The 95-Percent Question: What the MIT Study Really Says About AI

The 95-Percent Question: What the MIT Study Really Says About AI

· 10 min read

Auf Deutsch lesen →

Listen to this article
Speed

AI narration · automatically generated

Subscribe as a podcast
RSS
Personal note

Stored only locally in your browser — nothing is sent.

There are numbers that are too good not to share — and dangerous for exactly that reason. “95% of AI projects fail” is one of them. It confirms the skeptic (“knew it, all hype”) and the cynic (“the corporations are burning billions”) in equal measure, it fits into a tweet, and it carries the “MIT” seal. Almost everything about it is true. And yet nearly everyone who quotes it draws the wrong conclusion.

I will work through the number the way one should work through any viral statistic: What does the source really say? What do its critics say? And what is left when you lay both side by side?

What the study really says

The source is the report “The GenAI Divide — State of AI in Business 2025” from the MIT Media Lab, produced by the initiative Project NANDA, published in August 2025. The much-quoted figure reads precisely: roughly 95% of companies see no measurable effect of their GenAI pilots on the profit-and-loss statement; only about 5% achieve a noticeable acceleration in revenue. The context: an estimated 30 to 40 billion dollars have already flowed into such projects.

It pays to look closely right here. “No measurable P&L effect” is not the same as “failed.” It means: at the end of the quarter there is no line in the accounts that could be unambiguously attributed to the AI pilot. That may mean nothing happened — or that something happened which no one measured.

Far more important than the headline is the cause the study names — and it clears up the common misunderstanding. The bottleneck, the authors write, is not the quality of the models, but a “learning gap”: most systems retain no feedback, do not adapt to the operational context, do not improve over time. It is an organizational and integration problem, not a technology problem.

Three of the study’s findings make this concrete:

  • Buying beats building. Bought-in specialist solutions make it into production about 67% of the time; internally built ones only roughly a third as often. The expensive “we’ll do it ourselves” reflex is one of the best predictors of getting stuck.
  • The money sits in the back office, but it is spent in the shop window. Around half of budgets flow into sales and marketing tools — whereas the study found the measurable return in unglamorous process automation: saved agency and outsourcing costs, leaner workflows.
  • Adoption is not transformation. Large corporations run the most pilots and move the fewest into production (often nine months and longer); mid-sized firms make the leap in around 90 days. It is precisely this gulf the authors call the “GenAI Divide.”

This is a nuanced, usable study. It is just that almost nothing of it survives in the public memory except the one number.

What the critics argue — and they have points

It is too easy to simply parrot the 95%. Several serious voices have taken the study apart — and their objections belong in the picture:

First, the methodology is thinner than the authority “MIT” suggests. The trade publication Futuriom and the analyst behind the podcast Your Everyday AI point out that the 95% can barely be cleanly reconstructed from the report itself; the defensible observation is “a steep drop from pilot to production” — not “95% failure.” The sample is small and partly self-reported (around 300 public initiatives, a few dozen interviews, a good 150 survey responses), and the full raw data was never published. Anyone who puts such a steep number into the world ought to disclose the data.

Second, there is a definitional trick in it. The tech magazine Sify nails it (“misread by all”): “no measurable balance-sheet effect” does not mean “does not work”; at most it shows that many companies still run AI as a parallel tool alongside their processes rather than as an embedded operational change. That is an entirely different claim from the headline.

Third — and this is the most uncomfortable point — a conflict of interest. The report culminates in the recommendation to switch to “agentic,” learning AI systems. NANDA, as an MIT Media Lab initiative, offers exactly such systems itself; a corporate membership in the six-figure range is reported. A study whose diagnosis happens to prescribe precisely its own product deserves an extra dose of skepticism.

Fourth, the time horizon. ROI from fundamental AI integration usually takes years, not months. Measuring pilots against the balance sheet after a few months is almost methodologically guaranteed to disappoint.

What is interesting is where this critique leads. The Forbes writer Jason Snyder turns the finding around: the pilots fail, he writes, because companies shy away from friction — from the arduous work of genuinely redesigning workflows instead of placing a chatbot window next to them. That is not a contradiction of the study. It is its confirmation with the sign reversed: both say the problem sits in the organization, not in the model.

