AI is a component, not a project
“We should do something with AI” is not a brief, and it is the reason a lot of small-business AI spending produces nothing. AI is a component — a very good one for a narrow set of jobs — and it only pays when it is bolted into a workflow that already exists and already hurts.
The businesses that get value out of it did not start with the technology. They started with a task somebody does forty times a week and resents, and asked which part of that task a machine could take.
The four jobs a language model is genuinely good at
Nearly every useful small-business AI workflow is one of these four, or a chain of them.
- Drafting. A first reply, a quote covering note, a job description, a follow-up. Getting from blank page to 80% is where most of the time goes.
- Classifying. Reading an incoming enquiry and deciding which pile it belongs in, how urgent it is, and who should see it.
- Extracting. Pulling structured fields out of unstructured mess — dates, totals, addresses and reference numbers out of a PDF invoice or a rambling email.
- Summarising. Turning a long thread, a call transcript or twelve months of tickets into something a person can act on in a minute.
Notice what these have in common. In all four, a human sees the output before anything irreversible happens, and a mediocre answer costs a few seconds of correction rather than a customer.
The jobs it is bad at, and will stay bad at
Being straight about this saves money. A language model should not be your source of truth for arithmetic, stock levels, prices or dates — look those up in the system that owns them and hand them to the model, rather than asking the model to remember. It should not make policy decisions, sign anything, or be the only thing standing between a customer and a wrong commitment.
And it should not be asked to do a job that a simple rule already does correctly. If “send the reminder three days before the appointment” is the requirement, that is a scheduled job, not an AI feature. There is a whole category of that in our guide to simple automation for local businesses.
Choose the first workflow with a stopwatch
Rank candidate tasks on three numbers and the winner is usually obvious.
- Volume. How many times a week does this happen? Under ten, it is rarely worth wiring up.
- Minutes. How long does one instance take a person today, honestly measured rather than estimated?
- Cost of being wrong. If the output is wrong and nobody notices, what breaks? High answers here move a task to the back of the queue, not the front.
Volume times minutes gives you the prize. The third number tells you how much review to build around it. A task worth six hours a week with a low error cost is a good first project. A task worth six hours a week that could misquote a customer is a good second project, after you have learned how your team actually uses the first one.
Design the checkpoint before you design the prompt
The part that determines whether an AI workflow survives contact with a busy week is not the prompt. It is where the human sits.
Good patterns: the model drafts into the reply box and a person presses send. The model tags an enquiry and a person can re-tag it in one click. The model proposes, the system logs both the proposal and the human decision, so you can see six weeks later how often it was right.
Bad patterns: the model sends. The model updates a record with no history. The model produces an answer nobody can trace back to a source document.
Prove it with a before-and-after, not a demo
Before you switch anything on, write down the current numbers: enquiries per week, average time to first reply, hours spent on the task. Those three lines are the whole business case, and they take an afternoon to collect.
Four weeks later the same three numbers tell you whether to expand, adjust or stop. Teams that skip this step end up arguing about whether the AI is helping, which is an argument nobody wins. Teams that take the baseline usually find the honest answer is somewhere between the vendor’s claim and the sceptic’s — and useful either way.
If you are not sure whether the business is ready to run this properly, our AI readiness assessment scores you across the areas that usually decide it.
Common questions
What is the best first AI project for a small business?
Whichever repetitive task has the highest volume multiplied by minutes spent, where a wrong answer is annoying rather than dangerous. Enquiry triage, first-draft replies and document search usually score highest. Anything that touches money, contracts or safety should not be first.
Do I need to train my own AI model?
Almost certainly not. Training is expensive and rarely helps small businesses. Modern models do far better when you give them your actual documents at the moment of the question — retrieval, not training. Budget for wiring, not for GPUs.
How much does it cost to run an AI workflow?
The model itself is usually the smallest line. Per-task inference for typical business text is fractions of a penny. The real cost is the integration and the review process around it. Our AI Integration Sprint is £1,995 for a working, connected workflow.
What about hallucinations?
Assume the model will occasionally be confidently wrong, and design so that this is survivable. Ground answers in your own documents, cite the source, and keep a person approving anything that leaves the building. Workflows built that way stay useful even when a specific answer is bad.
Is my business data safe?
It depends entirely on the provider and the plan. Business and enterprise API tiers from the major providers do not train on your data by default, but the free consumer tiers often do. Check the terms for the specific tier you are on before pasting anything client-confidential.
Related guides
- AI readiness assessment — a free scored checklist before you spend anything
- Simple automation for local businesses — the same problems solved without AI
- AI strategy for SMEs — the longer strategic version of this guide
- Best AI automation ideas for 2026 — worked examples by business type