The most common pattern in small-business AI adoption is a pilot that impresses everyone, changes nothing, and is quietly forgotten. Avoiding that has less to do with technology choice than with sequence — and with picking a first project small enough to finish.
Week one: look, do not buy
1
Track where the hours go. A week of honest notes, in half-hour blocks.
2
Mark the repetitive ones. Same shape, different details.
3
Total them by month. Daily small tasks usually beat occasional large ones, and this is where intuition misleads most.
4
Use a general AI assistant on real work — drafting, summarising, sorting. Free, and it calibrates your sense of what these tools are actually good at.
5
Buy nothing. Anything bought this week is bought before you know what you need.
Month one: one workflow, end to end
Pick the single task with the most hours attached and automate only that. Not three. One.
- Measure before. Hours, response time, error rate. Without a baseline you cannot tell whether it worked, and you will end up arguing from impressions.
- Build or buy the narrowest thing that does the job.
- Use it for a month in real conditions, not in a test.
- Measure after, and be willing to conclude it did not help.
The discipline of one workflow is the whole method. A broad rollout gives you five half-working things nobody trusts, and no way to tell which part failed. One workflow that visibly works builds the confidence — and the internal evidence — for the next.
Quarter one: expand from evidence
- Add the next workflow from your original list, now with a working precedent.
- Connect systems where the same data is being entered twice.
- Decide your data boundary. What may go to external services and what stays in-house. Write it down — see AI and your business data.
- Train the people using it. Adoption fails far more often than technology does.
What to ignore
- Strategy documents about AI transformation. Automate one process. That is the strategy at this scale.
- Tools without a named problem. If you cannot say which hours it saves, it saves none.
- Chatbots as a first project. Visible, risky, usually low saving.
- Rebuilding what works. Modernising a functioning process is how you buy a problem.
- Comparison articles. Including this one, past the point where you know your top task. Reading is not a substitute for measuring your own week.
What good looks like at three months
One or two workflows running reliably, ten to thirty hours a month back, a clear sense of what these tools are and are not good at in your specific business, and a list of what to do next based on evidence rather than enthusiasm.
That is an unglamorous outcome and a genuinely good one. It compounds; a transformation programme usually does not.
Common questions
What is the first thing a business should do with AI?
Track where the hours actually go for a week, before buying anything. Almost every business is surprised by the result, and the surprise is what makes the first project the right one instead of the most fashionable one.
How much should a first project cost?
Small enough that failing is acceptable. The point of a first project is learning whether this helps in your specific business — a commitment large enough to need justifying is a commitment that will get defended past the point of usefulness.
How long before I see a return?
For a well-chosen single workflow, weeks rather than months. If a first project has not shown a measurable difference after a month of real use, something is wrong with the choice of task or with adoption — and both are worth diagnosing before adding anything else.
Should I get help or do it myself?
Do the looking yourself — nobody else can tell you where your hours go, and an external assessment that skips this step is guessing. Get help for building anything that connects existing systems, or where a silent failure would matter. A free audit is a reasonable way to get a second opinion on the sequence before committing.