Last updated: October 1, 2026
Bottom line. Stop quoting “95% of AI pilots fail.” It comes from a small, non-peer-reviewed study, and its own data show something different: most companies in the sample never ran a custom AI pilot at all. The enterprise agent problem is still real. The stronger evidence for it (McKinsey’s survey, Gartner’s cancellation forecast, the Klarna reversal) points to measurement, ownership and governance failures rather than a 95% collapse. If your team can’t state the denominator behind an AI statistic, it shouldn’t be in your board deck.
1. The statistic that moved markets
In August 2025, a report from MIT’s Project NANDA, The GenAI Divide: State of AI in Business 2025, said that despite $30–40 billion in enterprise generative-AI spending, 95% of organizations were getting no measurable return. Fortune’s coverage turned that into a headline about pilots failing. The number then spread through business media and became a standard line in “AI is overhyped” arguments.
I traced it back to what the report says about its own method. The picture is less dramatic, and more useful.
What the study was built on. According to a summary of the report’s methodology, it combined a review of 300+ public initiatives, interviews with 52 organizations, and survey responses from 153 senior leaders collected at four industry conferences. It was labeled a preliminary report, and I found no sign it was peer-reviewed. Some early coverage cited larger interview and survey counts than the report itself states.
What the funnel actually shows. An analysis by Rob Wiblin of 80,000 Hours, recorded in February 2026, reads the report’s own chart this way:
| Stage (share of all surveyed organizations) | Share |
|---|---|
| Investigated custom or task-specific enterprise AI tools | 60% |
| Actually ran a pilot | 20% |
| Successfully deployed in production | 5% |
Investigated 60% ██████████████████████████████
Piloted 20% ██████████
Deployed 5% ██▌The “95%” is simply 100% − 5%. That lumps together the 40% who never investigated these tools and the 80% who never piloted one. Recomputing the conversion rates from the same three numbers:
| Step | Calculation | Result |
|---|---|---|
| Investigated → piloted | 20 ÷ 60 | ≈ 33% |
| Piloted → deployed | 5 ÷ 20 | 25% |
| Investigated → deployed | 5 ÷ 60 | ≈ 8% |
This is my own arithmetic on figures as read by Wiblin. I did not independently re-extract the chart from the PDF, so treat the inputs as reported rather than verified.
A quarter of pilots reaching production is not a flattering number, but it is a very different one from “95% fail.” The study’s success bar was also demanding. As Wiblin reads it, a project had to show a “marked and sustained” productivity or profit-and-loss impact within six months. Under that rule, a project that breaks even, or is on track to pay back in year two, counts as a failure.
Evidence note: how much weight can 5% carry? If that 5% rests on the 52 interviews, it equals roughly two or three organizations. A standard confidence interval for 2–3 successes out of 52 spans about 1% to 16%. (My calculation, assuming that sample basis; the report doesn’t publish an interval.) The headline figure is too imprecise to anchor a strategy.
Two limits on this critique. First, Wiblin’s piece is commentary, not a peer-reviewed rebuttal. His claim about the authors’ commercial interests in agent technology is his reading, though the report does promote the NANDA agent infrastructure. Second, debunking the number doesn’t prove AI projects are succeeding. It only removes one bad piece of evidence.
2. A better way to rank the evidence
When sources disagree, ranking them is more useful than averaging them. Here is how I’d weight what’s available on enterprise AI agents, strongest first:
| Rank | Source | Type | Why it gets this weight |
|---|---|---|---|
| 1 | McKinsey, The State of AI in 2025 | Large survey (~2,000 participants, 105 countries, mid-2025 per summaries) | Broad sample; self-reported; consistent definitions across questions |
| 2 | Company statements on the record (e.g., Klarna’s CEO to Bloomberg) | Primary, first-person | Reliable as to what was said; not independent proof of outcomes |
| 3 | Gartner forecasts | Analyst prediction | A forecast, not a measurement; useful for direction |
| 4 | MIT NANDA preliminary report | Small mixed-method study | Useful qualitative themes; weak for rates |
| 5 | Vendor-run surveys and aggregator “stats roundups” | Self-interested or untraceable | Frequently can’t be traced to a primary page |
3. What stronger evidence says
McKinsey: a thin top tier, and what separates it
McKinsey’s 2025 survey reports that 23% of respondents are scaling an agentic AI system somewhere in their company and another 39% are experimenting. The widely quoted “62% experimenting with agents” is those two groups added together, not a separate finding.
