own
Before the research brief runs unattended again, the team needs a few answers.
What is the most one run should cost? Two numbers help: what a normal week costs, and what the brief is worth to you. The limit belongs between them. If a normal week costs a few dollars, a run that costs ten times that has probably gone wrong, and no run should cost more than the brief is worth.
What happens when a run reaches that number? If the answer is an email, ask who reads it and how quickly. If the agents keep working while the email sits unread, nothing has actually stopped.
What is the limit for the month? A job that repeats can stay inside its limit every week and still add up to more than you meant to spend.
Who can change the limit, and what do they need to see first? When the brief stops, someone should be able to see where the money went, fix what caused it and decide whether to start it again.
And who decides whether the job is still worth doing at all? A limit protects you from a bad week. It can't tell you whether the brief is still useful to the people reading it.
Money is part of managing a workforce
Most of the conversation about agents is about what they can do. Can an agent research, write, check its own work, use a tool? Even when the answer is yes, a group of capable agents doesn't become a workforce until someone can direct it, review the work and decide what it's allowed to spend.
Supanova is an autonomous workforce. Say the research brief ran on Supanova atoms. When the back-and-forth reaches a spending limit you've set, the work pauses instead of failing, and it can resume once the limit resets or you raise it. A bad week stops at the limit you chose.
If you never set a spend cap, you don't have one. Your wallet is the limit. If the money runs out, the work waits and your Chief of Staff tells you exactly what it needs. It asks once. Nobody has to catch an alert at 2 a.m. for the spending to stop.
This matters most to the people with the least room for a surprise. A small team like the one running the research brief can't absorb a bill nobody saw coming. We built Supanova so people could turn ideas into economic value without waiting for capital, headcount or permission. For a team that size, that means trying something new for an amount they've chosen, and deciding from what comes back whether it deserves more.
You already know what your agents can do. Before the next job runs overnight, find out who is watching what it spends, and what happens when nobody is.
Sources
- The 2025 AI Index Report, Stanford Institute for Human-Centered Artificial Intelligence (April 2025)
- The 2025 AI Index Report: the state of AI in 10 charts, Stanford HAI (April 7, 2025)
- The hidden AI tax: IDC research reveals nearly all organizations lose cost control when deploying GenAI and agentic workflows at scale, DataRobot press release on an IDC survey it sponsored (December 9, 2025)
- Over 40% of agentic AI projects will be scrapped by 2027, Gartner says, Reuters via The Star (June 25, 2025)