Why do most companies see no return from AI?
Most companies see no return from AI because they buy the tool before anyone has counted the job it is for, and then put it on top of information that is scattered and disagrees with itself. In McKinsey's 2026 State of AI survey of 1,719 people, only 37% said AI had added anything at all to their company's operating profit.
The return shows up where someone picked one dull, countable job first. At a debt reconciliation firm we built for, that took 120 hours a week of follow-up off the team, and the build returned 350%.
How many companies actually see a return from AI?
In McKinsey's 2026 survey, only about a third of respondents said AI had added anything to their company's profit, and about 6% saw a large return. Two studies, run two different ways, point the same way.
- McKinsey, August 2026: 37% of respondents attribute at least some EBIT impact to AI, about the same as a year earlier. The share McKinsey calls high performers, with 5% or more of their EBIT coming from AI, has stayed at about 6% (McKinsey).
- MIT NANDA, July 2025: the report behind the headline that 95% of organisations get zero return, against $30 to 40 billion of spending (MIT NANDA).
The MIT figure is the one everyone quotes, and it rests on the smallest base: interviews with 52 organisations, 153 survey answers and a review of over 300 public projects. I'd treat it as a rough signal. McKinsey's 37% is the steadier number, and it's bad enough.
The same report has a figure I find more telling. MIT found that only 40% of the companies had bought an official AI subscription, while workers at more than 90% of them were using personal AI tools for work. So in most of the companies MIT looked at, AI was already in use, but that use hadn't reached the accounts.
Why do AI projects fail to deliver a return?
AI projects fail to deliver a return because nobody counted the job, the information underneath disagrees, the AI was added to the old way of working, and nobody measured the result afterwards. The MIT authors say the divide between the few that pay and the rest "seems to be determined by approach". In my view the approach goes wrong in four places.
- Nobody counted the job first. The tool was bought because AI felt urgent, not because someone measured how often a task happens and how long it takes. Without that baseline there's nothing to hold the result against, so even a real saving can't be shown.
- The information underneath disagrees with itself. When two systems disagree about a number, an AI on top repeats both versions with the same confidence. I went through that step in the article on what an AI operating system is.
- The AI was added to the old way of working. McKinsey's high performers redesign the work around the AI instead of inserting it into the process they already had, and nearly three-quarters of them say they fundamentally redesigned workflows, up from 55% a year before.
- Nobody wrote the number down afterwards. A saving in a pitch deck is a projection. If nobody measures the job again once the system is live, it stays one.
The third one is the easiest to miss. On a call in October with another agency's founder, I said the worst part of any dashboard is adding the information by hand, because nobody keeps it up to date. AI tools go the same way. If using one means opening another screen and typing things in, it ends up as one more login nobody opens, and the hours never move.
Where does AI actually pay back in a business?
AI pays back fastest on a job that is already costing you money, happens every day, and has a clear before and after. Often it sits at the end of a process, where nobody is measuring.
At a debt reconciliation firm we built for, every customer who signed up had until the end of that day to sign a debit order mandate, or the product they had just bought was cancelled. Staff couldn't reach everyone in time, and many customers didn't know how to complete the form, so revenue the firm had already won was lapsing on the last step.
We built an AI agent that follows up with every customer who hasn't signed, well before the deadline. It reminds them, explains what's needed, and walks anyone who is stuck through it step by step. Only the genuine exceptions reach a person.
The result was 120 hours of manual follow-up gone every week, 50% more mandates completed, and a 350% return on the build, measured against the mandates it saved and the hours it replaced (case study). The case study doesn't publish the count from before the build, so I can't show you the full sum for this one.
It's nobody's idea of an exciting AI project, and I think that's why it paid. The job had a clear count, because a mandate is either signed by the deadline or it isn't, so the result could be measured. And the agent did the chasing itself, so the team only saw the exceptions.
How do you work out the return before you build anything?
You work out the return by counting the job with the people who do it, putting a price on the hours, and agreeing how you will measure it again afterwards. It's worth doing before you talk to any vendor, us included.
- Name the job and how often it happens. Per day or per week, from the records if you have them.
- Get the time per task from the person who does it, then multiply. Put an hourly rate on the hours.
- Check where the information for that job lives. If two systems disagree about it, that has to be settled first.
- Decide what changes in the work. Who stops doing what, and where the AI's output lands so people actually use it.
- Agree how you'll measure it once it's live. The same count, a month later, plus the business result the job exists for, like mandates signed or quotes followed up.
Here is the sum with made-up numbers. A task that happens 30 times a week and takes 20 minutes each time is 10 hours a week. Over 46 working weeks that's 460 hours a year, and at €28 an hour (about the average Dutch gross salary as MKB Servicedesk reported it for 2025, with holiday pay, before employer costs) it comes to roughly €12,900 a year. That's the hours side of what a tool for that one job could be worth, before it costs anything. The business result from step 5 comes on top.
This is most of what our AI Assessment does over about three weeks, across every job worth looking at, and it ends in a costed plan that also says what not to build. I wrote up how we score each one in an earlier article.
When is counting hours the wrong way to judge AI?
Counting hours is the wrong test when the value isn't hours, or when the job is about to change anyway.
- The value is a better decision. Leadership deciding on this week's numbers instead of last month's is real value, but it won't show up in an hours sum. McKinsey's high performers aim AI at growth and innovation alongside efficiency. If you only count hours, you'll undervalue that kind of work.
- The job is about to change. Counting today's hours on a process you're redesigning next quarter overstates the return.
The survey figures above also mix every size of company together, so they can't tell you much about your own business.
Frequently asked questions
Is the MIT study right that 95% of companies get no return from AI?
It points the right way, but its sample is small: interviews with 52 organisations, 153 survey answers and a review of over 300 public projects. Larger surveys find a better but still poor picture. McKinsey's 2026 survey of 1,719 people found 37% seeing any profit impact at all.
How do you measure the return on AI?
Measure the job before and after: how often it happens and how long it takes, from the people who do it, priced at a rate you write down. Add the business result the job exists for, such as mandates signed or invoices paid. A return is measured against that result once the system is live.
Is our data too messy to get a return from AI?
Often it is, at first, and that's the honest answer. If your systems disagree about what a number means, AI on top will give confident answers that contradict each other. Settling those definitions is part of the build, and it's where we start.
Why do companies struggle to get measurable results from AI investments?
Mostly because nothing was measured before the AI went in. Staff often use AI tools for work, but without a counted job and a baseline, the time saved never turns into a number anyone can show.
For a business of 20 or more people, we'll find at least €100,000 a year in savings, then build the AI operating system that delivers them. If you want to count your own first, book a free intro call.
Related notes.
What could your business answer for itself?
Book a free call and we'll look at the questions your leadership waits on, and whether an AI operating system would pay for itself.
Emil Visser · Co-founder