The short answer: pick one checkable task
If you run a small business, you are probably also the finance team, the HR team and the person who knows where the spare printer cable lives. After 13 years in finance, my rule for a first AI job is this: pick a task where you can tell in a minute or two whether the output is right.
One task. Repeatable, checkable, boring.
This post doesn’t cover which tool to buy. That is a separate question, and I’ve written about it in how to choose AI tools by task and data. This post starts one step earlier, with the job itself.
Why the back office is a better first test than marketing
Most AI advice for small businesses starts with marketing, because a bad social post is cheap to delete. The trouble is that marketing output is hard to judge. Did the post work, or did it just read well?
Back-office work can be checked. A reconciliation note matches the bank or it doesn’t. A policy draft agrees with your contracts or it doesn’t. Because you can check it, you find out quickly whether the tool helps.
You also do these jobs every week, so you have plenty of practice material. Take a slow job you already do and test whether an AI first draft makes it faster.
Three back-office jobs ranked by risk and checkability
I rank jobs on two questions. How easy is the output to check? How sensitive is the data that goes in? Here they are, lowest risk first.
1. Finance admin drafting. Think explaining a variance in plain English, tidying a month-end checklist or sketching a spreadsheet formula. The numbers stay in your ledger. The AI only touches words, structure and formulas, and you check every figure against the source. The manuals for bank reconciliation and cash-flow forecasting show how this works step by step.
2. General office admin. Meeting notes, templates, a first pass at a supplier email. You check these by reading them, and the risk stays low as long as you strip out names and account details first. The office managers hub collects these tasks.
3. HR templates. Policies and job descriptions are mostly structure and wording, so they are reasonable to draft with AI and then check against your contracts and current guidance (see HR policies and job descriptions). Anything about a named person is different: a grievance, a sickness record, a performance issue. That belongs in your HR system, not a chat window.
So the order is finance wording, then admin wording, then HR templates. Data about real people comes last, if it goes in at all.
What stays in your own systems
The rule on this site is simple. Numbers and decisions come from the ledger, the HR system or the bank. AI drafts words, structure and formulas. A person checks and approves.
There is a legal reason as well as a practical one. The ICO’s AI guidance gives “a detailed overview of how to apply the principles of the UK GDPR to the use of information in AI systems”. If personal data goes into a tool, data protection law comes with it. For the basics, the ICO runs a hub for small and medium organisations, including beginner guidance and a self-paced training programme called Data Protection Essentials.
If you are thinking about putting personal data into a new tool, read the ICO’s page on when you need a data protection impact assessment (DPIA). The general rule is that you must do one where processing is “likely to result in a high risk to the rights and freedoms of individuals”. The ICO notes that this guidance is under review after the Data (Use and Access) Act, so read the live version.
Business and consumer plans can treat your inputs differently. OpenAI is one example. Its business data page says that by default it does not use data from ChatGPT Business, ChatGPT Enterprise or its API platform for training. On personal plans, a setting called “Improve the model for everyone” decides whether your conversations are used for training. Other providers set their own terms, so read the terms for the tool and plan you actually use, on the day you sign up.
A one-week trial you can run on fictional data
By our estimate (not a measured figure), this takes 30 to 60 minutes a day for one working week. Adjust it to suit.
- Monday: pick your one task. Write down what a correct output looks like.
- Tuesday: build a small fictional dataset: invented suppliers, made-up invoices, an employee called Test Person. No real names and no real account numbers.
- Wednesday and Thursday: run the task three or four times with slightly different wording. Check every output against the answer you already know. Note each mistake, because the pattern of errors tells you more than the successes do.
- Friday: decide. Did it save time once you count the checking? If yes, try one real example with the sensitive details removed, and check it just as carefully. If no, drop it and pick a different task. Dropping it is a useful result.
What does AI cost a small business? Read the live pricing page
This post gives no figure. Plans and prices change, and a number in a blog post goes out of date fast. Go to the live pricing page for the tool you are considering. Read what is included, what is limited and what that tier says about your data.
If a free version exists, run your fictional-data trial on it before you pay. Then you know what the paid plan would actually improve.
The ‘30% rule’ for AI: what we could and could not verify
You may see people mention a “30% rule” for AI. We searched for a primary source and found no official or authoritative definition. The pages that mention it use the number in different ways. Some say AI does most of a task and people keep 30%. Others say you should automate 30% of the work first.
So it gets no number here as if it were fact. If someone quotes it to you, ask where it comes from. Your own rule is more useful: check everything the AI produces before it goes anywhere.
When to stop and ask someone
Stop if a task involves a named person’s data and you are not sure it is allowed. Stop if an output would reach a customer, a regulator or HMRC without a person reading it first. Stop if you cannot check the answer, because an answer you cannot check is just a guess in a nice font.
Then ask: your accountant, your solicitor or the ICO’s own guidance. Asking first takes less time than fixing a mistake later.