Short answer: AI projects usually fail for organizational reasons, not technical ones. The usual causes are the wrong problem, no owner, no baseline, data that is not ready, a pilot beside the real work, no agreed approval point and running costs nobody modeled. Each one has a prevention step you can check before you approve the budget.
AI projects usually fail for organizational reasons, not technical ones. The research keeps pointing to the same seven causes:
the project targets the wrong problem;
nobody owns it;
success is never measured against a baseline;
the data is not ready;
the pilot runs beside the real work instead of inside it;
people do not trust the output;
running costs arrive as a surprise.
Every one of these can be checked before you sign off a budget. Below is what the main studies actually found, the seven causes with a prevention step for each, and a one-page pre-mortem you can run in your next meeting.
What does the research actually say about AI project failure?
Several failure rates circulate, and they measure different things. Here is what each source is, and what it is not.
RAND (2024): “more than 80 percent.” RAND’s report on the root causes of AI project failure is the usual source for this number, but it is not RAND’s finding. The report introduces it with “by some estimates” and cites a 2022 Fortune article. RAND’s own work is qualitative: interviews with 65 experienced data scientists and engineers, conducted between August and December 2023, about projects that trained or customized models. Projects that only prompted a ready-made language model were out of scope. Read RAND for causes, not for a rate.
Gartner (2024 to 2026): predictions, then an estimate. In July 2024 Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. In January 2026 a Gartner article said that by the end of 2025, at least 50% had been abandoned after proof of concept. Gartner says it analyzed hundreds of implementations but does not publish the method on that page. In June 2025 it predicted that over 40% of agentic AI projects will be canceled by the end of 2027. These are analyst judgments, not audited counts.
S&P Global / 451 Research: a survey. In 451 Research’s survey of 1,006 respondents in North America and Europe, fielded in late 2024, the share of companies abandoning most of their AI initiatives before production rose from 17% to 42% in a year. The average organization scrapped 46% of its proofs of concept. Self-reported, but the sample is large.
McKinsey State of AI 2026: the value gap. In McKinsey’s 2026 survey of 1,719 participants, 37% attribute at least some EBIT impact to AI, essentially unchanged from a year earlier. This is less a failure rate than a measure of how few companies see results on the income statement.
MIT NANDA (2025): the 95% headline. The famous claim that “95% of organizations are getting zero return” comes from preliminary research built on interviews with 52 organizations, 153 survey responses and a strict definition of success. We explain why the number is weaker than its fame in The model is no longer the advantage.
So what is the real failure rate? Nobody knows precisely, and it depends on what counts as failure. A fair summary: in the better-documented sources, somewhere between a third and a half of AI pilots never reach production, and fewer still show up in the financials. The more useful finding is that the sources largely agree on why.
1. The project solved the wrong problem
RAND’s first root cause is that stakeholders misunderstand, or miscommunicate, what problem AI should solve. Models get optimized for the wrong metric, or do not fit the workflow they were meant for.
RAND also describes the reverse error: leaders asking for machine learning on problems that a few if-then rules would have handled, and technical teams (16 of 50 interviewees raised this) choosing new tools because they are new. RAND’s advice is blunt: “Successful projects are laser-focused on the problem to be solved, not the technology used to solve it.”
Gartner’s 2026 article names lack of business value as the most fundamental failure mode. On agents, Gartner’s Anushree Verma put it plainly: “Many use cases positioned as agentic today don’t require agentic implementations.” Gartner estimates only about 130 of the thousands of agentic AI vendors are real.
Prevent it: write the problem as a workflow with a start, an end and a number that should move. “Supplier invoices are matched against orders in under a day” is a problem. “AI for finance” is not. Then ask whether a rule, an existing software feature or a process change would do. Our guide to choosing your first AI workflow has a scoring method for this step.
2. Nobody owned it
In RAND’s interviews, failures driven by leadership decisions and expectations were “far and away” the most frequent. Eighty-four percent of interviewees named one or more of them as the primary reason projects fail. The patterns are familiar: priorities that switch every few weeks, projects dropped before they can show results, and leaders who expect weeks when the work takes months (14 of 50 interviewees reported this). One caveat: RAND notes that most of its interviewees were engineers, not executives, which may tilt the findings toward blaming leadership.
McKinsey’s 2026 survey points the same way from the other side. The roughly 6% of respondents it calls AI high performers are twice as likely as others to say their senior leaders demonstrate commitment to AI initiatives.
