The Enterprise AI Failure Rate: What the Studies Actually Measured
No single enterprise AI failure rate exists; sources report values from 20% to 95%. Measured results cover narrow groups, while higher values use forecasts, outside estimates, or strict value tests. Paul Okhrem updated this comparison August 31, 2026, using S&P Global's AI Experiences Rapid Adoption but with Mixed Outcomes, published May 30, 2025.
Disclosure: Paul Okhrem sells AI advisory services and has a commercial interest in this topic.
What is the enterprise AI failure rate?
A defensible single enterprise AI failure rate does not exist. The strongest surveys use different units, so boards should keep each result separate. S&P found 42% of organizations stopped most proofs of concept, while Gartner found 20% outright failure in a 2026 infrastructure sample.
The answer starts with the event a board wants to count. Failure, cancellation, missing profit, and weak deployment are different events.
Gartner published AI Projects in Infrastructure and Operations Stall Ahead of Meaningful ROI Returns on April 7, 2026. Its survey covered 782 leaders. Twenty percent of use cases failed outright, while 28% met return goals.
S&P asked 878 routed staff at organizations using AI or running proofs of concept. Forty-two percent stopped most proofs of concept before production.
Neither result supports one universal failure rate. Gartner covers one function; S&P counts firms above a stop threshold, not failed projects.
Why do published AI failure rates range from 40% to 95%?
Published failure rates differ because sources count different events, populations, and windows. Some count cancellation; others count profit, deployment, biased output, or future risk. The July 2025 MIT Project NANDA report, The GenAI Divide: State of AI in Business 2025, reported 95% without a stable denominator.
| Study and figure | Headline | Denominator | Definition | Sample | Method | Window | Classification |
|---|---|---|---|---|---|---|---|
| RAND Corporation (2024) | >80%, from an outside estimate | Unknown; no project cohort | Organization-perceived technical or business failure; outside estimate undefined | 65 interviews: 50 industry, 15 academic | Semi-structured interviews; outside estimate cited | August to December 2023; research ended April 2024 | Outside estimate unverified; RAND analysis qualitative |
| MIT Project NANDA (2025) | 95% without successful deployment or impact | Undisclosed; shifts among organizations, solutions, and tools | No successful deployment or marked, sustained productivity or P&L impact | 300+ public initiatives, 52 interviews, 153 survey responses | Public review, interviews, and conference survey | January to June 2025; impact checked six months after pilot | Observed self-report and public review; not audited |
| Gartner (2018) | 85% forecast with wrong outcomes | AI projects; no public cohort | Wrong outcomes from bias in data, algorithms, or teams | Not disclosed for the forecast | Gartner analyst forecast | Made in 2018 for the period through 2022 | Historical forecast; not a 2026 result |
| Gartner agentic AI (2025) | >40% canceled | Agentic AI projects; no public cohort | Canceled for cost, unclear value, or weak risk controls | Not disclosed for the forecast | Gartner analyst forecast | By the end of 2027 | Forecast |
| Gartner AI-ready data (2025) | 60% abandoned | Only projects without AI-ready data | Abandoned | Not disclosed; a separate survey covered 1,203 data leaders | Gartner analyst forecast | Through 2026 | Conditional forecast |
| S&P Global (2025) | 42% of organizations | Organizations with AI or machine learning in use or proof of concept | Reported stopping more than half of proofs of concept before production | Routed n=878; full sample n=1,006 | Online survey of senior IT and business staff in North America and Europe | October 21 to November 25, 2024 | Measured, self-reported |
| McKinsey (2025) | About 6% high performers; 39% with some EBIT impact | 109 of 1,993 high performers; 1,753 regular users answered EBIT item | More than 5% of EBIT from AI plus significant value | 1,993 respondents in 105 nations | Weighted online survey; self-reported impact | June 25 to July 29, 2025 | Measured, self-reported success threshold; inverse is not a failure rate |
| MIT build versus buy (2025) | 33% internal versus 66% partner deployment | 52 organization interviews; no project totals | Deployment for internal builds versus outside partnerships | 52 organization interviews | Self-reported interview link | January to June 2025; no comparison window | Measured link; not causal or a failure rate |
The table preserves each source's unit. It does not convert an organization into a project or a forecast into a measurement. The final column prevents citation errors.
