Summary
Since August 2025 one statistic has dominated every enterprise AI conversation: 95% of GenAI pilots fail. It has been used to kill budgets, to justify budgets, to sell consulting and to excuse inaction. It comes from a report that surveyed 153 people at four industry conferences and defined success as whether someone had "remarked" that a tool helped.
The number is not reliable. The underlying phenomenon — a large, real, expensive gap between pilot and production — is well evidenced elsewhere by organizations with proper sampling frames. This is not the contrarian case that AI works fine; most of the corrected numbers are still uncomfortable. The point is that you cannot manage a portfolio using a statistic that cannot tell you which of your pilots is in trouble.
---
1. What the MIT NANDA study actually claimed
The report is The GenAI Divide: State of AI in Business 2025, from Project NANDA, an initiative affiliated with the MIT Media Lab. Authors: Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari. Research period January to June 2025. The headline claim, verbatim:
> "Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return."
> "Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact."
Note what this says and does not say. It describes organizations getting zero return, not pilots technically failing, and the second sentence adds a further filter — "extracting millions in value" — a bar a genuinely successful but modest pilot would not clear. The report also describes a funnel: roughly 60% of organizations evaluated enterprise-grade systems, 20% reached pilot, and 5% reached production (via Virtualization Review). That funnel is the interesting structure, and it is almost never quoted.
---
2. The methodology, stated plainly
The report describes "a multi-method research design that includes a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences."
Each component has a weakness. Public disclosure is a selection filter: organizations announce AI initiatives when they are newsworthy, so quiet working deployments and quiet failures are both absent. The 52 interviews are qualitative — useful for mechanism, not rates. The 153 survey responses collected at four industry conferences are the load-bearing weakness: such conferences attract people with an active interest in enterprise AI, usually because they are mid-program, evaluating vendors or selling something. That is a convenience sample of a self-selected population, and it cannot support a population-level failure rate for "organizations" — precisely what the 95% is used to assert.
Then the success definition, stated openly in the report and almost never quoted:
> "We define successfully implemented for task-specific GenAI tools as ones users or executives have remarked as causing a marked and sustained productivity and/or P&L impact."
That is a perception threshold — whether someone said something positive, not whether an auditable benefit was recorded. A pilot delivering a real 4% cycle-time improvement that nobody remarked upon scores as a failure.
Three further problems compound it. It is not peer reviewed — no journal, no replication, no public dataset. The 95% is not traceable to the published data: Wharton's Kevin Werbach, quoted in a published teardown, found "there appears to be no further support for the 95% claim," and the conversion from "investigated" to "piloted" to "succeeded" is never shown (Futuriom). And the institutional name carries credibility that conflicts sit underneath — the authors are affiliated with major AI vendors, and the MIT association is via a Media Lab-affiliated initiative rather than a traditional academic channel. The claim then widened in the retelling: "95% of organizations are getting zero return" became "95% of AI pilots fail," then "95% of AI projects fail."
How to use it honestly. Cite it as directional evidence of a widely observed pilot-to-production gap, corroborated by better-sampled sources — never as a measured failure rate. When it lands in a board pack, the right question is not "is AI worth it?" but "what is our own conversion rate from pilot to production, and do we measure it?"
---
3. The four-to-five-fold spread in "AI adoption"
The 95% is not the only number being misused. The adoption figures are worse, because the disagreement between credible sources is enormous and almost entirely a measurement artifact.
| Source | Figure | What it actually measures |
|---|---|---|
| McKinsey State of AI (n=1,993, 105 countries, Jun–Jul 2025) | 88% of organizations regularly use AI in at least one business function | Self-selected executive panel, weighted to GDP; "at least one function" |
| Stanford HAI AI Index 2026, Economy chapter | 88% organizations; 79% for GenAI specifically | Reproduces the McKinsey data |
| US Census Bureau BTOS (all US firms, Dec 2025–May 2026) | ~19.8% national average; 37% of firms with 250+ staff; under 20% of firms with four or fewer staff; firms under 20 employees showed no significant change over the period | Full population survey of US businesses |
| UK ONS (businesses with 10+ staff, June 2026) | ~35%, up from ~12% in late 2023; 49% of 250+ firms; 28% micro | Official national statistics |
Every one of these is honestly produced. They disagree by a factor of four to five because they count different populations against different thresholds. McKinsey surveys an executive panel and asks whether AI is used in at least one function anywhere in the business — a low threshold on a high-propensity population, for which 88% is the correct answer. The US Census asks the full population of US businesses, including the plumbing contractor and the three-person accountancy, whether the firm uses AI to produce goods or services — a higher threshold on a representative population, for which roughly 20% is correct.
