Why two studies say the exact opposite
Because they did not ask the same question, and not of the same people. On this subject almost every number is self-reported: a few thousand executives get surveyed, and there is no audited P&L behind any of it. That is how one survey can measure a 95% failure rate and another can have three quarters of companies reporting a return.
The much-quoted 95% comes from a 2025 US report. It is worth knowing what sits behind it: a non-peer-reviewed preprint built on 52 interviews, 153 questionnaires collected at four industry conferences and a review of 300 public initiatives, with the data gathered in the first half of 2025. "Success" in that report means users or executives described the impact as marked and sustained — so it is self-reported too. The criticism of it is that the 95% figure itself has no workings anywhere in the document. No updated edition has appeared as of September 2026: the number quoted daily is sitting on data a year and a half old.
From the same report, though, come two more usable numbers. The first is the funnel: 60% of organisations evaluate some kind of enterprise AI tool, 20% get as far as a pilot, and 5% reach production. The second: a rollout started with an outside partner succeeds 67% of the time, an in-house build 33% — and the report itself adds that this may also reflect a difference in organisational capability.
The most recent large sample was collected in May and June 2026, across 97 countries, with 1,719 respondents. It produces the most important pair of numbers in this post:
| What is measured | 2025 | 2026 |
|---|---|---|
| Scaling AI across the enterprise | 38% | 44% |
| Reports at least some EBIT impact | 37% | 37% |
| EBIT impact of 5% or more, plus significant value | 6% | 6% |
Usage is growing, and the financial impact anyone can demonstrate has not moved in a year. The same survey also measures that 80% of respondents see an individual productivity gain. So the gap is not at the tool. It is at the point where an individual gain would have to turn into a company result.
Size matters, and in the wrong direction
The same survey splits large and smaller organisations out separately: 54% against 33%. In European statistics released in March 2026, adoption at large enterprises is 55% and at small ones 17% — so the gap is 38 percentage points, against roughly 30 a year earlier. It is not closing, it is opening.
Mid-sized companies have their own data, from March 2026, based on more than 200 executives: 57% are still in the pilot phase, only 15% have got as far as real scaling, and just 7% have a policy covering agents.
This is the most uncomfortable finding in the post: every publicly documented case is a large-company case. You will not find a named, publicly documented AI failure at a company of fifty to five hundred people — because that kind of thing does not reach a courtroom, a listed-company filing or the national news. So you can learn from the examples, but the correction for scale is yours to add.
Where to start at all is what strategy for a small company works through in more detail.
Five cases worth learning from
An Australian bank, 2025. It cut forty-five customer service roles on the grounds that voice AI had reduced call volumes. Call volumes had in fact gone up: overtime was offered, team leaders were put on the phones, and the decision was reversed. The bank admitted that it "did not adequately consider all relevant business considerations". The lesson is not about AI: the headcount decision was made on an assumed effect, not a measured one.
A Swedish fintech, 2024–2025. In its first month the AI assistant handled 2.3 million conversations, two thirds of all enquiries, with a resolution time under two minutes — the equivalent, the company said, of 700 agents. In May 2025 it started hiring people again. The nuance that matters: the bot was still handling two thirds of enquiries at that point. What failed was not the technology but full replacement as a strategy.
A Canadian airline, 2024. The chatbot wrongly told a passenger that the bereavement fare discount could be claimed retroactively. The company's defence was that the chatbot is "a separate legal entity that is responsible for its own actions". The tribunal rejected that: the chatbot is part of the website, and the company is liable for everything stated there. What the AI tells your customer, you told them.
The Australian arm of a large consultancy, 2025. A government report worth 440,000 Australian dollars contained a fabricated quotation from a Federal Court judgment and cited publications that do not exist. The firm paid back more than 97,000 Australian dollars. The model was not what went wrong: there was no checking step between the output and the client.
A US fast food chain, 2025. Voice AI running in hundreds of restaurants: inaccurate orders, interruptions, slower service, and a prank order that jammed the system. The company's digital chief conceded that at busier sites a human is better at taking the order. Load and environment decide, not whether the use case "is a good one".
