Which jobs in a small business AI is genuinely good at — and which it is not
Short answer: today’s AI is excellent at producing a good draft from messy input, and poor at anything that must be exactly right without a human looking. Sort your jobs by that one line and you will avoid most of the expensive mistakes.
The sorting rule
Ask two questions about any task. First: is the input mostly language — emails, documents, descriptions, notes? Second: is a 90-per-cent-right answer that a person then checks genuinely useful, or does it have to be right first time with nobody watching?
Language in, draft out, human checks: good fit. Numbers in, single correct answer out, nobody checking: bad fit. Almost every disappointment we hear about comes from putting a task in the second group into a tool built for the first.
Things it is genuinely good at
First drafts of anything written. Quote covering notes, follow-up emails, product descriptions, job adverts, policy text, tender answers. You edit rather than start from nothing, which is usually where the time goes.
Summarising a pile. Twelve supplier emails into one paragraph of what changed. A long thread into three bullets and an action. A call into a note someone will actually read.
Pulling structure out of mess. Reading a PDF delivery note or an awkward supplier price list and turning it into clean rows. This is often the highest-value, least glamorous use in a trading business.
Classifying and routing. Deciding whether an incoming email is an enquiry, a complaint or a supplier notice, and tagging it. Cheap, invisible, saves triage time daily.
Rewriting the same thing for different audiences. One set of product facts into a listing, a trade email and a social post.
Answering repetitive customer questions from your own documents, with a person reviewing anything unusual.
Things it is not good at, and should not be given
Being the system of record. Stock figures, prices, balances. Those belong in a system that is accountable for them. If your stock number and your AI disagree, neither is the problem — the architecture is. This is the same discipline as working out which stock number to trust.
Final arithmetic that goes to a customer or HMRC. Use it to draft, explain or check a method. Do not use it as the calculator of record, and do not take tax rates, thresholds or deadlines from it — those come from gov.uk or your accountant.
Decisions with consequences and no review. Approving credit, hiring, firing, pricing a tender, deciding what to order. It can prepare the brief for all of those; it should not make them.
Anything where it being confidently wrong is costly. If an error would reach a customer uncorrected, there must be a person in the loop. That is a process rule, not a tool setting.
Relationships. The awkward phone call, the apology, the negotiation. Small businesses win on exactly those, and automating them away is a strange thing to buy.
The grey middle
Some jobs sit in between and are worth piloting narrowly: drafting replies that a human sends, suggesting a reorder quantity a human approves, flagging unusual invoices for review. The pattern is always the same — AI proposes, person disposes. Keep that shape and the risk stays small while the time saving is real. It is the same thinking as deciding when not to chase an estimate: automation with judgement attached beats automation alone.
Sorting your own list
If you would rather have someone else sort your jobs into those piles, that is precisely what the AI Tools Report produces. Ten minutes of questions about your week, then a nine-slide costed report within minutes, including an impact–effort matrix and six quick wins written as problem and fix, each with a tool, a cost and a setup time. £59.99 inc VAT. It is deliberately boring about what AI can do, because being honest about the "not good at" column is what makes the other column worth acting on.
Questions, or want a tool we don't have yet? Email hello@grafto.co.uk — a real person replies.