What a leader should conclude. The 88% is an enterprise and large-cap phenomenon: across the actual business population, roughly four in five firms are not using AI at all. If your strategy assumes your competitors have all moved, check whether they sit in the McKinsey panel or the Census frame. Any business case resting on a single adoption figure is built on sand — ask which instrument, which population, which threshold. And the gap is the commercial opportunity, not the threat.
---
4. The better-sourced evidence on where value actually lands
Strip out the unreliable statistic and the picture that remains is still sobering — but it is measured, sampled and actionable.
| Finding | Source |
|---|---|
| ~67% have not begun scaling AI across the enterprise; only 6% are high performers | McKinsey State of AI (n=1,993) |
| Fewer than 20% have meaningfully scaled GenAI; only 7% run fully autonomous agents in production | Bain |
| 56% of CEOs report neither revenue increase nor cost decrease in 12 months; only 12% report both | PwC 2026 CEO Survey via Forbes |
| Only 17% of finance leaders say AI delivered "clear, measurable value" | Deloitte US CFO Insights 2026 |
| 39% report any EBIT impact; most attribute under 5% of EBIT to it | McKinsey State of AI |
| Nearly 40% that measured AI cost savings landed below 10%, having targeted 11–20% | Bain |
| Only 14% of CEOs define P&L impact for all AI initiatives | BCG AI Radar 2026 |
These say something different from "95% get zero return." Most organizations are early, most are not measuring properly, and among those that do measure, benefits are real but smaller than forecast. "95% fail" is a reason to give up; "40% of measured savings landed below 10% against an 11–20% target" is a reason to fix the forecasting.
The trajectory matters more than the level. Bain finds satisfaction climbing from 25% "strongly positive" at pilot stage, to 41% in full production, to over 60% among the most mature organizations (Bain). If pilots were fundamentally worthless, satisfaction would not rise with maturity. It rises because the value sits on the other side of the crossing. This is a crossing problem, not a value problem, and that reframing changes what you fund.
The clearest value lever is not technological. Redesign-first organizations achieve a 10–15% productivity lift and 10–25% EBITDA gains, while tool-first firms get what Bain calls "micro-productivity: faster reports, modest task savings" (Bain). BCG finds high performers 7x more likely to redesign workflows end-to-end (BCG); McKinsey independently finds them 3x more likely to redesign workflows and to have senior leadership ownership (McKinsey).
---
5. The NANDA findings that are robust
Dismissing the headline does not mean dismissing the report. Three findings survive scrutiny and are more useful than the number that made it famous.
Build versus buy. External partnerships with learning-capable, customized tools reached deployment approximately 67% of the time versus 33% for internally built tools (MIT NANDA). This is a simple ratio from the initiative review rather than the conference survey, which is why even critics accept it — and it is the highest-leverage decision in most portfolios, usually made early and by default.
Back office beats front office. Value concentrates in back-office and operational functions rather than the customer-facing deployments that attract budget and press attention. The reversal cases in section 7 are, without exception, front-office.
Shadow AI. Only 40% of companies bought an official LLM subscription, but workers at over 90% of surveyed companies reported regular use of personal AI tools for work (MIT NANDA). Independent research corroborates it: 49% of employees use unapproved AI tools and 51% have connected AI to work systems without IT approval, with leaders the worst offenders — 69% of presidents and C-suite accept shadow AI (BlackFog via CIO, n=2,000 in organizations with 500+ employees, Jan 2026). And 57% of employees hide their AI use and present AI output as their own (KPMG / University of Melbourne, n=48,000+, 47 countries).
That last figure has a consequence nobody discusses: if 57% of your people conceal their AI use, you cannot measure AI's productivity effect at all. Some share of the "zero return" finding is measurement failure caused by concealment — leaders estimate 4% of employees use GenAI for 30% or more of daily work while employees self-report 13% (McKinsey).
The report's root-cause diagnosis also beats its headline: the constraint is not infrastructure, regulation or talent but learning capability — "most GenAI systems do not retain feedback, adapt to context, or improve over time" (via Virtualization Review). A deployment with no feedback loop will not hold its performance.
