Practical AI
What to use AI for in a small business — and what not to
The question is never whether AI can do a job. It usually can, to some standard. The question is whether the standard it reaches is good enough for that particular job, and that depends on what it costs you when it is wrong.
Sort by the cost of being wrong
Ignore how impressive a task sounds. Ask one thing: if this output is wrong and nobody catches it, what happens?
| Cost of an unnoticed error | Example | Use AI? |
|---|---|---|
| You notice and redo it | A first draft, a summary, notes tidied up | Yes, freely |
| Mild embarrassment | A social post, an internal update | Yes, with a read-through |
| A customer is misinformed | A quote, a policy answer, a delivery promise | Only with a human check before sending |
| Money moves, or a rule is broken | Invoicing, contracts, anything regulated | No — or a person signs it off every time |
That table settles most arguments. It is not about the glamour of the task, it is about who absorbs the mistake.
The jobs where it genuinely pays
Anything that starts with a blank page
First drafts are where AI is strongest, because the alternative is staring at nothing. A mediocre draft you improve beats a perfect draft you never started, and editing is a far easier mode than writing.
Turning one thing into another
Notes into a summary. A long document into bullet points. A phone call into an action list. Data from one shape into another. The information already exists and is not being invented, which is exactly when the technology is most reliable.
The same question, answered fifty times
If you answer the same five customer questions every week, drafting those replies is nearly free once you have supplied your actual policies. Note the condition — supplied, not guessed.
Being the person who asks the obvious question
"Here is my plan. What have I not thought about?" This is underrated and costs nothing. It will not be brilliant, but it reliably catches the thing you were too close to see.
The jobs to keep
Anything where being specifically you is the point
The message to a customer whose order went badly wrong. The condolence note. The reply to a complaint that is really about something else. These land because a person wrote them, and a competent-but-generic version is worse than a clumsy sincere one.
Final numbers
Quotes, invoices, tax, anything that ends up in your accounts. Use AI to structure the working if you like; do not let it produce the figure that leaves the building.
Decisions you would have to defend
Hiring, firing, pricing changes, whether to take on a client. You can use it to think out loud. The decision is yours, and "the AI suggested it" is not a position you can hold in front of anyone.
The trap in the middle
The genuinely dangerous zone is not obviously-bad uses. It is the ones that work well enough for months and then fail expensively once.
Customer-facing replies are the classic case. Ninety-five per cent are fine. The five per cent where it states a policy you do not have, or a date you cannot meet, is a commitment you made without knowing it.
If you go anywhere near this, two rules: supply the actual policies rather than hoping, and instruct it explicitly to leave a marked gap rather than guess. That single instruction converts a silent wrong answer into a visible blank.
Where to start on Monday
- Pick the job you most resent doing that nobody sees — notes, summaries, the weekly update.
- Write down how you do it, properly, once.
- Use that document as the prompt.
- Check the output against what you would have produced.
- Only when it is reliably good should it go anywhere near a customer.
Small, invisible, repetitive. That is where the hours are, and it is where being wrong costs nothing while you learn.