The part that matters more: only about 39% report any enterprise-level EBIT impact from AI, and most of that is under 5%. Roughly 6% qualify as “high performers” (5%+ of EBIT attributable to AI). Per the same summaries, within any single function no more than about 10% are truly scaling agents.
What distinguishes the high performers is concrete: workflow redesign shows one of the strongest correlations with impact, and high performers are about three times more likely to be scaling agents. In plain terms, the winners rebuilt the process around the agent instead of bolting an agent onto the old one.
Gartner: cancellations driven by cost, value and risk
In June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. It also estimated that only about 130 of the thousands of vendors marketing “agentic AI” are real, calling the rest “agent washing”: rebranded chatbots, RPA and assistants.
Two details people drop: the forecast is over 40%, not “40%,” and it describes projects cancelled, which is not the same as projects that “failed.” Gartner’s 2026 follow-up, its first Hype Cycle for Agentic AI (April 2026), places the category at the Peak of Inflated Expectations, per secondary coverage. I could not access the full report.
4. The 2026 problem: “agents in production” ranges from 9% to 80%
This is the finding I’d most like more readers to take away. In 2026, you can find a credible-looking statistic for almost any share of enterprises running agents, depending on the definition:
| Claim | Reported figure | Source as I found it | My confidence |
|---|---|---|---|
| Fully deployed agents | 9% | HBR Analytic Services report (AWS-sponsored), via summary | Moderate; sponsored |
| Organizations that have deployed agents | 17% | Gartner 2026 CIO & Technology Executive Survey, via secondary blogs | Moderate; not traced to Gartner directly |
| Scaling an agent in ≥1 function | 23% | McKinsey 2025 | Higher |
| Enterprises with ≥1 agent in production | 31% | Attributed to Gartner by a secondary blog | Conflicts with the 17% above |
| Enterprises running agents in production | 54% | Ampcome, via a vendor blog | Low; vendor-sourced |
Note the 17% versus 31%: two secondary sources attribute different figures to Gartner’s 2026 work. I couldn’t resolve this without subscriber access to the primary reports.
The spread isn’t mainly measurement error. The definitions differ: “uses AI embedded in a SaaS product,” “has one agent in a sandbox,” “has an agent autonomously taking actions in a core process” and “fully trusts agents for core processes” are four different things. Vendor surveys tend to pick the broadest definition, because it produces the best headline.
On ROI, the same pattern holds. A 2026 WRITER survey of 2,400 executives and employees reports only 29% see significant ROI from generative AI and 23% from agents. That’s a vendor survey, so weight it accordingly, but it lines up with McKinsey’s picture far better than “80% report measurable ROI” claims do.
A rule I now apply. Before using any AI adoption number, ask four questions. (1) What’s the denominator: all companies, companies that tried, or companies that responded? (2) What’s the definition of “deployed” or “success”? (3) What’s the time window? (4) Who funded or sells against the finding? If the source can’t answer all four, don’t cite it.
5. Klarna: what is documented, and what isn’t
Klarna is the most-cited agent case, and the one most often retold wrongly.
The launch (February 2024, Klarna press release). Klarna said its OpenAI-powered assistant had handled 2.3 million conversations, two-thirds of its customer service chats, in its first month. It claimed the work equivalent of 700 full-time agents, customer satisfaction on par with human agents, a 25% drop in repeat inquiries, resolution time cut from 11 minutes to under 2, and an estimated $40 million profit improvement in 2024. These are first-party claims.
The reversal (May 2025). In a Bloomberg interview, CEO Sebastian Siemiatkowski said that because cost had been “too predominant” an evaluation factor, the result was lower quality, and that investing in human support was the way forward. Klarna began recruiting human agents again, per reports, in an “Uber-type” remote model. It had also reduced headcount by about 22% through attrition, by that outlet’s account.
