57% of senior AI leaders say their AI returns still don't outpace the costs: what to measure instead

A study of senior AI leaders in North America, the UK and Europe has a number worth sitting with. The share who say their AI returns still don't outpace the costs has held at 57% since 2025, even as production capability keeps climbing (Source: Domino Data Lab, Fifth Annual Enterprise AI Report, 2026). That is now a two-year plateau, and it tells you something a single survey never could.

Here is the tension. In the same study of 639 senior AI leaders, 93% reported improved production capability in 2026, up from 88% the year before (Domino Data Lab). So the engineering got better and the returns did not move. The report puts the cause plainly: a last-mile gap between AI models in production and the business users who need to get value from them (Domino Data Lab). Independent research points the same way. MIT's Project NANDA found that 95% of the generative AI pilots it studied showed no measurable profit and loss impact (Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025).

The study surveyed leaders at Director level and above, at organisations with annual revenues of US$100 million or more, across North America, the United Kingdom and continental Europe, fielded in April 2026 by BARC Research (Domino Data Lab). This is a serious, repeated sample, which is why the flat line matters more than any one-off headline.

Why the last mile is a measurement problem, not a model problem#Link to this section

When capability rises and returns sit still, the bottleneck has moved past the model. It now lives in whether people actually use the tools, whether the work gets faster or better as a result, and whether that shows up in money. Most organisations can't yet see that clearly, because their measures sit at the other end of the pipe.

The usual dashboard tracks licences bought, seats provisioned and models shipped to production. Those are useful counts. Whether a business user got value on a Tuesday afternoon is the question around them. Deploying a model and delivering its value to the people who need it are different things, and the report finds a wide gap between the two (Domino Data Lab).

The pressure to close that gap is real and rising. Kyndryl found 61% of senior business leaders feel more pressure to prove ROI on their AI investments now than a year ago (Source: Kyndryl, 2025 Readiness Report). And the money keeps flowing regardless. BCG reports companies plan to spend about 1.7% of revenue on AI in 2026, more than double the 0.8% in 2025, with 94% saying they will keep spending at current or higher levels even if this year's initiatives fail to deliver a return (Source: BCG, AI Radar 2026). Spending that grows while returns stay flat is exactly the pattern that turns into scrutiny in the next planning cycle.

What to track, and what good looks like#Link to this section

If you want the last mile on your dashboard, measure it directly from the people doing the work. Four things tell you most of what you need to know.

Track adoption as active use, not seats. A licence that sits idle is a cost with nothing to show for it. Compare the people using a tool in a normal week with the people you're paying for. Good looks like most paid seats in weekly use, and a clear, short list of tools that everyone actually touches.

Track perceived impact by task. Ask the people using each tool whether it saved them time, improved quality, or did neither, on the specific work they do. Look at it per tool and per team, because averages hide the tools that are quietly doing nothing. Good looks like a majority reporting real time saved on tasks they repeat often.

Track unused licences and overlapping tools as a live number. Combine idle seats, tools that do the same job, and tools people rate as no help. Look at it monthly, not annually, so you can act before renewal. Good looks like a shrinking figure quarter on quarter as you drop the tools that fail the test.

Track the gap between capability and use. This is the one the study exposes. If your teams say a tool is powerful but they still route around it, you have a workflow and enablement problem, not a model problem.

None of these studies says what the organisations whose AI returns do outpace the costs are doing differently. Measuring who uses what, whether it helps and what it costs is how you'd find out which side of that line you're on, and the quickest way to see your own last-mile gap is to ask the people living in it.