AI for small businesses: which tasks are worth automating?
Where AI actually pays for itself in a small business — and where it quietly burns money.
Most of the conversation about AI for business is written for large companies. A business of 10-200 people is in a different position: the budget is small, there is little room to experiment, and a badly chosen first project usually kills the second one too.
At this scale the right question is not "how do we use AI". It is: which of the tasks we already have is ready to be automated?
Four properties of a task worth automating
Ask four questions before a task goes on the list. If you cannot answer "yes" to all four, it should not be your first project.
- Does it repeat at least a few times a week? A monthly task will never repay the cost of maintaining the automation.
- Can the rules be written down? If the person doing the work cannot explain how they decide, a system cannot either.
- Is a mistake reversible? A wrongly tagged request can be fixed; a wrongly issued invoice cannot.
- Who checks the output? An automation with no owner quietly starts producing wrong results within months, and nobody notices.
What actually pays off at small-business scale
In practice the fastest payback comes from the unglamorous work:
- Classifying and routing incoming requests. Getting emails, forms and messages to the right person. Measurable gain: time to first response.
- Extracting data from documents. Invoices, delivery notes, order forms. Anywhere data is keyed in by hand, there is a clear gain.
- First answers to repeat questions. Not all of them — the repeating 40 per cent.
- First drafts of product and content copy. Drafts, not final text. An edited draft is faster than a blank page.
- Compiling reports. The same weekly summary pulled from several tables.
What burns money
The list of things not to do is just as clear:
- Pricing decisions. Margin, stock and competition are weighed together; this is not left to a model alone.
- Sending the final customer message unreviewed. First draft yes, sending no.
- One-off work. Automating a task performed once a year takes longer than doing it.
- Processes whose data is scattered. AI does not fix the mess, it is built on top of it. First find out where the data sits.
How to split the budget
At this scale a workable split is roughly: half the budget on the work itself, a quarter on getting the data in order, a quarter on corrections during the first three months. Projects that ignore the last item end up ownerless shortly after launch.
And a duration: the first project should not run longer than 6-8 weeks. If it does, the scope was chosen badly — the task is not narrow enough.
Where to start
Write down the repeating tasks you already have, apply the four questions to each, and from what survives pick the most boring one. The job of a first project is not to impress; it is to make the team say "this works".
We cover the sequencing and what the first week looks like in more detail here: AI for business: where to start?. If you would rather map these tasks together, the work we do on the AI side begins with exactly this distinction.