What I could not verify.
- Any public satisfaction figure showing a drop. The 2024 release claimed parity; the 2025 admission was about quality in general terms. Commentators have noted the tension between the two accounts.
- Any dollar figure for the cost of reversing course. Claims that the pivot “cost more than doing it right” are unsupported.
- Klarna’s 2026 position. I found no verified update.
The lesson that survives scrutiny. The CEO’s own explanation is an optimization problem: Klarna optimized the metric that was easy to measure (cost and volume handled) and under-measured the one that mattered (quality of resolution). That’s a metric design failure, not proof that AI can’t do customer service.
6. My position: the failure sits at the metric, not the model
Here is the claim I’d defend: the most common reason enterprise agents “don’t deliver” is that nobody defined, before launch, what delivering means. The evidence for this is indirect but consistent across sources:
- MIT’s study defined success as marked, sustained P&L impact within six months, then counted everything short of that as failure.
- McKinsey’s high performers are distinguished by redesigning workflows and tying AI to EBIT.
- Gartner attributes cancellations to cost, unclear value or weak risk controls: all things a pre-agreed metric would have surfaced.
- Klarna’s CEO attributes the reversal to cost dominating the evaluation.
This isn’t proven causation. It’s a reading of four sources that point the same way, and a skeptical reader can fairly say a company with a clear metric might simply be a company with better management in general. That confound can’t be separated with the data available.
A worked example: why “deflection rate” misleads
The numbers below are hypothetical, chosen to show the arithmetic. They are not benchmarks.
A support team handles 10,000 cases a month. A human-handled case costs $5. An AI-handled case costs $0.50. The agent “deflects” 70%, and 20% of those deflected customers come back and need a human anyway.
| Line | Calculation | Cost |
|---|---|---|
| All-human baseline | 10,000 × $5 | $50,000 |
| AI-handled cases | 7,000 × $0.50 | $3,500 |
| Escalated to humans | 3,000 × $5 | $15,000 |
| Recontacts after “deflection” | 1,400 × $5 | $7,000 |
| Total with agent | $25,500 |
The naive reading of 70% deflection suggests savings of about 63% ($31,500 of $50,000). Counting recontacts, the saving is about 49% ($24,500). And this still excludes churn from customers who give up rather than call back, which is the cost Klarna’s 2025 comments imply it ignored.
7. The 5-Gate Test (an original diagnostic)
This framework is my own synthesis, not a validated benchmark. It organizes the failure patterns above into questions you can answer before scaling. Thresholds are suggestions to adapt.
| Gate | Question to answer in writing | Fails if… |
|---|---|---|
| 1. Owner | Who is accountable, with budget, for the agent’s outcomes 12 months from now? | The answer is a committee or “the AI team” |
| 2. Metric | What single business metric, measured over what window, defines success, and what result means we shut it down? | Success is described as deflection, speed or “adoption” |
| 3. Error cost | What does one wrong action cost in dollars, legal exposure and customer trust? | Nobody has estimated it, or the agent acts autonomously where the estimate is high |
| 4. Visibility | Can you trace any single agent decision from input to action after the fact? | Failures would only be discovered through customer complaints |
| 5. Fallback | What happens when the agent is wrong or unsure, and who handles it? | The human path is slower or harder than before the agent existed |
Scoring: any failed gate is a stop, not a deduction. A pilot that passes four of five gates is not 80% ready.
8. Recommendations that change what you do on Monday
- Pre-register your success metric and evaluation window before the pilot starts, and write the kill criteria next to it. Most disputed pilots are disputes about moving goalposts.
- Replace deflection with cost per truly resolved case, including escalations and recontacts (see the worked example). Report it next to a quality measure sampled by humans.
- Set the autonomy ceiling by error cost, not by model confidence. Cases where one mistake is expensive start human-first, even if the agent is usually right.
- Interrogate “agentic” vendor claims. Ask which actions the product takes without human approval, what it does on failure, and how it is evaluated. Gartner’s agent-washing estimate (about 130 real vendors among thousands) means a rebranded chatbot is a live risk.
