gettyImagine a company that tracks every dollar of payroll to the cent but has never once run a performance review — leadership knows precisely what each employee costs but nothing about who's actually creating value. That's the position many enterprises are in with AI today."The CFO can see the bill, and the bill is real information," says Florian Douetteau, CEO of Dataiku. "It tells you nothing, though, about whether the work was worth doing." Token bills, the metered cost of every action an AI model takes, have become a board-level line item. The missing half of the ledger is whether any of it is working, and fewer than one in three organizations can tie AI spend directly to profit and loss outcomes, meaning most companies increasing their AI investment can't say what the increase is buying them.Why Agentic AI Can Burn Budget FasterA chatbot's economics are self-limiting: someone types a prompt, the model answers and nothing more happens until a person decides to ask again. That person, without anyone designing it that way, also served as the budget control."An agent removes that person from the loop," Douetteau says. "It runs by itself, chaining steps across several systems and making a decision at each one, so a single task can trigger dozens of calls before it finishes." Because it keeps going without waiting for anyone, token exposure compounds in a way simple prompting never produced, and non-determinism makes it worse — an agent can set off in a direction no one intended and spend real money walking down a path that leads nowhere. "When one person was typing prompts, a bad call cost you a few cents and a moment of your time," Douetteau says. "When an autonomous system makes the same bad call at three in the morning, across your systems, for a week, the arithmetic is different."Some organizations have burned through AI budgets surprisingly quickly when experimentation goes uncapped.As bills like that land, CFOs get pulled deeper into the budget conversation, and the instinctive response is to implement or tighten spending caps. That's short-sighted, says Douetteau. "A cap treats every token as equal. It cannot tell the agent that is clearing a quarter of your exception cases from the one that has been looping on itself for three weeks and producing nothing." Lower the ceiling, and both slow down together, throttling the work that was paying for itself along with the work that wasn't.Visibility — Not Budget Caps — Is the Answer"The CFO asks how much," Douetteau says. "The CIO has to ask how much, for what, and whether it is working." The CIO is closest to what each agent does, which systems it touches and which process it was meant to improve. But most can't answer that today, he says, because their organizations never built the visibility that would let them.Visibility into agent performance lets leaders redirect investment, spending more where agents are earning their keep and pulling back or retiring those that aren't.Douetteau says many companies today know their agents are active, but have no idea whether they're effective: "Uptime tells you an agent is running, which is not the same as knowing whether it should be."CIOs should begin by defining the business outcome an agent is meant to produce before it's deployed and continually measure it against that goal. That's because agent performance isn't static. An agent worth its cost in January can quietly stop earning it by June as the underlying model updates, the data shifts and the business context moves, all while the agent looks perfectly healthy from a systems standpoint.Consider a loan assessment agent built to triage exceptions and route the hardest cases to a senior reviewer. Three months in, the underlying model updates and the mix of incoming applications shifts. No one touches the agent's instructions, so it keeps running, except it's now flagging different cases than it was designed to catch. "By the time someone notices, it has been working that way for weeks," Douetteau says, "and the damage and the question arrive together. How did we let this run?"That's the moment a governance trail earns its place: a record of who approved the agent, what the business expected it to do, and when anyone last reviewed it. It's also what makes AI ownership real rather than nominal, closing the gap between the executive who claims to own the AI strategy and the decisions made hour by hour within these systems. Dataiku Agent Management is built to hold that record in a single view across agents, platforms, and tools, showing what's hitting its goal, what's drifted, and who signed off on it.The leaders who pull ahead won't be the ones who spent the most carefully. They'll be the ones who ran the performance review and can show the board what the investment actually bought.