Where the strongest evidence is
There is exactly one study on this subject that appeared in a peer-reviewed economics journal and is not built on a questionnaire. It was published in 2025: 5,172 customer service staff at a large company, with a staggered rollout, measured on real performance data. The result:
| Who was measured | How much faster |
|---|---|
| The average worker | +15% issues resolved per hour |
| Newer, less skilled | +30% |
| Experienced worker | effectively nothing, with a slight quality drop |
Customer satisfaction did not move meaningfully, but requests of the "get me a manager" kind fell significantly, and retention among new joiners improved.
This is the most useful sentence in the article: AI lifts the beginner towards the experienced. If you measure the effect on your best people, you will see nothing — and you will draw the wrong conclusion about the entire rollout.
Which areas produce measurable savings at all is something document processing shows with a concrete example.
The measurement almost nobody does
Here is the biggest gap in the whole body of research. In a July 2026 survey of chief executives, 56% name the difficulty of tying AI's impact to results as a barrier — but only 14% clearly define that impact for every initiative. Forty-two percentage points between thinking it is a problem and doing something about it.
In a survey published in April 2026, 65% of UK respondents said their company would keep investing in AI even without a tangible return. That is not stupidity — but it does mean the decision is not being made by measurement.
And there is evidence that self-reporting is unusable here. In a July 2025 experiment, 16 experienced developers were given 246 real tasks, with AI randomly permitted task by task. In the AI arm they were 19% slower — while beforehand they had expected a 24% speed-up, and afterwards still believed they had been 20% faster. They were thirty-nine percentage points wrong about their own experience.
One practical rule follows from this: measure a baseline before you roll anything out. Not a feeling, but a count, a turnaround time or an error rate. Without it, once the rollout is done you will have no way to decide whether anything improved.
The method behind that measurement is what measuring AI ROI turns into something you can actually run.
The hidden cost that self-reporting does not catch
There is a Danish study that followed roughly 25,000 employees per round, across 7,000 workplaces and 11 occupations, linked to administrative payroll records — so it never asked anyone how they felt.
- The average time saved is 2.8% of working hours (0.6% for teachers, 6.8% for marketing staff).
- The effect on earnings is exactly zero: the confidence interval rules out anything above one percent.
- And the most important part: 59% of the new tasks AI creates are the implementation and oversight of AI itself, and only 41% are productive AI use.
The work does not disappear, it changes shape. Anyone planning a rollout should budget for oversight as a standing cost, not as a one-off project task.
The cost side says the same thing: in a second-quarter 2026 survey across 20 countries, 49% of organisations delayed or narrowed an agent rollout once the cost exceeded the benefit, and only 35% have full visibility of what running AI costs them. The ones who do are five times more likely to reach a return.
What training actually delivers
This is the best-supported relationship in the whole body of material. A 2025 survey of 10,635 people across 11 countries breaks regular users down by how much training they received:
| Training received | Uses it regularly |
|---|---|
| More than 10 hours | 79% |
| 5–10 hours | 67% |
| Fewer than 5 hours | 54% |
Twenty-five percentage points of difference, and the format matters too: 82% after in-person training, 70% remotely. At the same time only 36% got the training they thought they needed — and that share had not changed by 2026.
The follow-up comes from the next edition of the same survey: 42% of regular users save a full working day or more per week, but two thirds of them get no guidance on what to do with the time that frees up. This is the point where a rollout quietly loses its result.
And the fear you have to plan for: 41% of employees worry about losing their job, and in a February 2026 survey of 23,000 US employees 18% think it likely their position will be eliminated within five years because of AI — 23% at companies that have adopted it. Actual AI-driven headcount reductions, meanwhile, were reported by 14% of organisations. So the fear is larger than the measured reality, and handled badly it leads exactly where you would expect: the team uses the tools in secret.
How to build the training is set out step by step in the AI literacy plan, ready to use.
The risk that is genuinely live today
Unauthorised use, measured rather than surveyed. According to a February 2026 telemetry analysis, 39.7% of the data movements heading into AI tools are sensitive data, the average employee puts something like that in once every three days, and — this is the important one — 32.3% of ChatGPT use runs on personal accounts. There the data processing agreement and the retention switch do not apply, because the data never travels under the company contract at all.
From which follows the cheapest, highest-impact rule in the entire rollout: company accounts, no exceptions. Not a ban but an alternative — instead of the forbidden tool, give people a corporate one that works.
Handling that unofficial usage is the subject of shadow AI, in considerably more depth.
