What AI can and cannot do for a small business

The single most useful piece of advice available to a small business considering AI is that the first step costs a few pounds a month and requires no supplier at all. Before commissioning anything, give the people doing repetitive work access to a competent general assistant and a fortnight to find out where it helps them. What comes back from that is a genuine list of use cases grounded in your actual work, which is a far better brief than anything produced in a meeting.

A great many AI projects are commissioned to solve problems the business has not yet confirmed it has. Doing it in this order costs almost nothing, takes two weeks, and frequently establishes that a subscription and better habits deliver most of the available value. Where it does not, you now have a specific brief instead of an aspiration.

These are the categories where small UK businesses reliably get value, in rough order of how quickly it shows up. Notice that none of them involves replacing a person.

These fail predictably, and they fail after money has been spent rather than during the pitch. Each is a category where the technology is genuinely capable and the business context defeats it.

• Anything requiring judgement about a specific customer. It will produce a confident answer and be wrong occasionally, and in a small business one badly handled account matters more than the time saved across fifty.
• Fixing a process that is already broken. Automation multiplies whatever exists — a poor follow-up process becomes poor follow-up delivered faster and to more people.
• Replacing content that needed expertise. A first draft is genuinely useful; a finished article generated end to end reads exactly like what it is, and increasingly so as more of the web fills with the same output.
• Anything depending on data you do not have. A model cannot know your stock levels, your margins or which customers are difficult unless something tells it, and building that pipeline is usually the actual project.
• Regulated advice. Financial, legal, medical and similar output carries obligations that do not disappear because a machine generated it, and the liability remains entirely yours.

Named specifically, because all three are commonly sold and all three have a cheaper version that works better.

The first is a custom chatbot trained on your website. The cheap version is a good FAQ page and a well-configured off-the-shelf widget, and it captures most of the value for a fraction of the cost. Custom builds become justified when the questions are genuinely complex, the content changes constantly, or handover to a human has to be reliable — and not before.

The second is fine-tuning a model. It is the most requested and least often necessary piece of AI work. Better prompting and supplying the right documents at the point of use solves the overwhelming majority of cases at a tiny fraction of the cost, and fine-tuning becomes worthwhile mainly when you need a specific format or tone reproduced at volume.

The third is an AI feature added to a product because competitors have one. If nobody can describe the job it does for a user, it will be built, announced, and quietly unused — while adding a running cost and a support burden that persist long after the announcement.

AI projects have a recurring cost that proposals frequently understate, and it behaves unlike software licensing because it scales with use rather than with seats.

• Usage costs, which rise with volume. A tool that costs very little in testing can cost meaningfully more once the whole team uses it daily.
• The maintenance nobody quotes. Models are deprecated and replaced, prompts that worked stop working, and something has to be re-tested when a provider changes anything.
• Human review, which is not optional for anything customer-facing. The saving is the drafting, not the checking, and a workflow with no review step is where the reputational damage comes from.
• The integration, which is usually the real project. Connecting to your CRM, your documents or your booking system is ordinary software work and it is priced as such.

If these five have clear answers, the project is probably real. If two or more do not, the honest position is that it is an idea rather than a requirement.

1. 1.What specific task, done by a named person, does this replace or shorten? Vague answers here predict projects that get built and not used.
2. 2.How many hours a week does that task currently take? If nobody knows, measure it for a fortnight before spending anything.
3. 3.What happens when it gets something wrong? Every AI system does eventually, and the answer to this question decides whether a human review step is mandatory.
4. 4.Have you tried doing it with a subscription first? If not, do that before commissioning anything custom.
5. 5.Who maintains it in a year? An AI workflow with no owner degrades quietly, and nobody notices until the output has been wrong for a while.