Prevent it: name one business owner, not an IT contact and not a committee. That person needs time in their calendar, the authority to change the process, and a seat at the go-live decision. The sponsor commits to a fixed period and a review date. RAND suggests committing a team to one problem for at least a year. For a mid-sized company the build may take weeks, but ownership has to outlast it, because someone runs the workflow after launch.
3. Nobody measured the starting point
A project without a baseline cannot prove anything, so it loses the first budget argument. Gartner’s 2026 article says projects without clear success metrics become easy targets when budgets tighten. McKinsey finds high performers twice as likely to have defined processes to measure the impact of their AI initiatives.
More metrics are not the answer. 451 Research found companies tracking significantly more KPIs year over year without their projects performing better against them.
Prevent it: before anything is built, record two to four weeks of normal work on four numbers: cycle time, hands-on minutes per case, rework or error rate, and backlog. Then write down two thresholds: the result that justifies continuing, and the result that stops the project. Agree on both before the pilot starts, when nobody is attached to it yet.
4. The data was not reachable or not ready
Data problems were the second most common root cause in RAND’s interviews, and business leaders were often surprised to learn their organization lacked the data a model needed. Gartner surveyed 1,203 data management leaders in July 2024: 63% of organizations either do not have, or are unsure whether they have, the right data management practices for AI.
Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. In 451 Research’s survey, 54% of companies with below-average failure rates considered data availability when prioritizing use cases, against 42% of those with above-average failure rates.
In most business workflows, the data is documents: contracts, SOPs, product information, emails. The problem is rarely volume. It is access, permissions and knowing which version is the approved one.
Prevent it: before approval, ask someone to pull 20 to 50 real cases from the actual systems the workflow will use. Confirm the access path, the permissions and the data protection basis. If getting the documents needs a new contract or an IT project, the project waits until that is done.
5. The pilot ran beside the work, and the team was never brought along
Many pilots live in a sandbox: a separate tool, a demo dataset, a few enthusiasts. Then they meet the real process and stall. RAND found trained models being deployed that do not fit the business workflow around them. MIT NANDA blames brittle workflows and misalignment with day-to-day operations. Gartner lists poor change management as a top failure point and advises making AI native to existing workflows rather than adding new tools.
People matter as much as plumbing. In 451 Research’s survey, 28% of respondents named staff resistance as a challenge. Companies with above-average failure rates were 36% more likely than the overall sample to cite it.
Prevent it: build the workflow into the tools the team already uses, such as the inbox, the ERP or the document system. Train people on the new workflow, not just the tool. Run it alongside the old process for a short period and compare outputs. Tell the team what to do with the time it saves.
6. Nobody decided where a person approves, so nobody trusted it
Trust is a design question. In 451 Research’s survey, confidence in accuracy tied with budget as the most common challenge, at 29%. McKinsey’s 2025 survey found that 51% of respondents from organizations using AI had seen at least one negative consequence, with nearly one-third reporting consequences from AI inaccuracy. The same survey found high performers more likely to have defined processes for when model outputs need human validation. Gartner adds that treating responsible AI as an afterthought leads to project shutdowns, and that risk failures worry boards in particular.
Prevent it: write one page with four headings. What the system does. What a person decides. When it escalates. Who can stop it, and how. Build a test set of past cases with known correct answers and measure the error rate before go-live, so the team knows what to check. In regulated work, sector rules and data protection apply on top; see our post on the EU AI Act for the general picture. This is general information, not legal advice.
7. The running costs were never modeled
Pilots are cheap. Production is not always. In July 2024 Gartner’s Rita Sallam warned that generative AI costs are less predictable than those of other technologies, and Gartner put the cost of some enterprise-scale deployment approaches at $5 million to $20 million. Most mid-sized projects are far smaller (see what AI implementation costs in 2026), but the pattern holds at every size.
Gartner’s 2026 article describes projects that looked viable in proof of concept turning into budget problems in production. McKinsey’s 2026 survey found one in five respondents saying their organization limits AI use because of operating costs. In 451 Research’s survey, budget constraints rose from fourth place to joint first among challenges.
The model bill is often not the largest line. A hypothetical example: a drafting workflow handles 1,500 cases a month, and a person spends three minutes reviewing each draft. That is 75 hours of review a month, before any maintenance of prompts, integrations and model updates.
Prevent it: before approval, estimate cost per case at full volume: model usage, review time, vendor fees and maintenance hours. Set a ceiling. Check actual cost per case monthly for the first quarter after launch.
Is a failed AI pilot program always a bad sign?