---
6. The change-management diagnosis
The strongest-evidenced claim in the field is that this is not a technology problem. Four independent research programs converge on it.
- 70% of AI value comes from people and process changes, 20% from data and technology, 10% from algorithms — BCG's "10-20-70" rule (BCG AI Radar 2026).
- Organizational factors — culture, manager support, talent practices — account for more than twice the AI impact of individual factors: 67% versus 32% (Microsoft Work Trend Index 2026, n=20,000, 10 countries, Feb–Apr 2026, with Edelman DxI).
- 63% of AI implementation challenges stem from human factors, not technical limitations; user proficiency is the largest failure category at 38% of failure points against 16% for technical challenges (Prosci, n=1,107).
- Nearly 60% cite knowledge and training gaps as the primary barrier to responsible AI (McKinsey State of AI Trust 2026).
Two mechanisms inside that are directly actionable.
The trust gradient runs the wrong way relative to where change budget goes. Prosci finds executives expressing strong confidence, team leaders moderate and cautious trust, and frontline workers minimal trust (Prosci). Most programs concentrate effort at the top, where belief already exists.
Manager modeling is the highest-return intervention found in 2026, and it is free. Where managers actively modeled AI use, employees reported a 17-point lift in perceived AI value, a 22-point lift in critical thinking, and a 30-point lift in trust in agentic AI; the highest-performing cohort's managers openly use AI (85% versus 64%) and set quality standards (83% versus 57%) (Microsoft WTI 2026).
Against that sits an incentive failure only the chief executive can repair: 65% fear falling behind without AI adoption, yet 45% find hitting current goals feels safer than redesigning work (Microsoft WTI 2026). Nobody redesigns a process while being measured on the output of the un-redesigned one.
---
7. The autonomy gap
This is the most expensive single mis-specification in enterprise AI, and it explains a large share of "AI didn't deliver ROI." Bain names it precisely: business cases assume full automation; production reality uses human oversight. Only 7% of companies run fully autonomous agents in production, 38% require human approval, and 32% operate with guardrails (Bain).
A case built on removing the human returns a fraction of forecast benefit when the human is still in the loop. The pilot did not fail. The forecast was wrong.
The remedy is short and high-payoff: map each process to one of four honest states — fully autonomous, guardrailed, approval-gated or human-only — and build the case on the state it will be in at go-live, not the state the vendor demonstrated. The consequences of skipping it are documented and named.
Klarna cut roughly 700 customer service roles on the basis that its AI assistant did the work of 700 agents, then reversed — the chief executive conceding publicly that cost-cutting had been over-prioritized and quality and retention had suffered (Forbes; Entrepreneur).
Commonwealth Bank of Australia made 45 roles redundant in July 2025, attributing reduced call volumes to its "Hey CommBank" chatbot. The Finance Sector Union pointed out volumes were in fact rising and management was already covering them with overtime. On 21 August 2025 the bank reversed: "We have apologized to the employees concerned and acknowledge we should have been more thorough in our assessment" (ACS Information Age; The Register; Bloomberg).
Taco Bell deployed AI voice ordering at over 500 restaurants across more than 2 million orders, then slowed expansion after peak-hour failures, misheard orders and an order for 18,000 cups of water that crashed the system, acknowledging human crew may be more effective at high-volume locations (eWeek).
All three are front-office deployments where the business case assumed an autonomy level the operation could not sustain. None is a story about model quality.
---
8. How to judge whether your pilot is on track
Lagging indicators — realized P&L, EBIT impact, headcount reduction — arrive too late to steer by, and are exactly what the unreliable statistics measure badly. Below is a leading-indicator set: each observable within weeks, each mapped to a documented failure mode.
