Why AI projects fail comes down to one thing more than any other: nobody defined the job the tool was meant to do. RAND interviewed 65 data scientists in 2024 and named five root causes, led by a misunderstood problem. A small firm can test for all five before spending anything.
What do the big studies say about why AI projects fail?
The large studies agree on a short list of causes, and most of them concern the job rather than the software. Three sources come up most often, and each measures something different.
RAND’s 2024 report, The Root Causes of Failure for Artificial Intelligence Projects, is based on interviews with 65 data scientists and engineers. Each had at least five years of experience.
RAND cites estimates that more than 80 percent of AI projects fail, twice the rate of IT projects without AI.
Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept. Gartner gave four reasons: poor data quality, inadequate risk controls, escalating costs and unclear business value.
MIT’s NANDA report is the source of the widely quoted 95% figure, which covers enterprise generative AI pilots with no measurable profit-and-loss impact. Its finding on bought versus home-built tools has its own write-up in why internal AI builds fail.
None of these studies isolates firms of 5-20 people. Treat the percentages as a direction of travel and the causes as the useful part.
What are the five root causes RAND found?
RAND found five leading root causes, and the first is the one most likely to hit a small business. The table below restates each in owner terms, with the check that catches it.
| Root cause (RAND) | What it looks like in a 12-person firm | The check before you spend |
|---|---|---|
| Misunderstood or miscommunicated problem | The owner buys a tool “for customer service” and each person assumes a different job | Write the job as one sentence with a number in it |
| Lack of the data to do the job | The tool needs past quotes, but they sit in three inboxes and a paper file | List where the tool’s input lives and who can export it |
| Chasing the technology instead of the user | The firm picks the tool that went viral, and staff were never asked | Name the person who does the job today and ask them first |
| Inadequate infrastructure | Nobody can say where files are kept, who has logins or how output reaches a customer | Draw the path from input to finished output on one page |
| A problem too hard for AI | The tool is asked to judge a difficult case with no clear rule | Run ten real past cases through it and count the misses |
RAND’s report focuses on machine learning, so treat the table as a translation, not a finding about small firms. Bedrock AI’s reading is that the same five patterns apply when a small team buys a ready-made tool.
How does a misunderstood problem sink a small business project?
A misunderstood problem sinks a project when the tool answers a question nobody was asking. RAND says misunderstandings about a project’s purpose are the most common reason for failure. Trained models are often optimised for the wrong measure.
In a small firm this looks like a tool that writes polished emails when the real delay is a missing price. The output is fine. The job stays slow.
Bedrock AI’s view is that most small-firm AI failures are decided before the purchase, in the sentence that describes the job. A vague sentence cannot be tested, so nobody ever learns the tool did not help.
For a method that finds the right job first, see where to start with AI.
Does a small firm have enough data for AI to work?
A small firm usually has enough data for an off-the-shelf tool, but not always in a place the tool can reach. RAND lists a lack of data as its second root cause, and Gartner puts poor data quality first on its own list.
For a business of 5-20 people, the data problem is rarely volume. It is that past quotes, client notes and price lists live in personal inboxes, on desktops and in people’s heads.
A quick test is to ask the person who does the job to find three past examples in under five minutes. If they cannot, the tool cannot either.
What does a failed AI project look like in numbers?
An AI project has failed when the fee keeps going out and the number it was bought to move has not changed. The example below is illustrative, with invented figures, so you can follow the sums.
A ten-person kitchen fitting firm pays for an AI tool “to improve customer service” at £40 a month. Three months later nobody can say what improved. The firm has paid £120 and learned nothing, because no baseline was measured.
Now rewrite the job as one sentence. “Cut the time spent chasing unanswered quotes from 5 hours a week to 2 within six weeks, owned by the office manager.”
At an assumed £22 an hour, three saved hours is £66 a week. The £40 fee is about £9 a week.
The net gain is roughly £57 a week, or about £250 a month. The method is set out in full in the ROI calculator guide.
The sentence also gives a stop rule. If chasing time is still above 4 hours after four weeks, the firm cancels, having spent about £40.
What is the one-sentence test that catches most failures?
The one-sentence test asks whether the job fits in a single sentence containing a verb, a number and a name. If it does not, the project is not ready to buy.
Use this order of checks before any purchase:
- Job. Write one sentence with a verb and a number, such as “cut quote chasing from 5 hours to 2”.
- Owner. Name the person who does the job today and will keep doing it.
- Baseline. Time the job for one week so the starting number is real.
- Input. Confirm the information the tool needs can be found and exported.
- Path. Write down who checks the output and where it goes next.
- Stop rule. Set the number and date at which you cancel.
Six lines on one page is enough. The AI implementation plan template turns the same checks into a filled-in page with a trial period.
Why do AI tools get dropped even when they work?
AI tools get dropped when they do not fit the way the team already works, even if the technology performs well. RAND names this as a project that does not fit the overall business workflow.
The abandonment pattern gets its own article, why AI tools get abandoned.
The practical fix is to involve the person doing the job before buying. They know which step is slow and which workaround everyone uses. A tool chosen without them tends to solve the owner’s version of the problem.
Name one owner for the tool, and give staff protected time to learn it in the first month. A rollout that relies on spare moments tends to fade.
When is a problem too hard for AI?
A problem is too hard for AI when the right answer depends on judgement that the business itself cannot write down. RAND’s fifth root cause is applying AI to tasks that even advanced models cannot automate.
The ten-case test is the cheapest way to find out. Take ten real past examples and run them through the tool.
The job’s owner then marks each result right or wrong. A tool that misses more than a couple still needs a human on every case, which removes most of the saving.
Costs matter here too. Gartner lists escalating costs among its four reasons for abandonment. A subscription hides most of the bill for a small firm, so count staff time for set-up, training and checking. A stop date limits what a wrong call can cost.
FAQ
What percentage of AI projects fail? RAND cites estimates that more than 80 percent of AI projects fail, twice the rate of other IT projects. MIT’s NANDA report puts enterprise generative AI pilots with no measurable profit impact at 95%. The definitions of failure differ, so treat these as a direction, never a forecast for your firm.
Do AI projects fail more often in small businesses? No study found for this article splits failure rates by firm size. A small firm usually loses less per failure, because it buys a seat or two rather than building a system. It also fixes the biggest cause, a vague job, faster than a large company can.
Is bad data the main reason AI projects fail? No. RAND lists a lack of data as its second cause, behind a misunderstood problem. Gartner names poor data quality as its first reason for abandonment. For a small firm, the practical question is whether the information the tool needs can be found and exported.
How long should I trial an AI tool before calling it a failure? Four weeks is a reasonable trial for a single job, with a stop rule written before it starts. Write down the baseline number first and the number that counts as success. If the number has not moved by the review date, cancel and choose a different job.
Get ahead with AI by getting The Business Brain | Would rather talk it through? Book a free 15-minute call