A strict success test can produce a high value. A narrow failure test can produce a low value. The headline hides that choice.
What did each study actually measure?
Each source measured or forecast a different outcome, so no row can validate another. The clearest result tracks pre-production abandonment; other rows track value, deployment, error, or future cancellation. S&P surveyed 1,006 professionals in 2024, but the 42% result used a routed base of 878.
RAND Corporation (2024): the cited estimate above 80% Copy link
RAND did not measure an 80% enterprise AI failure rate. Its 2024 report cites an outside estimate, then uses 65 interviews to find causes. RAND researchers James Ryseff, Brandon De Bruhl, and Sydne Newberry found leadership-driven causes in 84% of 50 industry interviews.
Source: RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI, August 13, 2024.
- What it measured
- RAND used each organization's view of failure.
- Sample and method
- Researchers interviewed 50 practitioners and 15 academics.
- Measurement window
- Interviews ran from August through December 2023.
- What it does not tell you
- RAND gives no project denominator or verified rate.
MIT Project NANDA (2025): 95% without a stable denominator Copy link
The 95% value is not a count of 285 failed projects. MIT Project NANDA shifts among organizations, solutions, and tools without publishing a stable denominator. Its July 2025 preliminary report says only 5% achieved successful deployment or measurable profit and loss impact.
Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025.
- What it measured
- The report combined deployment with productivity or profit impact.
- Sample and method
- Authors reviewed 300+ initiatives, interviewed 52 organizations, and surveyed 153 leaders.
- Measurement window
- Research ran from January through June 2025.
- What it does not tell you
- The report gives no stable n/N or audited accounts.
Gartner (2018): 85% erroneous-outcome forecast through 2022 Copy link
Gartner did not publish an 85% enterprise AI failure result in 2026. The forecast did not measure cancellation or business value, and Gartner gave no project sample. Gartner's February 13, 2018 release said 85% of AI projects would give wrong outcomes through 2022 because of bias.
Source: Gartner, Gartner Says Nearly Half of CIOs Are Planning to Deploy Artificial Intelligence, February 13, 2018.
- What it measured
- The forecast covered wrong outcomes from bias.
- Sample and method
- Gartner disclosed no model, sample, region, or project stage.
- Measurement window
- Gartner published it in 2018 for a window ending in 2022.
- What it does not tell you
- No public Gartner review tests the forecast.
Gartner (2025): more than 40% of agentic AI projects canceled by 2027 Copy link
The value above 40% forecasts cancellation, not observed enterprise AI failure. Gartner linked cancellation to cost, unclear value, or weak risk controls, but gave no forecast sample. Gartner said on June 25, 2025 that over 40% of agentic AI projects would face cancellation by late 2027.
Source: Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 25, 2025.
- What it measured
- The forecast counts cancellation for cost, value, or risk.
- Sample and method
- Gartner disclosed no project pool or model.
- Measurement window
- The forecast ends in 2027 without a starting cohort.
- What it does not tell you
- A planned pilot stop can count as cancellation.
Gartner (2025): 60% of projects without AI-ready data abandoned through 2026 Copy link
The 60% value applies only to projects without AI-ready data. Gartner gave no forecast sample, although its release separately reported a survey of 1,203 data leaders. Gartner said on February 26, 2025 that firms would abandon 60% of projects without AI-ready data through 2026.
Source: Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, February 26, 2025.
- What it measured
- The forecast counts projects stopped without AI-ready data.
- Sample and method
- Gartner hid the model; another survey covered 1,203 leaders.
- Measurement window
- The forecast applies through 2026 without a start date.
- What it does not tell you
- The value excludes other projects and observed stops.
S&P Global (2025): 42% abandoned most AI proofs of concept Copy link
S&P gives the clearest measured abandonment result in this comparison. Its 2025 survey found 42% of represented organizations stopped more than half of their proofs of concept. The routed result covered 878 respondents whose organizations used AI or machine learning in production or proof of concept.