Why the number should be taken seriously anyway

After so much criticism, one might be tempted to dismiss the 95% entirely. That would be the second mistake after the first. Because the genuinely striking observation is not the MIT number alone — it is its confirmation by entirely independent investigations:

  • The Boston Consulting Group, in “The Widening AI Value Gap” (September 2025), concludes that only about 5% of companies extract substantial value from AI at scale — and 60% no material value at all.
  • McKinsey, in “The State of AI 2025,” counts only around 6% “AI high performers” with a noticeable EBIT effect. 88% use AI somewhere, but only 39% see any bottom-line effect at all — mostly under five percent.

Three houses, three methodologies — and one has to look closely at what actually converges. It is not a single digit; the three measure different things, and each knows, alongside its strict 5%, a much higher “any-value-at-all” number: McKinsey names 39% with an EBIT effect (mostly under five percent), BCG around 40% with some material value. What coincides, then, is not a number but a shape: a narrow head of about 5% that uses AI transformatively; a broader middle around 40% that sees measurable but modest returns; and a majority with no effect. Transformative value is rare, incremental value more common, none at all the rule — on that the three agree.

And this agreement is remarkable precisely because everyone making the case has skin in the game. BCG and McKinsey are not neutral parties — they sell exactly the transformation consulting whose necessity their reports establish, just as NANDA sells the agentic systems it recommends. The same skepticism about who is speaking that one owes the MIT study therefore applies to its star witnesses too. That the findings converge despite these vested interests — and do so with a hard behavioral figure: S&P Global (via 451 Research) reports that the share of companies abandoning most of their AI initiatives jumped from 17 to 42 percent in 2025 — makes the picture more robust than any single source.

The apparent contradiction dissolves once you go one level deeper. Because that AI works at the task level is just as well documented — the break lies between the individual task and the balance sheet. Perhaps the cleanest experiment on this comes from Harvard and BCG (“Navigating the Jagged Technological Frontier”): 758 management consultants, tested under controlled conditions. Within the AI’s zone of capability, the AI users completed 12.5% more tasks, 25% faster, at around 40% higher quality. But — and this is the honest edge — outside that “jagged frontier” they performed worse than colleagues with no AI at all. That is exactly what explains the 95%: tool productivity at the individual desk simply does not translate automatically into an effect in the corporate P&L. In between lies all the work the study calls “integration.”

Honestly, this very finding blurs the clean separation of “technology” and “integration.” That AI fails without warning outside its zone — and that users do not recognize the boundary — is also a reliability problem of the model; hallucinations force a level of checking that can eat the productivity gain back up. “Not the technology,” then, does not mean “the technology is finished,” but more soberly: its maturity matters less than the craft around it.

(Two frequently cited counter-figures — “74% see ROI in the first year” from Google Cloud, “$3.70 returned per dollar” from IDC — I deliberately let stand only as a footnote: both are vendor-financed. Whoever scrutinizes the MIT number must apply the same skepticism to the friendly ones.)

What remains

Lay thesis and antithesis side by side and a surprisingly reconciliatory insight remains: they argue about the number, but not about the diagnosis. Whether 95 or 80 percent, whether “failure” or “no balance-sheet effect” — study and critics say the same thing: it is not the technology that fails, but the craft of adoption.

And the successful five percent, across all three studies, visibly do the same five things — with one caution up front: we see the winners, not those who tried the same and failed. The following patterns are not a recipe, but the best available correlation:

  1. They buy ready-to-use solutions instead of building everything themselves.
  2. They embed AI in real workflows instead of setting it up alongside them.
  3. They aim at the back office, where the return is measurable, not at the marketing show.
  4. They invest in processes and people, not only in models — BCG’s often-cited “10-20-70” rule of thumb places just 10% of success in the algorithms and 20% in the technology, but 70% in people and processes.
  5. They measure financial outcomes, not click counts — otherwise “95% without ROI” becomes a self-fulfilling prophecy.

Two things remain open in all this — and should stay open. Whether the precise 95 holds, nobody knows; NANDA never showed the raw data. And the fact that more companies abandoned projects in 2025 than the year before is a warning: “learnable” does not mean “easy.” What remains is therefore more modest than the headline — but more durable: the 95 percent are not a verdict on artificial intelligence, but on how organizations introduce it. And that lies, unlike the capabilities of a model, far more in one’s own hands than the headline suggests.

If this piece gave you something to think about, feel free to share it — and the next time a viral AI percentage comes around, ask yourself whether you are reading the number or its headline.

Sources (selection):