| Leading indicator | Why it matters |
|---|---|
| A written value hypothesis with a named P&L owner outside IT | Only 14% of CEOs define P&L impact for all initiatives (BCG) |
| The honest autonomy state is documented — autonomous, guardrailed, approval-gated or human-only | 7% autonomous, 38% approval-required in production (Bain) |
| The underlying process was examined before the tool was inserted | Redesign-first delivers 10–15% productivity vs "faster reports" (Bain) |
| A feedback loop is designed in, with an owner | Systems that do not retain feedback decay (via Virtualization Review) |
| Line managers in the affected team use the tool visibly | 17/22/30-point lifts from manager modeling (Microsoft WTI 2026) |
| Frontline trust is measured, not assumed | Executive confidence is high; frontline trust is minimal (Prosci) |
| Usage is measured on outputs created, not licenses issued or courses completed | 57% hide their AI use (KPMG / Melbourne) |
| Build-vs-buy was a deliberate decision, not a default | 67% vs 33% deployment rate (MIT NANDA) |
| A scheduled kill gate exists, with a named decision-maker | High performers review on 12-week cycles (BCG) |
| Incentives for the affected team have been checked | 45% find current goals safer than redesigning work (Microsoft WTI 2026) |
A pilot scoring well on eight or more of these is on track regardless of its early P&L. A pilot scoring below five will appear in someone's failure statistic, and no amount of model improvement will change that.
Two further disciplines. Gartner forecasts that over 40% of agentic AI projects will be canceled by end-2027 on grounds of escalating cost, unclear business value or inadequate risk controls, and estimates that of the thousands of vendors claiming agentic capability roughly 130 are genuine (Gartner, 25 Jun 2025). So: a portfolio that cancels nothing is not disciplined, it is one where nobody has authority to stop things. And demand production references when buying — a demonstration is not evidence of production readiness.
---
9. What to say when the 95% comes up
The honest position, in three sentences:
The 95% comes from a non-peer-reviewed study using 153 conference survey responses and a perception-based success definition, and its critics — including at Wharton — cannot trace the number to the published data. The pilot-to-production gap it describes is nonetheless real and better evidenced elsewhere: roughly two-thirds of organizations have not begun scaling, fewer than one in five have meaningfully scaled GenAI, and 56% of CEOs report no revenue or cost movement at all. What separates the minority who get value is not better models — it is workflow redesign, honest business cases, manager-led adoption and a governance cadence that will actually stop things.
That is a harder message than either "AI is transformative" or "AI is a bubble." It is also the only one of the three that tells you what to do next.
---
EZ Consulting's AI Readiness Assessment is a fixed-fee, two-week engagement delivered under our Advisory practice: a scored assessment against a published framework, a use-case inventory with data-readiness scoring, a prioritized 90-day plan with a measurement baseline, and a board-ready summary. ez.consulting
---
Sources
The NANDA study and its critique
- MIT NANDA — The GenAI Divide: State of AI in Business 2025 (full PDF)
- Futuriom — Why We Don't Believe MIT NANDA's Weird AI Study (methodology critique, incl. Kevin Werbach)
- AIGL — State of AI in Business 2025: methodology summary
- Virtualization Review — MIT report finds most AI business investments fail (19 Aug 2025)
- Forbes — MIT finds 95% of GenAI pilots fail because companies avoid friction (26 Aug 2025)
Adoption measurement
- McKinsey — The state of AI (n=1,993, 105 countries, Jun–Jul 2025)
- Stanford HAI — 2026 AI Index, Economy chapter (PDF)
- US Census Bureau — AI use in businesses (May 2026)
- UK ONS — Artificial intelligence in UK businesses, 2023 to 2026
Value, ROI and scaling
- Bain — How companies create value with AI: redesign, not tools
- BCG — What's driving AI value (AI Radar 2026)
- BCG — How CEOs can scale AI value across the enterprise (2026)
- PwC 2026 CEO Survey — analysis via Forbes
- Deloitte — US CFO Insights: AI cost, risk, ROI (2026, PDF)
- Gartner — Over 40% of agentic AI projects will be canceled by end of 2027 (25 Jun 2025)
Change management, trust and shadow AI
- Microsoft — 2026 Work Trend Index (n=20,000, 10 countries, Feb–Apr 2026)
- Prosci — Why AI transformation fails (n=1,107)
- McKinsey — State of AI trust in 2026
- McKinsey — Leaders underestimate employees' AI use
- KPMG / University of Melbourne — Trust, attitudes and use of AI (n=48,000+, 47 countries)
- BlackFog shadow AI survey via CIO (n=2,000 in organizations with 500+ employees, Jan 2026)
Documented reversals
- Forbes — Klarna reverses AI push
- Entrepreneur — Klarna CEO reverses course by hiring more humans
- ACS Information Age — CBA reverses AI-driven job cuts, admits error
- The Register — Bank reverses decision to replace 45 staff with chatbot
- Bloomberg — Australia's biggest bank reverses plan to replace jobs with AI
- eWeek — Taco Bell scales back AI plans