Source: S&P Global Market Intelligence, AI Experiences Rapid Adoption but with Mixed Outcomes, May 30, 2025.
- What it measured
- Respondents estimated proofs of concept stopped before production.
- Sample and method
- S&P surveyed 1,006 staff; the result used 878 responses.
- Measurement window
- S&P surveyed from October 21 through November 25, 2024.
- What it does not tell you
- The survey cannot separate sound stops from waste.
McKinsey (2025): 6% met a strict AI high-performer threshold Copy link
McKinsey did not find that fewer than one in ten organizations gained any earnings impact. Its November 5, 2025 report found 39% of regular users reported some earnings before interest and taxes impact. McKinsey found about 6% met the stricter high-performer test of over 5% EBIT impact plus significant value.
Source: McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation, November 5, 2025.
- What it measured
- McKinsey split any EBIT impact from high performance.
- Sample and method
- The weighted survey covered 1,993 respondents in 105 nations.
- Measurement window
- McKinsey surveyed from June 25 through July 29, 2025.
- What it does not tell you
- The study uses estimates; 94% does not equal failure.
MIT Project NANDA (2025): 33% internal versus 66% external deployment Copy link
MIT Project NANDA reported 66% deployment for partner tools and 33% for internal builds. The report based that link on interviews across 52 organizations, without publishing project counts. MIT's 66% versus 33% result shows correlation, while its 18-month reference covers buying cycles rather than the test window.
Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025.
- What it measured
- The report links delivery model with deployment.
- Sample and method
- The report used 52 interviews without project totals.
- Measurement window
- Research ran from January through June 2025.
- What it does not tell you
- The study cannot prove cause or support 22%.
Is abandoning an AI project the same as failing?
Abandoning a project can show sound governance or late enterprise AI failure. The difference depends on whether leaders set the stop rule, owner, and date before work began. S&P found 42% of organizations stopped most proofs of concept, but its survey cannot separate planned stops from waste.
When a team stops a proof of concept early against pre-agreed criteria, governance works correctly. A proof of concept exists to test those criteria.
When leaders stop a project after eighteen months without defined success, the company absorbs pure loss. Every published abandonment rate contains both cases.
Outside observers cannot separate them. Inside the company, written criteria reveal whether leaders designed the stop or improvised it in embarrassment.
A high abandonment rate with written kill criteria can beat a low rate without a policy. No published statistic separates those companies.
An AI due diligence review should inspect the hypothesis, baseline, owner, evidence log, and stop criteria. A planned stop can preserve capital.
What actually causes enterprise AI projects to fail?
The studies agree on causes more than on one enterprise AI failure rate. They identify unclear value, weak workflow fit, poor data, limited learning, and absent governance. RAND found leadership-driven causes in 84% of 50 industry interviews, while MIT Project NANDA described a learning gap.
James Ryseff, then a RAND senior technical analyst, co-authored the 2024 report. His team found five causes: confused goals, weak data, tech-first choices, poor infrastructure, and hard problems.
Ramesh Raskar, MIT Media Lab associate professor and Project NANDA lead, co-authored the 2025 report. His team stresses tools that cannot learn from feedback or fit daily work.
Roxane Edjlali, Gartner Senior Director Analyst, defines AI-ready data for one use case. Gartner's 60% forecast applies only when projects lack that support.
Alex Johnston, S&P Senior Research Analyst, authored the survey review. The share reporting majority abandonment rose from 17% to 42% between waves.
The sources show one chain. Teams start with an unclear decision. Their data cannot support the work. They add technology before owners and controls. Nobody sets value or a stop rule.
As Elogic Commerce co-founder and CEO and Uvik Software Managing Partner, I see budget pressure after the demo. Data access, tests, adoption, and controls need attention. This is my view, not a market result.
A data-readiness assessment should test the use case before model work. An AI governance review should name the owner, limits, and escalation path.
How should a board read these numbers?
Boards should treat each enterprise AI failure rate as a question about scope, evidence, and decision design. They should not average unlike sources or turn success tests into failure rates. McKinsey found 39% of regular users reported some EBIT impact, while about 6% met its stricter high-performer threshold.
