Last month, someone showed me a business case: Copilot saved each employee 30 minutes per day. Impressive, on paper. I asked where those 30 minutes appeared in the budget. The room went quiet.
That is not a facetious question. It is the only question that matters.
Saved time feels like a return, but it is not automatically a return. If an employee saves thirty minutes a day and then spends thirty minutes doing something else, nothing has changed in your costs or revenue. The time has shifted, not been converted into value. Yet in almost every AI business case, those thirty minutes are multiplied by the number of employees and an hourly rate, then presented as a hard saving. That creates a return that never appears in the accounts.
The figure every board member should know
If one figure shows the scale of this problem, it is this one. Research by the MIT NANDA initiative (opens in new window), titled The GenAI Divide, found that around 95% of organisations achieve no measurable result from their generative AI pilots. Only around 5% achieve rapid revenue acceleration. The rest are left with little or no demonstrable effect on the profit and loss account.
Notice the wording: no measurable effect on the profit and loss account. That does not mean the 95% did nothing. They built, tested and rolled out. What was missing was a return visible in the numbers. A large part of that gap comes from the way results are measured. Anyone who records saved time as a cost saving is measuring an effect that never existed in the accounts.
McKinsey found something similar. Its State of AI 2025 (opens in new window) reports that 88% of organisations use AI in at least one function, yet only 39% report any effect on EBIT, and for most of them that effect is less than 5%. Broad adoption, narrow returns. That gap is the subject of this post.
Why saved time is so tempting
Saved time is a popular metric because it appears easy to measure and always produces a positive outcome. You ask people how much faster they complete a task, multiply the result and end up with an impressive figure. Nobody has to make a difficult choice.
That is exactly the problem. A return that requires no difficult decision is usually not a return. Saved time turns into money in only two ways. Either you convert it into lower costs, meaning less external hiring, less overtime or a task that is no longer performed. Or you convert it into more output at the same cost, meaning the same people do more work that creates genuine value. Both require a decision. Until that decision has been made, the saved time is a feeling, not a result.
That may sound harsh, and I do not mean it that way. Saved time is real, pleasant and sometimes reason enough to use a tool. But do not call it ROI. Call it convenience, job satisfaction or quality. ROI is a specific thing: euros out compared with euros in.
There is a subtler trap. Even when saved time is converted into money, that rarely happens with the first employee who saves five minutes. It happens at team level, when enough small savings together mean that a task can be organised differently. Five minutes per person is noise. Five minutes per person across an entire team, followed by a real reorganisation of the work, can mean something. The difference is whether someone actually carries out that reorganisation. Without that step, the saving remains fragmented across individuals who do nothing with it, and fragmented time is not money.
The formula is simple; filling it in honestly is not
The calculation itself is not advanced mathematics.
ROI = (benefits in euros minus the full costs in euros), divided by those full costs.
The difficulty is not in the formula. It is in filling in both sides honestly. On the benefit side, count only what appears in the accounts. On the cost side, count everything, including the hidden items. I wrote earlier about the full cost side, the total cost of ownership of AI, because if you include only token costs, you compare half the cost side with the entire benefit side and every business case looks too good.
One more condition is often skipped: measure beforehand. Without a baseline, you cannot know afterwards whether the effect was real, so you simply believe it, because believing is easier than measuring. Choose a period, record the starting point and compare against it. Otherwise, your ROI is a story, not a number.
What the successful 5% do differently
If 95% achieve no return, the interesting question is what the other 5% do differently. Three patterns emerge from the same MIT research, and none of them concerns better models.
- First, they buy more often than they build. Organisations that bought a targeted solution from a specialist were roughly twice as likely to succeed as organisations that tried to build their own tool. Building it yourself feels cheaper and sounds smarter, but has proved to be the least reliable route.
- Second, they target the back office rather than the shop window. Returns were found in repetitive, structured processes, not in the visible customer-facing applications that received most of the budgets. Boring beats sexy.
- Third, and this comes from the McKinsey data, the small group that genuinely scales redesigns workflows instead of placing AI as a layer on top of existing work. They change the work, not just the tool. That creates an effect you can see in the accounts because a redesigned process genuinely needs fewer hands or genuinely delivers more.
Follow me on LinkedIn
"But not everything can be expressed in euros"
At this point, someone often raises an objection, and it is a good one. Not all the value of AI can be expressed in euros. Better quality, less tedious work and faster responses to customers all matter, yet they do not fit neatly on a line in the budget. Am I defining that value out of existence?
No. I am saying that you should give it a different name. There are returns that appear in the accounts, and there is value you notice in other ways. Both are real. But they deserve separate treatment because they lead to different decisions.
If an application delivers a hard return, you can defend it with a number and continue investing on the basis of that number. If an application mainly delivers soft value, you have to make a judgement: is this job satisfaction or that faster response worth the money? That is a perfectly sound decision, provided you make it consciously. Things go wrong only when you disguise soft value as a hard return, because then you take a matter of judgement and pretend it is a calculation. You claim certainty you do not have, and one day someone asks where that certainty appears in the numbers. Then the room goes quiet, just as it did in my opening anecdote.
An honest measurement in four questions
When someone presents an AI business case to me, I ask four questions. They are simple, and they remove most of the hot air.
- First: where does the benefit appear in the numbers? Not in time, but in euros, on a line someone can point to.
- Second: which decision is attached to that benefit? If the saving is "less external hiring", who decides to hire fewer people? Without that decision, the benefit is theoretical.
- Third: are you counting the full costs? Including oversight, integration and remediation, not just the invoice.
- Fourth: do you have a baseline? So that in three months you can prove what you are promising now.
Four questions. If a business case survives all four, you probably have something real. If it already fails at question one, you have a feeling, and feelings do not belong in an ROI calculation.
Pilots measure something different from production
There is another reason so many ROI figures are wrong: they come from a pilot, and a pilot measures something different from reality. In a pilot, a small, enthusiastic team works with an application it chose itself. Conditions are ideal, motivation is high and the outcome is almost always positive. That is encouraging, but says little about what happens when you roll out the same thing across the entire organisation.
Everything changes during a broader rollout. The users are no longer all enthusiastic. The use cases are messier. Oversight has to scale, and costs scale with it. An ROI that shone in a pilot can evaporate in production simply because the ideal conditions have disappeared. Anyone who uses a pilot result to justify a large rollout is projecting the best scenario onto a situation in which that scenario no longer exists. So measure not only whether something worked in the pilot, but whether it keeps working when conditions become ordinary.
Why this is a board-level conversation
I understand that this can feel uncomfortable. You want to move quickly with AI, and then someone asks whether the benefits really are benefits. But that discomfort is precisely where the value lies. The 95% that achieve no return are not stuck because they bought the wrong models. They are stuck because nobody established in advance how the return would reach the numbers, then verified afterwards that it actually did.
That is a board-level question, not a technical one. The broader framework, from controlling your costs to calculating the full bill, is in my guide to AI costs and ROI. This post covers its sharpest point: dare to stop calling saved time a return, and measure what the profit and loss account recognises too.
The business case with the thirty minutes? I would have rebuilt it in no time. Not by removing the saved time, but by adding the missing question: and then what? If the answer was "less overtime in team X", we had something. If there was no answer, it was a reason to use the tool, not a reason to claim a return. That difference is the whole story.
Sources
* 95% of organisations achieve no measurable result from generative AI pilots (MIT NANDA, The GenAI Divide, via Fortune): fortune.com (opens in new window) * 88% use AI, while only 39% see any EBIT effect (McKinsey, State of AI 2025): mckinsey.com (opens in new window)