AI return on investment: why it is taking so long

If you have read that an overwhelming majority of AI projects bring in nothing, you have read something partly true. The serious studies on AI return on investment in business all point the same way: only some companies get visible value from generative AI, even if their share is growing. But they also say why: the problem rarely lies in the model. It lies in the organisation, in the lack of measurement, and in adoption that happens bit by bit, with no overall plan.

That is rather good news: an organisational problem can be fixed.

The figure that frightened people: “95%”

The figure comes from a report by the NANDA project at MIT Media Lab, “The GenAI Divide: State of AI in Business 2025”, published in July 2025. Its most quoted sentence: despite $30 to $40 billion invested by companies in generative AI, “95% of organizations are getting zero return”.

Before repeating it, you need to know how it was produced: a review of more than 300 publicly disclosed AI initiatives, interviews with representatives of 52 organisations, and 153 responses from executives collected at four industry conferences, between January and June 2025. It is not an accounting measurement: the authors themselves write that their figures are “directionally accurate”, but based on interviews and not on company accounts.

Several weaknesses are worth spelling out:

• “Zero return” does not mean “money lost”. A success there is a project whose users or executives report a marked and lasting effect on productivity or the profit and loss account. The 95% therefore counts projects with no reported impact, not measured losses.
• It is a preliminary version (0.1), whose only named reviewer is one of its authors, and whose circulating copy is not hosted on an MIT page. It contradicts itself at least once: on a single page, the share of budgets going to sales and marketing is half in the summary box, then about 70% a few lines further down. And it promotes an agent protocol backed by its own authors.
• The 5% applies to a specific scope: specialised tools taken all the way to production, at the end of a funnel that the report describes as follows.

So this report is not proof. It is a signal, to be checked against others.

Do the other studies say the same thing?

Roughly, yes, with different figures because the definitions differ.

IBM’s 2025 study of CEOs, run with Oxford Economics among 2,000 executives in 33 countries: the CEOs surveyed say that only 25% of AI initiatives have delivered the expected return in recent years, and that only 16% have been rolled out across the whole company. Useful to know: IBM sells AI.

The Boston Consulting Group painted a less gloomy picture on 30 September 2026: it places 48.5% of the companies surveyed among those that get value from AI, up from 26% in 2024. Most of that value is still concentrated in a small group of leading companies, and BCG counts only 7.5% among the most advanced. The survey collects the statements of 1,330 executives in more than 60 countries, and more than half of the companies surveyed have revenue above 5 billion dollars: nothing like an SME. The sample changes from one year to the next and the scoring bar was raised on nine criteria: the rise shows a trend, not a precise measurement. Here too, the author sells AI transformation consulting.

Three sources, three sponsors with an interest in companies investing in AI, and yet one finding: value does not arrive everywhere, and it is concentrated.

Why: organisation before technology

It is on the cause that the studies agree most clearly. BCG has turned it into a rule, the “10-20-70” rule: devote 10% of the effort to algorithms, 20% to technology and data, and the remaining 70% to people, organisation and processes. That is a recommendation, not a measurement. Its 2026 survey does, however, observe where the companies it judges most advanced put their priorities:

These are statements, and a correlation: nothing proves that this split explains their lead. But it matches what the other studies describe.

The MIT report puts it differently: the main obstacle is neither infrastructure, nor regulation, nor skills: it is learning. Most systems do not retain feedback, do not adapt to context and do not improve over time. The same report describes a “shadow AI”: only 40% of companies say they have bought an official subscription to a language model, while employees at more than 90% of them say they use one regularly. In other words, people use AI, but the company does not organise itself around it.

IBM adds a piece to the puzzle: half the CEOs surveyed acknowledge that their rapid investments have produced “disconnected, piecemeal technology”. An AI placed on top of a poorly maintained CRM, for example, quickly produces wrong answers, because the raw material is wrong.

Yet the gains exist when the task suits it

If we stopped there, we would conclude that AI is of no use. The productivity measurements say the opposite, with one condition.

• In a customer service department, a study published in the Quarterly Journal of Economics on real data from 5,172 agents shows an average 15% increase in issues resolved per hour. The least experienced gain the most; among the most highly skilled, quality drops slightly.
• On professional writing tasks, an experiment published in Science with 453 professionals measures an average time reduced by 40% and quality up by 18%.
• On 18 realistic consulting tasks designed with BCG, a pre-registered experiment published in Organization Science with 758 professionals measures 12.2% more tasks completed, 25.1% faster, with better quality. But on a task outside the AI’s capabilities, participants were 19% less likely to produce a good answer.