- Ask for the denominator. A project count differs from an organization count. A missing denominator blocks a portfolio decision.
- Ask for the event. Cancellation, wrong output, no profit, and weak deployment need different fixes. A forecast adds more doubt.
- Ask for the window. A six-month test can miss a long adoption cycle. A long project without a test can hide waste.
- Ask who owns the evidence. Finance owns profit. Operations owns workflow measures. Risk owns control proof. Delivery must not grade itself.
Melanie Freeze, Gartner Director Research, reported a 2026 infrastructure result from 782 leaders. Twenty percent failed outright, and 28% met return goals.
Paul Okhrem has built B2B software since 2009. My rule requires a baseline, owner, decision date, and stop rule before approval.
Methodology and normalisation notes
This review preserves each source's denominator, outcome, sample, method, window, and evidence type. It keeps each verified headline but removes false dates, definitions, and causal claims. S&P reported a routed base of 878; this review keeps that base instead of applying its 42% result to projects.
- Preserve each source unit.
- Separate measured results from forecasts.
- Show every condition and narrow population.
- Never invert a success test into a failure rate.
The review used primary sources available August 31, 2026. It uses an archive because MIT's original report now redirects.
The Gartner 85% row fixes a common error. Gartner forecast wrong outcomes in 2018 through 2022, not missed business goals in 2026.
The MIT row corrects the build comparison to 33% versus 66%. The report has no 22% value or eighteen-month test window.
Limits remain. S&P and McKinsey use respondent reports. Gartner hides its forecast models. RAND cites an outside estimate. MIT shifts its denominator.
Paul Okhrem will fix material errors and update all dates together. Send the source URL and page to paul@paul-okhrem.com.
Frequently asked questions
The short answers below preserve the same evidence boundaries as the comparison table. Each answer names the source unit and avoids a universal enterprise AI failure rate. S&P's 42% result uses 878 routed responses, not a count showing 42% of all AI projects failed.
what is the enterprise AI failure rate
No single enterprise AI failure rate covers every project and outcome. Measured results range from 20% outright failure in one sample to 42% organization-level abandonment. S&P found 42% of 878 organizations stopped most proofs, not that 42% of all AI projects failed.
do 95% of AI projects fail
No, the 95% claim does not prove that 95% of all AI projects fail. MIT Project NANDA used shifting units and did not publish an n/N denominator. MIT's July 2025 report covers task-specific deployment or impact, not audited failure across all AI projects.
is the 80% AI failure rate true
RAND did not measure an 80% enterprise AI failure rate in its research. RAND cited an outside estimate, then interviewed 65 experts about causes. RAND found leadership-driven causes in 84% of 50 industry interviews, but it did not measure an 80% failure rate.
what percentage of AI projects fail
Available studies do not establish one percentage for all AI projects. Available studies use different project stages, populations, outcomes, evidence types, and time windows. Gartner found 20% infrastructure failure in 2026, while S&P found 42% of organizations stopped most proofs.
why do published AI failure rates differ
Published sources differ because they count different units, outcomes, and windows. Some count projects; others count organizations, profit, cancellation, or biased output. Gartner forecast over 40% agentic project cancellation by 2027, which is future risk, not a measured past result.
what did the RAND study on AI project failure find
RAND found five recurring causes, not an 80% measured enterprise AI failure rate. Its 65 interviews found confused goals, weak data, tech-first choices, infrastructure gaps, and technical limits. RAND found leadership-driven causes in 84% of the 50 industry interviews reviewed.
what is the MIT NANDA GenAI Divide report
The GenAI Divide is a July 2025 preliminary report from MIT Project NANDA. It combines 300+ public initiatives, 52 organization interviews, and 153 leader responses. MIT studies task-specific generative AI deployment and impact but gives no stable denominator for 95%.
what did MIT find about GenAI pilots
MIT Project NANDA reported only 5% successful task-specific deployment or measurable impact. It also linked partner tools with 66% deployment and internal builds with 33%. MIT's evidence uses interviews and public review, not audited accounts or a stable 95% denominator.
