Most AI advice for small businesses is either breathless or paranoid. The useful position is in between: start with one narrow task, keep a person responsible for the output, and decide in weeks — not quarters — whether it earned its place.
What AI is genuinely good at right now
- Drafting a first version of something you will edit anyway — an email, a summary, a job description.
- Reformatting and extracting: turning a messy document into structured rows.
- Summarising a long thread or meeting transcript into decisions and action items.
- Answering questions against your own documents, when set up properly.
What it is still bad at
- Facts it was not given. It will produce confident, plausible, wrong answers.
- Arithmetic and totals you cannot check.
- Anything requiring accountability — it cannot sign off, and it should not.
- Judgement calls about your customers, your staff, or your money.
Pick your first task with this test
A good first task is repetitive, low-stakes, text-based, and has an obvious right answer you can check at a glance:
| Good first task | Bad first task |
|---|---|
| Summarising a weekly report | Calculating payroll |
| Drafting a reply you will edit | Sending replies automatically |
| Tagging incoming requests | Deciding who gets credit |
| Extracting fields from documents | Making hiring decisions |
The rule that prevents most problems
A person remains responsible for anything that leaves the building. AI can draft; a human approves.
This is not a legal technicality. It is what stops a plausible-sounding error from reaching a customer, and it keeps someone accountable for the output.
Decide where your data may go, before you paste
This is the question that matters most and gets asked last:
- Which tools are approved, and which are banned? Write it down.
- Free consumer chatbots may use your input for training. Assume anything pasted there is public. Business tiers usually contractually exclude training — verify, do not assume.
- Never paste customer personal data, ID numbers, payment details, contracts, or credentials into a tool you have not approved.
- Keep a simple register: tool, owner, what data is allowed, renewal date.
Measure something honest
Do not measure ‘AI adoption’. Measure whether the task got better:
- Time taken, before and after — timed on three real examples, not estimated.
- Error rate — did quality hold up?
- Whether the person doing the task wants to keep using it.
If it did not help, stop. Turning it off is a valid result and costs nothing.
A two-hour first step
- Pick one repetitive text task and one volunteer.
- Write down today’s time and error rate on three real examples.
- Run the same three with AI assistance plus human review.
- Compare honestly. Keep or drop. Write one paragraph about what you learned.
That paragraph, repeated monthly, is worth more than any AI strategy document — because it is based on what actually happened in your business.
Want help choosing the first task and setting the boundaries? Talk to a specialist.