Mark Tauschek, VP, Distinguished Analyst & Research Fellow, Info-Tech Research Group.gettyWhat happens when the CFO walks in and asks, "What did we get back for the few million dollars we spent on AI last year?" Let's assume somebody pulls up a dashboard showing that adoption is climbing and ticket volumes are down. However, the CFO asks the question again because none of that was an answer. They wanted a dollar figure, and nobody has one.We're hearing this type of scenario playing out in boardrooms around the world, and the response is always the same. Organizations go hunting for a better formula (another ROI calculator, another framework, another business case template), but it rarely helps because the formula was never the problem; accountability is the problem. What most organizations are missing is one person with the authority and the incentive to say what the AI's returning, and that's an organizational failure. Appointing an accountable owner is harder than buying another tool.You can't measure savings you never measured before.Most AI projects get pointed at knowledge work—the drafting, summarizing and ticket triage that nobody ever timed. Nobody clocked how long it took to review a contract or clear a Tier 1 ticket before the tools arrived, so when a vendor tells you its product saves four hours a week per user, you have nothing to check it against. A July 2025 MIT research report found that 95% of enterprise GenAI projects showed no measurable ROI. That gets quoted as proof the technology is a bust, but I see it differently. Most of those projects launched without ever deciding, in hard dollars, what success looked like.Your vendor's ROI calculator is a sales tool.The vendor earns more when your usage climbs, but no one there is measuring what it did for you. Multiplying hours saved by a loaded hourly rate can get you a big figure for the slide, but saved hours only become saved dollars if you cut headcount, move people to measurably higher value work or improve throughput to something revenue generating. If you don't do any of that and quantify it, the figure is just slide candy.Somebody needs to own the numbers on both sides.AI is getting plenty of attention now as CEOs and CFOs sit in these conversations, unlike a couple of years ago. The operational metrics, adoption and tickets deflected stats sit with IT or the business unit, while finance owns the financial side. However, whoever could hold both and say this cost us X and returned Y is rarely on the org chart.According to a 2025 IBM survey of 2,000 CEOs, only a quarter of AI initiatives had delivered the expected return over the previous three years. One IBM consultant said he has yet to meet a client running fewer than 60 random acts of AI, none of them measured the same way.Now the meter is running.All of that assumed stable costs in the form of per-seat licenses, a line item you could forecast, but that era is over as usage-based pricing has become the norm. The average enterprise AI spend tripled from 2024 to 2026, even as the price of a token fell by 98%. Those should cancel out, but they don't because agentic workloads devour far more compute than the chatbots they replace. A May 2026 Goldman Sachs report projected agentic AI to push token consumption up 24-fold by 2030. OpenAI CEO Sam Altman has said that he expects intelligence to be a metered utility, similar to electricity or water.What happens when the meter is turned on in an organization that couldn't prove ROI on a flat subscription? The flat rate pricing is gone, and it now runs in real time across every team and agent with almost no one watching the dial. Uber gave us a demonstration of the result in 2026 when, after ranking engineering teams on a usage leaderboard (higher usage was better), it burned through its annual AI budget by April. Heavy users were running $500 to $2,000 a month in tokens, and its president admitted he couldn't connect the spending to features customers noticed. Uber now caps spend at $1,500 a month per tool. It encouraged people to consume but never tied the consumption to an outcome, and it's now rationing.What do the winners do differently?The companies seeing real returns didn't get there with a fancier spreadsheet. A couple we've talked to measure simple baselines as a start. What they share is discipline:1. One named person, not a committee, owns the tie between the operational and financial numbers, with authority to shut down what's not paying off.2. They capture the before state, cycle time, error rate and cost per task and treat it as a gate. The initiative doesn't launch without the baseline.3. They build cost-per-task visibility now before the variable bill ambushes them the way early cloud bills did.4. They write contracts with vendors around outcomes they can verify, not seats or tokens.The discipline must be sustained, though. Some of what AI delivers is real and slow, such as capability that compounds or the bad outcome that never happened. Demand a dollar figure this quarter for everything, and you mix the patient bets with the vanity ones. The owner’s actual job isn't killing whatever can't be measured by Friday; it's telling a slow investment apart from expensive noise and defending it.Here's the bottom line.Investors have lost patience while finance starts paying attention, and the cost side is now a meter that runs whether or not anyone reads it. Not knowing what your AI returns are was a slow strategic risk, but it now shows up on a quarterly statement. Drop the search for a perfect formula and ask who has both the authority and the reason to tell the truth about what your AI delivers and whether they know what it costs.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?