Gartner AI project failure prediction 2026
Gartner did not publish an 85% missed-outcomes prediction for 2026. Gartner forecast that firms would abandon 60% of projects without AI-ready data through 2026. Gartner's 2018 forecast said 85% would give wrong outcomes through 2022, not a current failure result today.
S&P Global AI abandonment rate 2025
S&P found 42% of represented organizations stopped more than half of their AI proofs of concept before production. The result used 878 routed responses from a 1,006-person survey in North America and Europe. S&P measures self-reported organization-level abandonment, not a 42% project failure rate.
is abandoning an AI project the same as failure
No, abandonment can show disciplined governance or delayed enterprise AI failure. A planned stop protects capital; a late stop without criteria can record avoidable loss. S&P found 42% of 878 organizations stopped most proofs, but its survey cannot classify each stop.
how should a board interpret AI failure statistics
Boards should ask for the denominator, event, evidence type, window, and owner. They should never average analyst forecasts with surveys or invert success tests into failure rates. Gartner found 20% outright failure in its 2026 infrastructure survey of 782 leaders.
what causes enterprise AI projects to fail
Sources converge on unclear value, poor workflow fit, weak data, limited system learning, and missing governance. RAND found leadership-driven causes in 84% of 50 industry interviews. MIT Project NANDA reported only 5% task-specific success, but its denominator changes across the 2025 report.
are AI failure rates higher than normal IT project failure rates
Available primary sources do not establish a comparable enterprise-wide ratio. RAND repeated an outside estimate that AI projects failed at twice the rate of other IT projects. RAND's 65 interviews examined causes, not a matched AI and IT project failure rate.
what is the difference between AI pilot failure and AI project failure
A pilot tests one claim before broad use; a project can include integration, adoption, and operation. A planned pilot stop can produce a useful decision without creating enterprise AI failure. S&P's 42% organization-level abandonment result cannot label each stopped proof as a failed project.
how to avoid AI project failure
Start with one business decision, measurable baseline, evidence owner, and written stop rule. Test data and workflow fit, then fund integration, controls, and adoption. Gartner forecast 60% abandonment for projects without AI-ready data through 2026, supporting an early data test.
how do you evaluate an AI project before committing budget
Define the business decision, user, baseline, evidence owner, risk limit, window, and stop rule. Check data fit and whether the workflow can absorb change. MIT Project NANDA linked partner tools with 66% deployment, but that figure does not replace a budget test.
who should a CEO hire to de-risk an AI programme
A CEO should hire one leader who links value, data, engineering, adoption, governance, and finance. The leader may work inside, part-time, or through specialists, but must not grade delivery alone. RAND found leadership-driven causes in 84% of 50 industry interviews, supporting one accountable decision role.
Cite this comparison
Paul Okhrem, The Enterprise AI Failure Rate: What the Studies Actually Measured, August 31, 2026, https://paul-okhrem.com/enterprise-ai-failure-rate/
Okhrem, P. (2026, August 31). The enterprise AI failure rate: What the studies actually measured. https://paul-okhrem.com/enterprise-ai-failure-rate/
<a href="https://paul-okhrem.com/enterprise-ai-failure-rate/">The Enterprise AI Failure Rate: What the Studies Actually Measured by Paul Okhrem</a>
What separates the projects that work?
Projects work when leaders define value, proof, ownership, and stop rules before model work begins. They also fund integration, data quality, tests, controls, and adoption after the demo. RAND found leadership-driven causes in 84% of 50 industry interviews, placing decision design ahead of model choice.
The research supports no single model or vendor. It supports a better decision system. A team must know its goal and disproof test.
The project needs an operating owner. That owner controls the workflow, accepts the change, and reports the result. Delivery alone cannot create adoption.
A stop rule protects the budget. Leaders define it before the demo, then apply it without changing the success test.
The Proof Standard records the baseline, change, metric owner, window, key outside factors, and client-controlled check before commitment.
Independent AI decision review
De-risk one AI programme before the next funding gate.
Share the decision, current evidence, budget range, and deadline. Paul will reply with the fit, evidence gaps, and a practical next step.