Quick answer (updated August 2026): AI projects fail because tools never fit into real workflows, not because the models are weak. The verified numbers: MIT Project NANDA’s preliminary research found 95% of organizations getting zero return from enterprise GenAI pilots, S&P Global found 42% of companies abandoning most AI initiatives, and McKinsey found more than 80% seeing no EBIT impact. Every study converges on the same root cause: workflow fit and integration, which is exactly the part a small business can control.

What the failure studies actually say

Six research organizations have put numbers on AI failure, and almost every version you see quoted online garbles at least one of them. Here is what each study measured, who it measured, and the catch that headlines drop.

Finding Source and date Who was studied The catch
95% of organizations getting zero return from GenAI pilots MIT Project NANDA, The GenAI Divide (July 2025) 52 interviews, 153 leader surveys, 300+ public AI initiatives, Jan–Jun 2025 Preliminary, not peer-reviewed; authors call figures “directionally accurate”
42% of companies abandoned most of their AI initiatives, up from 17% S&P Global / 451 Research (fielded Oct–Nov 2024, reported 2025) 1,006 companies, North America and Europe Measured, but enterprise-weighted; “most initiatives,” not all
“By some estimates, more than 80 percent of AI projects fail” RAND Corporation (Aug 2024) Interviews with 65 experienced data scientists and engineers The 80% is RAND quoting others’ estimates, not RAND’s measurement
More than 80% of respondents see no tangible EBIT impact from gen AI McKinsey, The State of AI (survey July 2024, published Mar 2025) ~1,491 global survey participants No profit impact yet is not the same as project failure
Only 26% of companies can move beyond proofs of concept to tangible value BCG (Oct 2024) 1,000 C-suite and senior executives, 59 countries Capability gap, not a failure count
At least 30% of GenAI projects abandoned after proof of concept by end of 2025 Gartner (July 2024) Analyst prediction A prediction, never a measurement

Notice what the populations have in common: enterprises running custom pilots. None of these studies measured whether a 12-person business gets value from a proven off-the-shelf tool.

Two of the findings deserve their exact wording. S&P Global’s 451 Research survey of 1,006 companies is the one measured, non-preliminary abandonment figure: “The proportion of companies that abandon most of their AI initiatives has increased from 17% to 42%, with the average organization scrapping 46% of its proof-of-concept projects prior to production.” And McKinsey’s finding is about profit, not projects: “More than 80 percent of respondents say their organizations aren’t seeing a tangible impact on enterprise-level EBIT from their use of gen AI,” even though 78% report using AI in at least one business function.

Read the fine print on the 95%

The MIT number deserves the closest reading because it travels the furthest. The report’s exact sentence: “Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.” Even that sentence needs a footnote: the $30–40 billion is the report’s own estimate, asserted without a source.

The report is honest about its limits, and quoting that honesty is the credible way to use it. Its own caveat: “These figures are directionally accurate based on individual interviews rather than official company reporting.” It also concedes its six-month observation window may understate success for longer implementations, and that gains from employees using personal AI tools never count toward the ROI figure.

Outside critics went further. Wharton professor Kevin Werbach noted “there appears to be no further support for the 95% claim” inside the document itself, and Marketing AI Institute’s Paul Roetzer was blunter: “This is not a viable, statistically valid thing.” The report is a preliminary, interview-based snapshot. It still matters, because four independent studies found the same direction.

The misquotes, corrected

Because we cite these numbers in our own work, we keep a list of the circulating versions that do not check out.

“MIT proved 95% of AI pilots fail.” Triple overreach. It is a preliminary, non-peer-reviewed report from Project NANDA, an MIT Media Lab-affiliated initiative, not “an MIT study.” The report says 95% of organizations got zero return, never that 95% of pilots failed. And per Werbach, the 95% cannot be reconstructed from the report’s own exhibits.

Fortune’s methodology numbers. Fortune’s August 2025 coverage described “150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public AI deployments.” The report’s own appendix says 52 interviews, 153 leader surveys, and 300+ public initiatives. Use the appendix.

“Gartner found 85% of AI projects fail.” Traces to a 2018 Gartner prediction about analytics projects producing erroneous outcomes. It was never a measured AI failure rate. Gartner’s real, current claim is the 30% abandonment prediction above.

“RAND found 80% of AI projects fail.” RAND wrote “by some estimates,” passing through a figure from journalism. RAND’s contribution is its root-cause interviews, not the 80%.

The root cause every study converges on

Strip away the disputed percentages and the studies agree on something more useful than any single number: what kills these projects. RAND’s interviewees put leadership-driven failure first, meaning stakeholders “misunderstand—or miscommunicate—what problem needs to be solved,” so teams ship models that optimize the wrong metric or “do not fit into the overall business workflow and context.” The MIT NANDA report reaches the same verdict from the other direction: the divide between winners and losers is not model quality but integration, whether the tool learns the organization’s actual workflow or sits beside it.