Data protection: the real risk is your own usage. The most instructive Hungarian data protection case dates from 2022: a bank used AI voice analysis to score the emotional state of customer calls and the performance of its own staff. The authority imposed a fine of 250 million forints (roughly 670,000 euros at the time), because neither the customer nor the employee knew about the profiling, nothing was said at the start of the call about the purpose or the right to object, and the legitimate interest assessment had been done as a formality. An aggravating factor was that the emotion was not recognisable in 91.96% of cases.
What is a legal obligation today: supporting AI literacy has applied since February 2025, and the transparency rules since 2 August 2026 — with a chatbot you have to say that the person is talking to a machine. The rules for high-risk systems, by contrast, have slipped to 2 December 2027. And one common mistake: the fundamental rights impact assessment is mandatory only for public bodies, private entities providing public services, and deployers doing creditworthiness scoring or life and health insurance pricing — an ordinary company using recruitment AI does not fall under it.
Exactly what applies to you today is unpacked by the AI Act as it stands, date by date.
And on what you may upload to a chatbot in the first place, there is GDPR and AI, which takes that relationship apart.
Security: the deepfake is not the first thing to fear. According to a July 2026 study of 602 organisations, 25% of malicious incidents are AI-assisted, and those cost 6 million dollars on average. For executive fraud, though, it is worth looking at the official numbers: in the 2025 US crime report, reported business email compromise losses came to 3.05 billion dollars, of which the AI-linked share was "more than 30 million" — roughly one percent. Fraud is a large problem, but in 2025 it was not AI that made it one.
The order the evidence points to
1. BASELINE FIRST
Count, turnaround time, error rate - TODAY.
Without it "better" is only an impression.
2. PICK THE PROCESS, NOT THE TOOL
One process with a lot of repetition and a
measurable output. Not three at once.
3. MEASURE THE EFFECT ON BEGINNERS
The strongest evidence puts it at +30% there,
and near zero for experienced staff.
4. COMPANY ACCOUNTS AND A SHORT POLICY
One page. A green-amber-red tool list,
what not to type in, and who is responsible.
5. AT LEAST TEN HOURS OF TRAINING, IN PERSON
Above ten hours 79% become regular users,
under five hours 54%.
6. A HUMAN BETWEEN THE OUTPUT AND THE CUSTOMER
Most documented failures fell over here,
not at the model.
7. BUDGET FOR OVERSIGHT AS A STANDING COST
59% of the new AI tasks are the rollout
and the oversight itself. That does not pass.
8. SAY WHAT HAPPENS TO THE TIME THAT FREES UP
Two thirds of users get no answer to this.
Without it the saving never becomes a result.
Workflow redesign, the single strongest predictor of success, is covered at length in redesigning the workflow, with the counter-tests.
And for the policy there is a fill-in template: the one-page AI policy contains it.
Common misconceptions
- "95% of rollouts fail." This is a non-peer-reviewed number sitting on year-and-a-half-old data, and its own document shows no workings for it. The better-anchored figure: 44% of organisations are scaling, but only 37% can show any impact on results.
- "If the staff are happy, it works." In a controlled experiment developers got 19% slower while feeling faster. Self-reporting is not evidence here.
- "We will test it on our best people." That is exactly where the effect is smallest. The measured gain shows up with beginners.
- "AI replaces customer service." The best-known example stepped back — but even there the bot was handling two thirds of enquiries. Replacement failed, not the tool.
- "A ban solves shadow AI." It does not. A third of the data movement goes through personal accounts; a corporate alternative is the only thing that brings it back.
- "Compliance first, rollout afterwards." What is live today is transparency and supporting AI literacy. The high-risk rules have slipped to December 2027.
What to take away
The numbers that contradict each other are not in conflict because one of them is lying. They differ because they measure different things, on different people, with a different question. Anyone who pulls a single number out of this to decide with will decide badly.
What they do agree on: usage is growing everywhere, the demonstrated impact on results has not moved in a year, and the gap between large and small organisations is widening. The difference is not at the tool. It is whether someone redesigns the process, measures a baseline, and gives the team ten hours of training.
And the three cheapest things that would have prevented the documented failures: measurement before the decision, a checking step between the output and the customer, and a company account instead of a personal one. None of them is a technology question.
The most frequent traps, one by one, are what the ten mistakes gathers up in a single list.
Where does your company stand in its AI rollout?
Request a free 60-minute AI audit: three quick wins and a one-page action plan.