Somewhere in almost every large company right now, a version of the same meeting is taking place. Two years ago, chief executives mandated the widespread adoption of artificial intelligence. Therefore, budgets were found, licences bought, pilots launched, and agents were placed in every workflow that would allow it. Now, chief financial officers are being asked what the returns are, and there isn’t a clear answer.
This is not a failure of technology, but a failure of measurement, and it is nearly universal. McKinsey’s 2026 State of AI survey found that while eight in ten respondents said AI was improving their individual productivity, only 6 percent reported significant financial value from AI.
The instinctive reading is that the tools don’t work. The more uncomfortable reading, and the correct one, is that almost nobody deploying enterprise AI actually knows whether it works, because the way returns are being measured cannot tell them.
To say a tool delivered a return is to assert that outcomes with it are better than outcomes would have been without it. That is a statement about a world that didn’t happen, e.g. a counterfactual, and establishing it takes the same discipline a bank already applies to credit models and capital estimates. Instead, AI returns are typically asserted from a productivity heuristic, measured over a few weeks, at a single desk, with no control group, and credited entirely to the tool. Together they produce numbers that feel authoritative but mean almost nothing.
Agentic AI systems don’t just draft but act, executing multi-step work with limited supervision, makes this both more compelling and more dangerous. The productivity story is more dramatic, the spend is larger and grows in ways budgets don’t predict. A system that acts on its own generates real risk exposure alongside its benefit. If you are going to hand work to agents at scale, you had better be able to measure what they return.
Hours saved is not money made
Start with the metric almost every AI business case rests on: hours saved, multiplied by loaded salary. It is a manufacturing concept measuring output over input incorrectly applied to knowledge work where it does not hold. An analyst who finishes in two hours what used to take four has not created value if the work is weaker, if the errors surface later, or if the freed time is redeployed into unproductive activities.
That last point is where most of the phantom returns hide. Freed time is not saved money.
The discipline that fixes it is unglamorous and decisive. Separate the effect (the measured change, in minutes or error rates) from the benefit (the financial consequence, once capacity has actually been redeployed) from the value (benefit net of every cost the deployment created). Report effects as though they were value and you will overstate returns every time.
The single question that collapses most inflated cases is the plainest one a CFO can ask: which specific mechanism converts this saved time into money, and when will we audit whether it did?
The counterfactual you don’t have
Suppose you measure the right thing. You still have to prove the tool caused the change, and here the standard practice of tracking a team’s performance before and after roll-out, and then attributing the movement to the AI, is worthless. In any live business, everything moves at once.
How badly intuition misleads here was captured by a striking 2025 experiment. The research group METR ran a randomised trial with experienced software developers using modern AI coding assistants. The developers expected the tools to speed them up by roughly a fifth, and afterwards believed they had. In fact they were about nineteen percent slower. Without the control group, every one of them would have reported a productivity gain in good faith, and every one would have been wrong.
The solution is straightforward. It follows the same principle used in drug trials for decades: give the tool to one randomly selected group, compare the results with a similar group that does not use it, and measure the difference. It’s harder in an office than a laboratory because treatment and control sit side by side and swap tips, everyone knows they’re being watched, and AI tools have a learning curve that short trials mistake for weakness. These are reasons to design the experiment well, not to skip it.
The balloon problem
Even if AI improves one part of a workflow, that does not mean the business as a whole benefits. Knowledge work is a chain, as one person’s output becomes someone else’s input. If AI speeds up junior drafting, for example, it may simply create more work for the senior reviewer, who was the real bottleneck all along. The savings made at one stage can quickly reappear as a higher cost somewhere else.
Agentic systems intensify this precisely because they are fast: automate a whole sub-process brilliantly, and you can generate more work for the human checkpoint that remains, not less. The value of an agent is a property of the whole chain in its specific setting, not of the tool in isolation, and a number measured at the point of automation is an upper bound masquerading as an estimate.
The bill you didn’t count
Then there is the cost side, where counting the licence fee and stopping is the norm. For agents the omissions are large and systematic. Usage-priced inference scales with how hard the agent works, not how many seats you bought, so per-seat budgets understate the true run-rate of anything genuinely autonomous. Every agent needs oversight, quality assurance, knowledge-base upkeep, security review, and re-validation as models drift, which are real recurring costs the tool itself creates.
There is a cost most business cases ignore entirely, which is the risk the agent consumes. A system that acts can err in ways that carry financial consequence. For example, faulty decisions, conduct and fairness exposures, the concentration risk of everyone depending on the same underlying model, and, for autonomous systems, the distinctive failure modes of software that pursues a goal past the lines no one thought to draw.
In a bank, that exposure has a financial cost, with money set aside for expected losses and additional capital held against more extreme risks. A return stated gross of that price is not a return; it is half of one. Benefit and risk are two sides of a single decision, and evaluating them in separate documents before separate committees (as almost every institution still does) guarantees that returns are approved without the risks they generate.
The number that clears the hurdle
The value of an AI deployment is its validated benefit, minus its true cost of ownership, minus the expected loss it creates, minus the cost of the capital it ties up expressed not as a single confident figure but as a distribution with a range. Divide the risk-adjusted return by the capital consumed and you get a number directly comparable to every other use of the firm’s capital, tested against the same hurdle rate. AI stops being a special case pleading for indulgence and becomes one more claim on capital, evaluated like the others.
Work backwards before launching a pilot and you can establish a clear break-even point: the minimum improvement the tool needs to deliver to justify its cost, risk and capital requirements. That gives every pilot a measurable target agreed in advance.
Look across the whole AI programme and a familiar pattern often emerges. A small number of deployments generate most of the proven value, while a long tail continues to absorb money, oversight and resources without enough evidence to justify the investment.
It also rehabilitates the pilot. A rigorous trial that returns a clean negative is a losing investment killed before it scaled, at a quantifiable saving. The firms that will compound real advantage from AI are not the boldest deployers or the most cautious. They are the ones with the patience to run experiments that tell them something true, and the infrastructure to convert what is true into a number a board can stand behind.
What the CFO should ask
None of this requires the CFO to become a data scientist. It requires refusing to accept a return figure that cannot withstand basic scrutiny. That means establishing the counterfactual, so you know what would have happened anyway; measuring the impact across the full workflow, not just at one step; showing how time saved translates into real financial value, who owns that value and when it will be audited; and accounting for the risk side, including any expected losses or additional capital requirements. Only then can you judge whether the return still clears the hurdle. Every one of those questions is easy to ask and nearly impossible to answer with the numbers currently in circulation. The institutions that close this gap will know which of their AI bets are paying off and which are quietly draining capital, and will be able to prove it, to a board, a regulator, and themselves. The rest will keep mandating adoption, celebrating dashboards, and wondering why the returns everyone can feel never quite arrive in the accounts.
