AI projects fail for five main reasons, a 2024 RAND study found. The team solves the wrong problem. The data the job needs is missing. Someone picks the tool before the task. Nobody plans how to keep it running. Or the job needs more than AI can do. You can check your plan for all five before you spend.
I delivered software projects for nine years from the vendor's side, at Upstack Studio. That experience shapes how I read a plan. I want to know what could stop it working once real people use it.
For its 2024 study, RAND talked to 65 people who build AI. It looked at work on building AI models. Using a ready-made model was outside its scope.
Test with real work too. Even a sound plan can't rule out every failure.
Five reasons AI projects fail
Each of RAND's five reasons below comes with a check for your plan.
Five reasons, five checks
- 1Wrong problemCheck: one number to change
- 2Missing dataCheck: a real sample
- 3Tool before taskCheck: the simplest fix
- 4No running planCheck: who runs it each day
- 5Beyond what AI can doCheck: which errors you can accept
Settle all five before you spend.
1. The team is solving the wrong problem
A tool can work as built and still miss what your business needs. You may get a new screen without fixing the task that wastes your team's time.
Does the plan name the problem and show how you'll judge the result? A list of features won't answer that.
Write down one number you want to change, and what it is now. You could track hours spent on quotes or days to follow up a lead. Use a number you can check.
If you can't choose one yet, work out where the problem costs you most. This guide to choosing where AI goes first walks through that choice.
2. The data the job needs is missing
Your records can work well for you and still lack what a new tool needs.
Check a sample of the records the system would use. Are the facts there? Are they right? Can the team explain what each field means?
A made-up example: say you want a tool to draft quotes. Your weekly sales report shows sales by customer. It doesn't show the items you quoted, the price, or why a deal was lost. The tool needs those facts. Someone on your team decides whether to start logging them or pause the test. If you start logging, price that work before the full build.
Made-up example: a quote tool
Check a sample before the build. Then decide: start logging the facts, or pause the test.
Ask for a test with a sample you're allowed to share. Remove private details first. Check the result against the source, including any gaps the tool should flag.
3. The team picked a tool before a task
An AI pitch may name the tool first. Ask what job it would do and why you need it.
Write down the simplest fix you could try. You could change the process, buy a tool, or make no change. Compare the cost and what each option would solve.
If you're leaning towards a custom build, work through the build vs buy questions. Use AI when its result is worth its full cost.
4. There's no plan to keep it running
The build needs a place to run and a way to reach the right data. Which tools does it need to connect to? What happens if those links stop working?
Name the person on your side who will check the work each day. Agree how faults get raised and who fixes them. Include those costs in the price you compare.
A working demo helps you judge the idea. You still need a plan for daily use.
5. The job needs more than AI can do
A tool may make too many errors for the task you have in mind. A wrong answer could cause harm before anyone spots it.
Decide which errors you can live with and which you can't. Then ask how each kind will be caught.
Test hard cases as well as the easy ones. Keep a person checking work where a mistake could harm a customer. If there's no safe way to check it, keep AI out of that task.
Sort the errorsWhich can you live with? Which can't you?
TestHard cases as well as easy ones
Could a mistake harm a customer?Keep a person checking
No safe way to check?Keep AI out of that task
Use this checklist on your next plan
Write an answer for each row. A row you can't answer shows what to settle before you go ahead.
| Check | Ask | Fix the gap if |
|---|---|---|
| Goal | Which number should change? What is it now? | The team can't name it |
| Data | Have we checked a real sample? | The facts are missing or untested |
| Choice | What's the simplest fix that could work? | A cheaper option would solve enough |
| Daily use | Who checks it? Who fixes faults? | The plan ends at launch |
| Errors | What could go wrong? Who catches it? | A mistake could reach a customer unchecked |
The checks work for projects built in-house too.
When should you skip an AI project?
Wait if the project would cost more than the problem it solves. Check these points before you commit:
- Would fixing the process solve most of the problem?
- Will the running cost outweigh the value you expect?
- Does someone have time to own the work after launch?
- Can you agree how to judge the result?
The AI Profit Opportunity Audit ranks the options for where to spend. It includes what to leave alone. Your team, or people you choose, do the building.
How often do AI projects fail?
Published failure rates vary. Each source counts different work and tests success in its own way. None of them gives the odds for your project.
- RAND, August 2024: cites an older estimate of more than 80% of AI projects failing. RAND's own work used interviews to study the causes. It didn't measure that rate itself.
- Project NANDA, July 2025: its early report found a gap between trials and wider use. The authors call this a rough guide based on what people told them. The samples and tests of success vary.
- Gartner, July 2024: forecast 30% or more of generative AI projects ending after trials. That was a forecast for the end of 2025. It doesn't tell us what happened in 2026.
Questions owners ask next
Does the '95% of AI pilots fail' headline mean AI gives no value?
No. The NANDA report separates general chat tools from tools built for a set task. Its success test includes a lasting gain in work output or profit. Time saved can count, even without a profit gain. The report's figures aren't the odds for every AI project.
Is the supplier to blame if it fails?
Sometimes. You still need to agree on the goal, supply usable data and check the work. Put those duties in the plan so each side knows what it must do. Before you sign, use these questions to ask an AI or software vendor.
Should we start with a small test?
Yes, if you know what you'll check and what follows a good result. Set the scope and write down what would make you stop, change or go further.
Do we need to commit for a year?
RAND advises a year of focus on one problem for the model-building work it studied. That isn't a minimum length for every trial with a ready-made AI tool. Choose a test you can judge. Plan for the work needed if it succeeds.
Do small companies fail at the same rate?
These sources don't give a firm rate for a company like yours. Some mid-sized firms in the NANDA report moved faster than large ones. Those were the top performers among mid-sized firms, not typical ones. Their pace doesn't set your deadline. Check your data, team and task before you choose a date.
What if our AI project has already failed?
Start with the gaps in the checklist. Find which ones match what happened. Fix them before you spend again, or decide the problem isn't worth more work.
Before you spend
Take the plan you're weighing now and fill in the five rows. Bring the gaps to the people doing the work. Agree what needs fixing before the next step.