NANDA’s most practical finding follows from that. In its sample, externally purchased or partnered tools reached deployment about 67% of the time versus roughly 33% for internal builds. The figures are self-reported, and the report itself cautions that the correlation does not prove causation. We walk through what that means for a 15-person company in our build vs buy guide. The short version: the failure literature is an argument about fit, and fit is decided before any tool is purchased.

What this means at small-business scale

First, the honest disclaimer: none of these studies is about you. They measured enterprise pilots with six-figure budgets and integration teams. A small business subscribing to a proven scheduling or drafting tool faces none of that machinery, which is one reason the enterprise numbers should never be recited as “95% of AI tools fail for small businesses.”

But the mechanism transfers, and the small-business data shows a milder version of the same gap. The Federal Reserve Banks’ 2025 Small Business Credit Survey (6,525 employer firms) found 46% of small businesses use AI while just 7% of those AI users have fully integrated it. Broad adoption, rare integration: the enterprise pattern at owner scale. Our owner’s field guide covers what healthy adoption looks like; the rest of this article covers the five ways it goes wrong. At owner scale the failure modes are cheaper, quieter, and entirely avoidable, because the fix is discipline rather than infrastructure.

The five failure modes at owner scale

1. Buying before mapping

The owner-scale version of RAND’s top root cause: picking a tool before naming the problem. A subscription bought on enthusiasm gets pointed at whatever work is nearest, not the work that costs the most hours. The antidote is a week of honest measurement. Track where your and your team’s hours actually go, then rank the workflows by payback before evaluating a single vendor. Our guide to what to automate first sequences that ranking; the point is that the map precedes the purchase, always.

2. Shipping raw AI output unreviewed

AI drafts confidently and wrongly with the same tone. An unreviewed draft that reaches a customer, a filing, or an invoice converts a time-saver into a liability. The antidote is a standing review gate: AI drafts, a named human signs. The rule costs minutes and survives every model upgrade. If a workflow cannot afford a human check, it is not a candidate for automation yet, whatever the demo looked like.

3. Automating a broken process

Automation is an amplifier. Point it at a process that produces errors, rework, or angry customers and you produce them faster. This is the small-business cousin of the enterprise integration failure: the tool faithfully executes a workflow nobody should have kept. The antidote is sequencing. Redesign the process on paper first, run the fixed version manually for a week, then automate the version that works. Fixing first sometimes reveals the automation is no longer needed, which is the cheapest outcome of all.

4. Tool overwhelm

Three trials running at once, each half-configured, none owned by anyone. This is how the Fed’s broad-adoption, rare-integration gap looks from inside a small company: subscriptions that outlive their use and a team that quietly reverts to the old way. The antidote is a hard cap. One workflow, one tool, one month. Nothing new enters the stack until the current tool is either standard procedure with a written three-line process or cancelled. A confused team does not adopt.

5. Quitting too early

The mirror image of overwhelm: a tool judged and abandoned in week one, before anyone changed a habit. Most of the value of any workflow tool arrives after the routine forms, and routines take weeks, not days. The antidote is a fixed 30-day trial with a baseline measured on day zero, so the verdict is “it saved four hours a month” or “it saved none,” not a feeling. The realistic ROI math only works when there is a before-number to compare against.

Bottom line

The verified failure numbers, read carefully, are enterprise pilot statistics: MIT NANDA’s preliminary 95% zero-return finding (self-described as directionally accurate), S&P Global’s measured 42% abandonment rate, McKinsey’s 80%-plus with no EBIT impact, BCG’s 26% reaching tangible value, and Gartner’s 30% abandonment prediction. Every one of them points at workflow fit, not model quality. That makes them an argument for mapping your work first and buying proven tools second, not for avoiding AI. Mapping is the whole job of an AI opportunity assessment: diagnose which workflows pay back, prescribe tools that already work, and skip the failure modes the studies keep counting.

Sources

  • MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (preliminary, v0.1), July 2025: report PDF
  • S&P Global Market Intelligence, “Generative AI shows rapid growth but yields mixed results,” October 2025: spglobal.com
  • RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, August 2024: rand.org
  • McKinsey & Company, The State of AI: How organizations are rewiring to capture value, March 2025: mckinsey.com
  • Boston Consulting Group, “AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value,” October 2024: bcg.com
  • Gartner, “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025,” July 2024: gartner.com
  • Federal Reserve Banks, 2026 Report on Employer Firms (2025 Small Business Credit Survey), March 2026: fedsmallbusiness.org
  • Fortune, “MIT report: 95% of generative AI pilots at companies are failing,” August 18, 2025: fortune.com (cited for the methodology discrepancy)
  • Futuriom, “Why We Don’t Believe MIT NANDA’s Weird AI Study,” August 2025: futuriom.com (Werbach critique)
  • Marketing AI Institute, on the MIT NANDA study: marketingaiinstitute.com (Roetzer critique)