No. A pilot exists to find out cheaply whether something works, and a “no” is a valid answer. Gartner’s 2024 release says that if results fall short, it may be necessary to explore alternatives. RAND calls experimenting with new technology important, as long as it is a conscious choice balanced against the organization’s other goals. And 451 Research found that time-to-deliver did not correlate with failure rates, so racing to a demo does not protect a project.
The abandonment statistics count two very different things the same way. Healthy failure is cheap, time-boxed and stopped by criteria agreed in advance, and it leaves something behind: cleaner data, a test set, a clearer view of where AI fits. Unhealthy failure drifts for months, nobody decides, and nothing is written down.
Our view: a company that never stops a pilot is probably not testing honestly. What matters is how much each stop costs, and whether it was planned.
The pre-mortem: what to check before approving an AI project
A pre-mortem takes 45 minutes. Gather the owner, someone who does the work, someone from IT and whoever controls the budget. Ask everyone to imagine it is nine months from now and the project has failed. Each person writes down why. Then go through the checks below. Any “no” on owner, data or approval means the project is not ready for approval yet.
Check | Red flag if… |
|---|---|
Can we describe the workflow in one sentence, with a start and an end? | The scope is a department or a technology |
Have we ruled out a rule, an existing feature or a process change? | Nobody has asked |
Is one named business owner accountable, with time and authority? | It is a committee, or IT alone |
Do we have two to four weeks of baseline numbers? | “We’ll measure after launch” |
Have we written down the result that continues it and the result that stops it? | No stop criteria, no review date |
Has someone pulled 20 or more real cases from the real systems? | Access needs a new contract or IT project |
Will the output appear in a tool the team already uses? | A separate app people must remember |
Is it written down what a person approves, when it escalates and who can stop it? | “The AI handles it” |
Do we have past cases with known answers to test against? | Accuracy judged by demo |
Are training and a parallel run in the plan? | Launch is an email |
Do we know the cost per case at full volume, including review time? | Only the license price is known |
Who measures results in month three, and who maintains it? | Nobody after the vendor leaves |
Where to start
Start with one workflow that passes every check above, not five that pass half of them. Our guide on how to choose your first AI workflow walks through the scoring, the baseline and the approval page. If you are weighing whether to build or license the system, read build vs buy AI.
About to approve a project? Book a fit check and bring the pre-mortem. We’ll tell you whether it’s worth a diagnostic.
Frequently asked questions
What share of AI projects fail?
Nobody knows precisely, because the sources count failure differently. In the better-documented sources, somewhere between a third and a half of AI pilots never reach production, and fewer still show up in the financials. The widely quoted figure of more than 80 percent is not RAND’s own finding.
What is the most common cause of AI project failure?
In RAND’s interviews, failures driven by leadership decisions and expectations were the most frequent. Eighty-four percent of interviewees named one or more of them as the primary reason. The prevention is one business owner with time, the authority to change the process and a seat at the go-live decision.
How do you prevent an AI project from failing?
Check the seven causes before approval. Write the problem as a workflow with a number that should move, and name one owner. Record a baseline and pull 20 to 50 real cases from the actual systems. Build inside existing tools, write down where a person approves and estimate cost per case at full volume.
Is a failed AI pilot always a bad sign?
No. A pilot exists to find out cheaply whether something works, and a no is a valid answer. Healthy failure is cheap, time-boxed and stopped by criteria agreed in advance. Unhealthy failure drifts for months while nobody decides.
What is an AI project pre-mortem?
A 45-minute meeting before approval. The owner, someone who does the work, someone from IT and the budget holder imagine the project has failed nine months from now and write down why. Any no on owner, data or approval means the project is not ready yet.
Related reading
Sources
RAND Corporation, James Ryseff, Brandon F. De Bruhl and Sydne J. Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RR-A2680-1, August 13, 2024. Interviews with 65 data scientists and engineers.
Gartner, Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, press release, July 29, 2024.
Gartner, Arun Chandrasekaran, Why 50% of GenAI Projects Fail, January 26, 2026.
Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release, June 25, 2025.
Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, Q&A with Roxane Edjlali, February 26, 2025.
S&P Global Market Intelligence (451 Research), Alex Johnston, Generative AI shows rapid growth but yields mixed results, October 27, 2025. Online survey of 1,006 respondents, October 21 to November 25, 2024.
McKinsey & Company, The state of AI in 2026, August 25, 2026. Online survey of 1,719 participants in 97 nations.
McKinsey & Company, The state of AI in 2025: Agents, innovation, and transformation, November 5, 2025. Online survey of 1,993 participants in 105 nations.
MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Preliminary findings.