Why do most AI projects fail, and how do you avoid it?

What does the research say about AI project failure?

Four sources are quoted whenever this question comes up: a RAND Corporation report, an MIT NANDA study, two Gartner predictions and McKinsey's annual State of AI survey. They measure different things, so their numbers cannot be added up or compared directly. Read together, they point in one direction: most AI initiatives do not reach production, and most of those that do are not yet producing a financial result anyone can measure.

Is it true that 95% of AI projects fail?

Not in the way the headline suggests. The MIT NANDA report says that 95% of organizations in its sample were getting zero return from generative AI, and that most enterprise tools stalled before production: 60% of organizations evaluated them, 20% reached a pilot and 5% reached production. Its own methodology note defines success as deployment beyond the pilot with measurable KPIs, with ROI measured six months after the pilot, and calls its figures "directionally accurate" because they come from interviews rather than company reporting.

Critics, including academics who reviewed the report, point out that it is preliminary, not peer reviewed, based on a small self-selected sample and blind to gains that do not show up in a P&L within six months. All of that is fair. It does not make the finding useless. The report's explanation for the stall (tools that do not fit day-to-day workflows, do not learn from feedback and are not integrated into the systems people use) matches what RAND heard from practitioners a year earlier. The same report also found that projects built with external partners reached deployment about twice as often as internal builds. That finding suits firms like ours, so we quote it with the same caution as the rest.

Why do AI projects fail? Seven recurring causes

RAND names five root causes, and the most common one is that the people involved misunderstand or miscommunicate which problem AI should solve. Gartner's predictions add cost, unclear business value and weak risk controls. In our own engagements the same causes show up in a more practical form. Each one below comes with what to do instead.

1. No defined problem and no baseline

The project starts from the technology ("we should use AI") instead of a process with a cost. Without a measured starting point (hours per week, error rate, cycle time) nobody can say later whether the project worked, which is exactly the "unclear business value" Gartner lists. What to do instead: name one process in one sentence, measure it as it runs today and agree the success number before any build starts.

2. Automating a broken process

AI applied to a process that does not work when a person does it produces the same mess, faster and at a higher cost. What to do instead: fix the steps first. Remove the ones that should not exist, then decide what to automate. Often the right first project is process optimization and AI comes after.

3. The data is not ready

RAND's second root cause is that the organization lacks the data needed to make the system work, and Gartner names poor data quality first. In smaller companies the issue is usually access, not volume: the inputs sit in PDFs, inboxes and spreadsheets that no system can read. What to do instead: confirm, on real records, that the data can be exported and reached by a system, and collect around fifty checked examples of a correct output before choosing a model.

4. Nobody owns it

A pilot sponsored by a committee is abandoned the week the sponsor gets busy. What to do instead: one named owner with the authority to change the process, and one person who will supervise the result after launch. If nobody can be named, the project is not ready to start.

5. The pilot is never wired into production

This is the stall MIT NANDA describes: a demo that works in a sandbox but is not connected to the ERP, CRM or ticketing tool where the work happens, so people keep doing the task the old way. RAND's fourth cause, missing infrastructure to deploy and manage models, is the same problem from the engineering side. What to do instead: plan the integration from day one, and treat a pilot as the first version of a production system, not as a separate experiment.

6. No human in the loop

A system that acts on money, customers or legal documents without review fails once, publicly, and is switched off. Gartner lists inadequate risk controls in both of its predictions. What to do instead: define where a person approves or corrects the output, keep an evaluation set to test every change, and loosen review only as the results justify it.

7. Costs were never modeled

Escalating cost appears in both Gartner predictions, and in McKinsey's 2026 survey one in five respondents cited AI operating costs as a constraint on deployment. The usual surprise is not the model bill but the build, the people time and the upkeep. What to do instead: estimate all of it before you start, from real volumes. Our guide to AI implementation cost shows the method.

How do you avoid AI project failure?

Avoiding failure is mostly a matter of order: decide what to change and how to measure it before paying to build it. That is how we structure our work, described on how we work:

• Diagnose first. The Diagnosis maps how the work actually moves, where it stalls and what it costs, and ends in a written case. No code is written during it, and weak use cases are dropped there, where it is cheap.
• Measure the baseline. The success number is agreed before the build, from the process as it runs today, so the result can be checked rather than argued about.
• Ship to production. The first version is connected to the systems the team already uses, with review points and budget limits, so it is used from week one instead of living in a demo.
• Hand it over. The system is documented so the client's own team can run it without us, including the evaluation set and what to do when the model is wrong.

Before any of that, it helps to know whether the company is ready at all. Our AI readiness checklist covers the prerequisites in twelve questions. When the company is ready, our AI integration service applies these four steps to one workflow at a time.

A pre-mortem checklist

A pre-mortem asks the team to imagine the project has already failed and to name why. Run these questions before the budget is approved. Every "no" is a likely cause of failure.

• Can we name the process in one sentence, and do we know what it costs today in hours or errors?
• Does the process work when a person does it?
• Can a system reach the data without someone copying it by hand, and do we have examples of a correct output?
• Is there one named owner, and one person who will supervise the result after launch?
• Is the integration with our existing systems in the plan and the budget, not left for after the pilot?
• Do we know where a person must review the output before it acts?
• Have we estimated the build, the running cost and the upkeep from real volumes?
• Have we agreed the number that decides success, and the date we will check it?

When should you not start an AI project?

When the task runs only a few times a month, when a plain rule or integration would do the job, or when the checklist above has more than two "no" answers. In those cases the useful project is process or systems work, and AI becomes a sensible second step. Saying no early is not a failure. It is the cheapest outcome an AI project can have.

Summary

• The widely quoted failure rates come from different measurements: RAND cites estimates above 80%, MIT NANDA found 95% of organizations with no measurable return from generative AI, Gartner predicts high abandonment rates and McKinsey reports 37% of respondents seeing any EBIT impact.
• The causes are consistent across studies: an unclear problem, data that is not ready and pilots never integrated into real workflows.
• Seven practical causes recur: no baseline, a broken process, unready data, no owner, no production integration, no human review and unmodeled costs.
• Avoid them by diagnosing first, measuring the baseline, building for production from the start and handing over a system your team can run.

Sources

• RAND Corporation, J. Ryseff, B. F. De Bruhl, S. J. Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RR-A2680-1, August 2024.
• MIT NANDA, A. Challapally, C. Pease, R. Raskar, P. Chari, The GenAI Divide: State of AI in Business 2025, July 2025, preliminary findings.
• Gartner, Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, press release, 29 July 2024.
• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release, 25 June 2025.
• McKinsey & Company, The state of AI in 2026: On the road to ROI, 25 August 2026.
• All figures checked against the sources on 7 October 2026.