This last point is the heart of the matter, and we devote an article on the role of humans facing AI to it: AI helps a great deal on some tasks and degrades the work on neighbouring ones.

In July 2025, a randomised trial by the organisation METR measured that 16 experienced developers took 19% more time with AI, while they estimated they had saved 20%. METR has since said these results are outdated: its update of February 2026 judges the effect of AI to be probably more favourable in early 2026, without being able to put a reliable figure on it. The lesson is not that AI slows people down: it is that the perceived gain and the measured gain can diverge.

And in Europe?

The previous studies say almost nothing about small European companies, which official statistics cover better. According to Eurostat, 20.0% of EU enterprises with 10 or more employees used AI in 2025, up from 13.5% in 2024.

Eurostat’s database mainly shows the gap by size. It also details the obstacles cited by those that considered AI without adopting it.

The first obstacle is therefore a lack of expertise, and judging AI useless for the business is the least cited reason. These figures cover enterprises with 10 or more people, excluding agriculture, mining and finance.

Lessons for an SME

Here is what we draw from it, distinguishing what is established from what is only an observation.

1. Start from a task, not from an “AI project”. The gains measured by these studies all concern specific tasks: answering a customer, writing a professional text, producing an analysis. “Doing AI” is not a task.
2. Measure before, measure after. Time spent, error rate, response time: one figure before deployment, the same figure three months later. Without a starting point, you will never know whether the tool pays off, and gut feeling is not enough, as METR showed.
3. Start with the back office. The MIT report observes that the clearest return often comes from internal functions (operations, finance, document processing) rather than visible marketing projects. That is an observation, not proof. Our own expense-entry agent falls into this category: in production on Telegram since July 2026, it goes from the photo of the receipt to the entry in the accounting system, after human validation.
4. Change the way of working at the same time as the tool. A tool placed on top of an unchanged process will produce the same result as before, at a higher cost. That is also where the most advanced companies put most of their priorities, according to BCG: on people and processes.
5. Train the teams. It is also an obligation: Article 4 of the EU AI Act, rewritten by Regulation (EU) 2026/1744 and in force since 27 July 2026, asks companies that deploy AI to take measures to support the development of AI literacy among their staff, no longer requiring them to ensure a sufficient level of it as its 2024 version did.
6. Pick the cheapest model that does the job. Return on investment is also decided by the cost per request: a large model to sort emails is overpaying. We explain how the cost of AI in business is calculated and forecast.
7. Do not build everything alone. According to MIT, projects run with an external partner reach deployment about twice as often as tools built in-house. Correlation is not causation, the authors warn, and we are ourselves an external partner: read this with caution.

The limits of everything above

None of these studies measures return on investment in company accounts: the surveys collect statements, the experiments measure productivity on a single task. The samples include few small companies. Scaling up from measured gains to a whole company has not been demonstrated. And we do not have, in October 2026, a publishable before-and-after measurement from our own engagements: we would rather say so than invent a case.

What we think

The words “taking so long” in the title matter: they describe a delay, not a failure. The MIT report itself acknowledges that six months of observation may underestimate the success of a complex deployment. The value will not arrive with the next model: it will arrive when companies pick a task, measure it and reorganise the work around it.

To go further, see our articles on business AI, private and at the right cost. And if you would rather start from a task with figures than from a vague project, that is the approach of our digital and AI maturity audit: knowing what to automate before spending. If AI is already running and the bill keeps rising without you knowing what pays off, our AI at the right cost offer measures the cost of each use and hands each task to the model that is good enough.

Frequently asked questions

Sources

1. “The GenAI Divide: State of AI in Business 2025” (new tab)mlq.ai
2. IBM’s 2025 study of CEOs (new tab)newsroom.ibm.com
3. The Boston Consulting Group (new tab)bcg.com
4. A study published in the Quarterly Journal of Economics (new tab)doi.org
5. An experiment published in Science (new tab)doi.org
6. A pre-registered experiment published in Organization Science (new tab)doi.org
7. A randomised trial by the organisation METR (new tab)arxiv.org
8. Has since said these results are outdated (new tab)metr.org
9. Its update of February 2026 (new tab)metr.org
10. Eurostat (new tab)ec.europa.eu
11. Eurostat’s database (new tab)ec.europa.eu
12. Rewritten by Regulation (EU) 2026/1744 (new tab)eur-lex.europa.eu