You bought the tool. Maybe your team tried it for two weeks. Now it's a tab nobody opens.
That's the most common way an AI project fails. It doesn't blow up. It fades.
The research on why is consistent, even though most of it comes from bigger companies than yours. Here's what it says as of November 2026, and the six things the projects that work do differently. We're careful below about which numbers are firm and which are directional. For the money side, see whether AI saves small businesses money.
How often projects miss
A few numbers to set the scene.
IBM's 2025 survey of 2,000 CEOs found only 25% of AI initiatives delivered the expected return, and only 16% scaled across the whole company. Half said rapid investment left them with disconnected, piecemeal technology.
Closer to your size, Goldman Sachs surveyed alumni of its 10,000 Small Businesses program. Only 14% said AI is fully embedded in their core operations, while 73% said they'd benefit from more training and implementation help. That group isn't a random sample, so treat it as one signal. The Federal Reserve's Small Business Credit Survey shows a similar picture: among small firms using AI, about half were still experimenting, 44% had partly integrated it, and 7% had fully integrated it.
On predictions, Gartner has said, according to its June 2025 release as reported by others, that over 40% of "agentic" AI projects will be canceled by the end of 2027 because of cost, unclear value or weak risk controls. That's a forecast, not a measured rate, and we couldn't open the release ourselves. We mention it for direction only.
One more warning. You may have seen "95% of AI projects fail." That comes from MIT NANDA's 2025 preliminary report, which is based on 52 interviews and 153 survey responses and calls itself preliminary. It's a study of company pilots, not a universal failure rate. We don't repeat the headline. What's useful in it is the reason it gives: most tools in the study didn't retain feedback or adapt over time.
Why they fail
1. Nobody trained the team
A BCG survey of more than 10,600 workers found regular use is much higher with at least five hours of training plus coaching. Only about a third said they'd been properly trained. That's a link, not proof of cause. But it fits the NBER support study, where new and lower-skilled agents gained 34% in issues resolved per hour and experienced ones barely moved. People who don't know the job or the tool get the most from guidance, and the least from being handed a login.
2. Nobody owns it
When AI is "everyone's job," nobody decides what it's for, who can use it, or whether it's working. BCG found strong leadership support lifts positive sentiment from 15% to 55%, but only about a quarter of frontline employees get it. Again, a survey link.
3. It started too big
"AI for the whole business" is a plan to buy ten tools and use none. The failures people remember are broad rollouts. Klarna's CEO said in 2025 that automation had gone too far and the company hired humans again, and Taco Bell has said it's moving toward a hybrid with people for busy stores, according to press coverage. Those are big companies, but the pattern holds on a small scale.
4. Nobody measured anything
People are bad at judging their own gains. In a METR randomized trial, experienced developers using AI took 19% longer but believed they were about 20% faster. METR says those results are out of date, and developers are a narrow group. The lesson is about humans, not code. An NBER paper found executives perceive larger productivity gains than the measured data show. If you don't have a number, your feelings will decide.
5. No human handoff, and nobody checked the answers
In the Fed survey above, accuracy was the top problem among small firms using AI, at 46% according to the Fed's survey. When a bot answers customers alone, its mistakes become yours. Air Canada's chatbot gave wrong refund information and, according to coverage of the tribunal's 2024 ruling, the airline was held liable. New York City's business chatbot told owners wrong rules. We cover where a person has to stay in the loop in where AI works and where it doesn't.
6. Nobody kept it running
AI isn't software you install once. A Stanford and Berkeley study found the same model scored 84% on one task in March 2023 and 51% on the same questions in June. That's a 2023 result on one task, and critics say part of the gap was how it was measured, so read it narrowly: hosted models change, and what worked last quarter can quietly stop working. Models also get retired on a schedule. We cover that in your AI tools will change under you.
What the ones that work do differently
The research doesn't prove a recipe. Nobody has run a head-to-head trial of "start small" against "go big." But the factors that show up in the successes mirror the causes above.
- They train people, on their own workflow. Not a generic webinar. The five-hour mark in the BCG data is a useful floor. See how to train your team on AI.
- One named owner. One person who answers for it. Usually the owner or an office manager.
- One small task first. A repeated job where a mistake is cheap. Add the second one after the first is working.
- One number. Hours on the weekly report. Missed calls answered. No-show rate. Write down the "before." Check in 60 days.
- A person in the loop. AI drafts, a person approves, and any bot has a clear way to reach a human. Customers care about this: in a Gartner survey from 2024, as reported by others, the top worry about AI in service was that it would be harder to reach a person.
- Upkeep on a schedule. Retest when the tool changes. Check the bill. Review who has access. Look at the number again.
A quick self-check
Answer yes or no for any AI tool you've bought:
- Does a named person own it?
- Can you say in one sentence what it's for?
- Is there one number you checked this month?
- Has your team had real time with it, not just a login?
- If it gave a wrong answer to a customer, would a person catch it?
- Do you know who checks it when the tool changes?
Three or more "no" answers and you're in good company: IBM found only 25% of initiatives delivered the expected return. The fixes are cheap. They're just not exciting.
Frequently asked questions
Do 95% of AI projects really fail?
No, not in that form. The number comes from MIT NANDA's 2025 preliminary study of company pilots, based on 52 interviews and 153 survey responses, and the study itself says it's preliminary. IBM's CEO survey found 25% of initiatives delivered the expected return. Treat both as warnings about big companies, not a forecast for your shop. See our guide to AI implementation.
What's the most common reason an AI tool gets dropped?
Nobody was trained and nobody owned it, so it stopped being used. The research points to training, a named owner and a narrow first task as the common thread in the projects that work. That's a pattern from surveys, not a controlled trial. A short audit usually finds the one task worth starting with.
Can I do this myself without a consultant?
Yes, for small things. Our post on questions to ask an AI consultant helps you decide. The MIT NANDA report says outside partnerships reached deployment about 67% of the time versus about 33% for internal builds, but that's a self-reported deployment rate, not a success rate, so don't lean on it.
How long before I know if it's working?
Pick one number and check it at 60 days. For something simple like reminders or after-hours message taking, you should see a change in the number or a clear "no" by then. If you can't tell, the measure was too vague. Pick a sharper one.
How Di-Hy fits
The six causes above are the four steps of our work run backward. We start with a free audit to pick the one task. We set it up, train your team on it, and keep it running through Keep It Running. For a plan before you spend anything, see Get a Plan. We're in Ybor City and train in person around Tampa Bay or on video.
Next step: tell us which AI tool you bought and stopped using. A short call will usually show which of the six it was.
Most figures here are surveys and studies of bigger companies, with caveats noted. Press coverage items (Gartner, Air Canada, Klarna, Taco Bell) are marked as such. Read in November 2026.


