Why the first AI project in a small business usually fails
Short answer: first attempts rarely fail because the technology could not do it. They fail because the project was too big, aimed at the wrong job, measured against nothing, and owned by nobody.
1. It was a transformation, not a change
The commonest version: an owner decides to "get AI into the business", and the project becomes everything at once. Three months later there is a lot of trying and nothing running.
The fix is almost insultingly simple. One job. One tool. Two hours of setup. Working by Friday. Then the next one. A small business cannot absorb more than about two new habits at a time, and that is a constraint about people, not software.
2. It was pointed at the wrong job
People pick the job that sounds impressive rather than the one that is costing them. Or they pick a job that needs a guaranteed-correct answer with nobody checking — which is the category AI is worst at. Sorting your jobs properly up front avoids most of this; what AI is and is not good at is the sorting rule.
3. Nobody wrote down what success was
"It seems quite good" is not a result. Without a before number — hours a week, days to respond, errors per month — you cannot tell whether it worked, which means you cannot defend keeping it or justify stopping. Thirty seconds of writing down the current figure at the start saves the whole argument later.
4. The saving was never costed honestly
Two hours saved, one hour spent checking the output, and a subscription nobody counted. Net, that might be a small win or a small loss, and it matters which. Count the checking time. Subtract the tool cost. Use a real hourly figure rather than a flattering one — how to work out what an hour of your time costs covers how to get that number right.
5. It had no owner
If a new process belongs to everyone it belongs to nobody, and it quietly stops within a fortnight. Someone has to be the person who notices it has stopped. In a five-person business that is usually the owner, which is fine as long as it is explicit.
6. The prompt lived in someone’s head
The most reliably fatal one. A member of staff works out a good way to do something, does not write it down, and the knowledge leaves with them or with their memory of last Tuesday. Written-down instructions, stored where the team can find them, are the difference between a trick and a process.
What a project that sticks looks like
Narrow enough to describe in a sentence. One named job, with the current time cost written down. One tool, preferably one you already pay for. Instructions written out and saved. One named owner. Set up in under two hours. Reviewed after a month against the number you wrote at the start.
That is unexciting, which is exactly why it works. And it is why our AI Tools Report ends in a four-day plan where every task takes under two hours rather than a roadmap: ten minutes of questions, then a costed nine-slide report within minutes with six quick wins as problem and fix, the tools and their setup times, and a monthly net return built from your own numbers and rounded down. £59.99 inc VAT.
The report will not stop you overreaching if you are determined to. It will at least tell you which small thing to do first, and what it should be worth to you.
Questions, or want a tool we don't have yet? Email hello@grafto.co.uk — a real person replies.