Talk to enough people in tech right now and you’ll hear the same word come up frequently: fatigue. Fatigue with headlines, the rollouts, the relentless promise that this year is the year everything changes. The conclusion most people come to is that we’ve simply had too much AI, too fast, and the appetite is wearing thin.
Look at the numbers, though, and a different picture starts to form. MIT’s NANDA initiative, in its widely circulated GenAI Divide report, analysed 300 public AI deployments alongside dozens of executive interviews and found that around 95% showed no measurable impact on profit or loss. Only 5% drove a clear, rapid return.
Yet the same body of research found that employees are using AI tools anyway, with or without formal approval, because they personally find them useful. The technology isn’t being rejected. It’s being used widely and judged unevenly.
That gap is the real story, and it isn’t AI fatigue. It’s ROI fatigue: the exhaustion of being asked, over and over, to produce a single number that proves it was all worth it, with no one having agreed in advance what that number should look like.
Until we stop confusing the two, every new wave of AI is going to produce the same tired headlines, no matter how good the technology gets.
Two Different Kinds of Tired
It’s worth pulling these apart, because they have different causes and different fixes.
AI fatigue is the one most people picture. They have too many tools, too many tabs, and too much “have you tried our new AI feature” popping up in software that just used to work. It’s real, and it’s measurable.
A recent Section survey of 5,000 white-collar workers across the US, UK and Canda, found a sizeable gap between what leaders believe AI is saving and what staff actually experience. Close to four in five C-suite respondents say AI saves them at least four hours a week, while two-thirds of employees put that figure at two hours or less.
That mismatch alone is enough to leave people feeling like their being sold something their day-to-day doesn’t back up.
ROI fatigue is less noticeable, and it sits a level above. It’s the exhaustion of being the person who has to justify the spend at the next board meeting, with no consistent way of showing what’s actually been gained.
WRITER’s 2026 enterprise AI survey, run with the research firm Workplace Intelligence across 2,400 executives and employees, found that 97% of executives said they’d personally benefited from AI, while only 29% said their organisation had seen a significant return on it. Nearly half admitted AI adoption had been a disappointment overall.
People are getting value from AI. Organisations are struggling to prove it. Those are two different problems, and right now, most of the commentary treats them as one.
Why The ROI Number Never Arrives
The proof problem isn’t isolated to one survey or one industry.
Gartner’s most recent CIO research found that the majority of organisations are currently breaking even or losing money on their AI investments.
PwC’s latest global CEO survey, covering more than 4,400 chief executives across 95 countries, found that only 12% say AI has delivered both cost and revenue benefit so far, and over half report no significant financial benefit at all yet. Among IBM’s own enterprise research, fewer than a third of leaders say they can confidently measure AI’s return at all.
In my experience working alongside B2B tech companies across tax, security, infrastructure and construction technology, three habits explain most of this, and none of them are about the technology itself.
One formula, many shapes of value. AI doesn’t create value in one consistent way. Sometimes it reduces errors. Other times it improves a customer’s experience without touching the bottom line for months.
Treating all of these as if they should produce the same kind of number is a bit like measuring a kettle and a car by how fast they go. The question doesn’t fit the thing you’re asking of it.
Tellingly, MIT’s NANDA researchers found that more than half of generative AI budgets are going into sales and marketing tools. Yet the strongest, most provable returns were sitting in unglamorous back-off automation, the kind of repetitive, rule-bound work that’s easy to baseline and easy to measure.
No “before” to compare to. It’s hard to prove improvement if no one captured what things looked like before the tool went in. Plenty of AI projects launch on enthusiasm rather than a baseline, which means that six months later, there’s nothing solid to measure the “after” against, just a feeling that things seem a bit better, or a bit busier.
Activity standing in for outcome. Adoption rates, login numbers, “percentage of staff using the tool weekly” – these get reported because they’re easy to pull and easy to put in a slide. None of them tell you whether the business is actually better off.
Usage is not the same as value, even though it’s often presented that way. Gartner’s most recent survey of chief sales officers found that nearly a third now cite difficulty proving ROL as a top challenge for the year ahead. This isn’t because the tools don’t work, but because nobody handled them a sensible way to measure what “working” means.
Three Questions That Cut Through It
The fix doesn’t need to be complicated. Before declaring any AI initiative a success or a failure, it’s worth asking three simple questions:
- What, specifically, were we trying to reduce or improve? Not “be more efficient”, a particular cost, delay or error rate.
- What was that number before we started? If there isn’t one, that’s the first job, before any tool goes live.
- What’s that number now, measured the same way? Not a survey of how people feel about the tool. The actual figure, on the same terms as the baseline.
It’s a small shift, but it changes the conversation completely. Instead of chasing one mythical “AI ROI” figure across the whole business, you end up with a portfolio of smaller, honest answers, some strong, some still pending, all of them real.
The pattern shows up clearly in Grant Thorton’s 2026 AI Impact survey of 950 business leaders. Organisations with fully integrated AI were four times more likely to report revenue growth than those still stuck in pilot mode.
The gap between them, the report notes, isn’t really about better technology. It’s about better accountability, knowing what was measured, who owns it, and what happens when something doesn’t work.
Some Sectors Have a Head Start
This plays out differently depending on what the underlying work looked like before AI ever turned up.
Workflows that were already metric-driven adapt fastest. In tax technology, turnaround time on a filing or a compliance check was always a number someone tracked. So, when AI speeds that up, the improvement is visible immediately, because the “before” already existed.
The same goes for digital trust and cybersecurity, where detection rates and false positives were being measured long before AI entered the picture, and for data centres and construction technology, where energy use, uptime and project timelines have always been tracked closely. It’s exactly the kind of process-heavy, rule-bound technology where MIT’s researchers found the real returns were hiding, even while the budget and headlines went elsewhere.
Contrast that with work that’s always relied on judgement rather than a number, strategic advice, creative direction, and complex negotiation. AI can clearly help here too, but the absence of an existing metric makes the ROI conversation genuinely harder, not because the value isn’t real, but because no one was measuring this kind of work in hard numbers to begin with.
There’s a regulatory nudge coming that will accelerate all of this regardless. Financial regulators, including the UK’s FCA, are increasingly expecting companies to produce documented, reviewable business cases for their AI deployments, not just confidence that it’s working.
For sectors handling sensitive data or regulated processes, rigorous measurement is about to become a requirement rather than a nice-to-have.
The Fatigue Lifts When the Question Changes
None of this is an argument against AI, and it’s not an argument for being cautious to the point of doing nothing. It’s an argument for being a bit more honest about what’s being measured, and a bit more patient about how long good measurement takes to set up.
The fatigue most people are describing right now isn’t really about the technology losing its shine. It’s about being asked, again and again, to answer a question that was never properly defined in the first place.
Define it – one use case, one baseline, one honest number at a time. Do that and the exhaustion has a funny way of easing off. Not because AI proves itself overnight, but because everyone finally knows what they’re looking for.