- For non-differentiating workflows, default to buying. The MIT report’s own data, per secondary summaries, reportedly favor external partnerships over internal builds (roughly two-thirds versus about one-third reaching deployment). The sample is small, so treat this as a tie-breaker, not a law.
- Redesign the workflow, not just the task. McKinsey’s high-performer correlation is with workflow redesign. An agent dropped into an unchanged process captures the least value.
- Retire the 95% figure from your materials. If you need a statistic for urgency, McKinsey’s “about 6% are high performers” is better supported.
9. What we still can’t answer
- What share of AI agent pilots reach production and pay back? No source here can say. Resolving it needs a larger, definition-consistent sample with outcomes tracked past six months, ideally with the underlying data released.
- How many of Gartner’s predicted cancellations will be good decisions? Some cancelled projects will be rational pruning, not failure. Cancellation rates alone can’t distinguish the two; post-mortems that record why each project stopped could.
- Did Klarna’s reversal pay off? I found no public financial analysis of the cost or benefit of the pivot.
Frequently asked questions
Do 95% of enterprise AI pilots fail?
No source I reviewed supports that as stated. The MIT NANDA report found roughly 5% of all surveyed organizations had deployed custom AI tools successfully, but about 80% had never piloted one, so the figure isn’t a pilot failure rate. By the same chart, around a quarter of those that piloted reached deployment.
What share of companies are scaling AI agents?
McKinsey’s 2025 survey found 23% scaling an agentic system in at least one function, plus 39% experimenting. Figures for “agents in production” in 2026 range from single digits to over half, depending on the definition and the sponsor of the survey.
Did Klarna abandon AI customer service?
No. It reversed an automation-heavy approach and began rehiring human agents in 2025, with its CEO citing lower quality, while still using AI. I did not find a verified 2026 update on the current mix.
What does “agent washing” mean?
Gartner’s term for vendors rebranding existing chatbots, RPA or assistants as “agentic AI” without substantial autonomous capability. Gartner estimated about 130 of the thousands of agentic vendors are genuine.
What is the best first step before scaling an AI agent?
Write down the success metric, the evaluation window and the shutdown criteria, and name one accountable owner. Then run the 5-Gate Test in section 7.
How this article was checked
What I did. I searched for and read the primary or closest-to-primary material available on each major claim: the MIT NANDA report as described by multiple outlets and by a detailed critique, McKinsey’s 2025 survey as summarized by several outlets, Gartner’s June 2025 press release as reproduced by multiple publishers, Klarna’s February 2024 press release, and 2025 interviews and reporting on its reversal.
What I did not do. I ran no tests or experiments of my own, and no figure here comes from original data collection. The calculations in sections 1 and 6 are my arithmetic on the cited inputs. I did not have subscriber access to full Gartner or McKinsey reports, so several figures rest on press releases or secondary summaries and are labeled that way.
Known limitations. Several 2026 statistics trace only to vendor blogs or aggregators, and two conflict with each other (section 4). The MIT critique I rely on is commentary rather than peer-reviewed rebuttal. The 5-Gate Test is untested. If you find a primary source that contradicts something here, tell me and I will update the page.
Sources
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025): report PDF (hosted by a third party); summaries: Virtualization Review, Let’s Data Science
- Fortune, MIT report: 95% of generative AI pilots at companies are failing (Aug 18, 2025)
- Rob Wiblin, 80,000 Hours, The story behind the bad AI stat that moved markets and misled millions (Apr 28, 2026)
- McKinsey / QuantumBlack, The State of AI in 2025: Agents, Innovation, and Transformation (primary page not retrieved); summaries: BIC Magazine, Kanerika
- Gartner press release, Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 25, 2025): reproduced copy
- Gartner, Hype Cycle for Agentic AI 2026 (Apr 2, 2026): secondary coverage
- Klarna, AI assistant press release (Feb 27, 2024)
- Bloomberg, Klarna Turns From AI to Real Person Customer Service (May 8, 2025); coverage: Entrepreneur, CX Today
- WRITER, Enterprise AI adoption in 2026 survey (vendor survey)
