Why AI projects fail in a small business

Most AI projects in a small business do not fail because the technology was wrong. They fail because the people who were meant to use it were never brought along, and the first clear sign of that, someone quietly refusing to use the new tool, gets treated as a problem to manage instead of information to act on.

If that sounds soft, the research is not. When RAND studied why AI projects fail, the pattern was not technical.

What actually goes wrong

Here is the version most owners will recognise. You buy a tool. You set everyone up with a login. Maybe you run a short session showing people how it works. Three weeks later, two people are using it properly, a couple have opened it once, and everybody else has drifted back to the way they did it before.

So you conclude the tool did not work, and you start looking at what else is out there.

That conclusion is the expensive part. It sends you shopping for a second tool to solve a problem the first tool never had, and you pay for the licence twice while the actual blocker sits untouched.

Is it really the technology?

Usually not. In 2024 RAND published a study of why AI projects fail, based on interviews with 65 experienced data scientists and engineers. The single most common cause was not the models, the tooling or the budget. It was leadership: 84 per cent of the people interviewed named a leadership-level root cause as the primary reason projects failed. Data problems came second.

That study was aimed at large organisations, and RAND opens it by citing outside estimates that more than 80 per cent of AI projects fail, roughly twice the rate of IT projects that do not involve AI. Worth noting that the 80 per cent figure is quoted from elsewhere rather than measured by RAND, so treat it as context rather than gospel. The 84 per cent is theirs, and it is the number that matters here.

Scale it down to a business of five or fifty and the same shape holds. Leadership at your size is not a committee. It is you deciding what the tool is for, whether anyone has time to learn it, and whether you have said out loud what it means for people's jobs.

Why the person refusing to use it is worth listening to

This is the part we think most owners get backwards. When someone will not use the new tool, the instinct is to push harder. Another training session. A gentle mandate. A reminder in the team meeting.

What works better is slower and costs nothing. Ask them what they are worried about, and take the answer at face value.

Nine times out of ten the objection turns out to have very little to do with AI. The job is not shaped the way the tool assumes: the software expects a clean sequence, and the actual work has three exceptions in it that nobody documented. Or the output still needs checking, and nobody has time to check it, because a draft that needs verifying is not a time saving to the person doing the verifying, it is a second job. Or nobody has said whether this replaces part of their role.

That last one deserves its own line. If you have not answered it, out loud, unprompted, then people are answering it for themselves, and they are not answering it optimistically.

The framing we use

Resistance is data, not obstruction. It is unfiltered feedback on your rollout, delivered by the person closest to the work, and it is free. Most businesses throw it away because it arrives in an inconvenient form.

If you want the practical version: the first hour of any AI rollout should be spent finding out what actually stops people, not demonstrating what the tool can do.

What to do instead

You do not need a change management programme. You need four things, in order.

Pick one job, not five tools. Standardise a single AI habit across the team, done properly, before you add anything else. If you are not sure which one, our piece on where to start with AI covers how to choose. The broader case for sequencing over shopping is in the next step, not more tools.

Say the quiet part first. Tell people plainly what the tool is for, and what it is not for. If it is not intended to reduce headcount, say so. If you cannot say so honestly, say what you can.

Ask the non-users, not the enthusiasts. Your champions will tell you it is great. The people avoiding it will tell you why it is not being used, which is the thing you actually need.

Budget for the running cost, not just the licence. Adoption stalls faster when a tool turns out to be expensive to use. We covered that trap in what AI actually costs to run.

None of this requires a bigger budget. It requires asking one question and then not arguing with the answer. That is also, in practice, most of what a readiness engagement does. If you want to see how we structure that work, how it works walks through it.

Ready to find out where you stand?

If you want an honest read on where your team actually is with AI before you spend anything else, our free AI Readiness Check takes about five minutes. No cost, no pitch.