Start with one job, not a tool. Pick a task that comes round every week, takes a few hours of someone's time, and produces something a person reads before it goes anywhere: meeting notes, first drafts of proposals, summaries of long email threads, the commentary on the monthly numbers. Write down what information that job needs, and one rule for what staff may and may not paste into an AI tool. Then run it for a month with five or so people, on the AI you probably already pay for (Copilot Chat comes with Microsoft 365 business plans at no extra cost). Measure the time it saves, keep what works, and pick the next job.
Why start with a job instead of a tool?
Because the tools are already in the building. The ONS reported on 8 October 2026 that 38% of UK businesses with ten or more employees were using at least one AI technology in September 2026, up 26 percentage points since September 2023 (ONS, 8 October 2026). The government's UK Business Data Survey, published 18 June 2026, found 51% of small businesses (10 to 49 staff) and 58% of medium businesses (50 to 249 staff) that handle digital data used AI (DSIT, UK Business Data Survey 2026, as of October 2026).
What's usually missing is direction. In the same DSIT survey, only 17% of small and 22% of medium AI-using businesses had a formal, written AI policy. The ONS found in July 2026 that only 11% of businesses with ten or more staff said more than half their workforce had received any AI training (ONS, 20 July 2026). And Microsoft UK's research with 2,003 employees (13 October 2025) found 71% had used unapproved consumer AI tools at work (Microsoft UK Stories, as of October 2026).
So people are already using AI, on their own initiative and their own accounts. Choosing a job gives that use a direction: one place where the business agrees AI should help, with an approved tool and a clear rule. Starting from the tool ("we've bought it, go and use it") tends to produce scattered use and nothing you can measure.
And it's worth doing. In the FSB's June 2026 research, 55% of small firms said they now use AI, up from 20% in 2023, and 59% reported productivity gains (FSB, June 2026, as of October 2026).
What makes a good first AI job?
The best first job is boring, frequent and already checked by a person:
- It happens at least weekly. A twice-a-year task won't build anyone's habit.
- It's mostly reading and writing. AI tools are strongest at summarising, drafting, rewording and finding things, which is where a lot of office time goes.
- Someone already checks the result, so a mistake gets caught before it leaves the building.
- The information lives somewhere the tool can reach: your email, Teams, SharePoint, a folder of past proposals.
- Someone actually wants it. A director who'll use it daily beats a team that's been told to.
Leave a few things for later: decisions about people (hiring, pay, performance), figures that go to the bank, HMRC or the board without a check against the source, and anything inside a regulated process. AI can help there eventually. The first job should just be one where a mistake is cheap and easy to spot.
Which job should we start with?
Four jobs that tend to pass the test in a business of 20 to 200 people. Pick the one with a keen owner and a before-and-after you can measure.
| First job | Who it helps | What the tool needs to read | What stays out | Tool you may already have | How you'll know it worked |
|---|---|---|---|---|---|
| Meeting notes and actions | Directors, account managers, project leads | Teams meeting transcripts | Board, HR and disciplinary meetings at first | Copilot in Teams (paid Microsoft 365 Copilot seat) | Notes and actions out the same day, without anyone writing minutes |
| First drafts of proposals and tenders | Sales, bid writers, the MD | Past proposals, standard terms, the client's brief | Other clients' confidential details | Copilot in Word, ChatGPT Business, Claude Team | Time from brief to first draft, measured before and after |
| Summaries of long internal and supplier email threads | Managers who inherit a thread | The thread, in the mailbox | Anything a contract says stays private | Copilot Chat in Outlook, working on the open thread (included) | People reply without reading 40 messages first |
| First draft of the month-end commentary | Finance director, management accountant | This month's signed-off numbers, last month's commentary | Unreleased results in any unapproved tool | Copilot in Word and Excel, or a business AI plan | Every figure in the draft ticks back to the pack before it goes out |
Source: a structured decision table, not survey data. Tool availability is from each vendor's pages as of October 2026; check the plan you're on.
None of these asks the AI to decide anything. They take the first draft off someone's plate and leave the judgement with them, which is exactly where AI is most useful today.
What information does the first job need?
Name it on paper: which mailbox, which SharePoint site, which folder. Two reasons.
First, the tool has to reach it, and that's also where the risk sits. Microsoft says Copilot "only surfaces organizational data to which individual users have at least view permissions" (Data, Privacy, and Security for Microsoft Copilot, as of October 2026). In plenty of businesses a salary spreadsheet or board paper was shared with everyone years ago and nobody noticed, because nobody went looking. An AI tool goes looking. Before paid seats go out, have your IT provider check sharing on the finance, HR and board sites.
Second, numbers. Ask any AI tool about last quarter's margin and it will answer confidently from whichever spreadsheet it finds. Sage's research, published 23 September 2026, found 42% of UK SMEs say their financial information is too spread out to make better use of AI, and only one in ten can see their full financial picture in one place (Sage UK, as of October 2026). So the first job should work on words, or on figures a person has already signed off. Questions about your own numbers come later, once there's one agreed version of them.
What's the rule for what staff may paste in?
One sentence and a short list: "Company work goes into our approved business AI tool, signed in with your work account, never a personal or free account. These never go into any AI tool: client or customer personal data, payroll and HR records, passwords, unreleased financial results, and anything covered by an NDA."
The account matters because of the terms. Microsoft says Copilot prompts and responses aren't used to train its foundation models, OpenAI says business workspace data is excluded from training by default, and Anthropic lists training as "none by default" on Claude Team. Free and personal versions are a different contract.
Then one more line: the person who sends something owns it, however it was drafted. The biggest single AI worry in the FSB's research, at 54%, was inaccurate responses, and a named person checking the output answers most of that. For the full version, my one-page AI policy template has ten sections you can adopt. For the first job, the sentence above is enough.
Which AI tool should the first job use?
The one you already pay for. On Microsoft 365, Copilot Chat is included with business plans at no extra cost (Microsoft UK Copilot pricing, as of October 2026). On Google Workspace, Gemini is built in. On neither, ChatGPT Business and Claude Team are the usual choices.
For the people who'll use it most, a paid seat puts Copilot inside Word, Excel, PowerPoint, Outlook and Teams, working across their own email, meetings and files. On Microsoft's UK page as of October 2026, Microsoft 365 Copilot Business is £16.10 per user per month paid yearly, ex VAT (£19.32 monthly), and existing business customers committing to a year between 1 July and 31 December 2026 can get the first year at £13.80. Don't buy for everyone; the pilot tells you who'd miss it. I've compared the options in ChatGPT, Claude or Copilot: which should your business pay for?
Is it really an AI job, or an automation job?
Worth asking first. If the job follows the same steps every time, from data already in a system (the overdue invoices list each morning, a nudge when a quote has gone quiet, the weekly sales email), a plain automation does it the same way every time, and you can check it once and trust it. AI earns its place where the input is messy: long text, a mix of documents, something a person has to read and interpret.
I've written about what a small business can automate separately, and my automation check (£295, credited in full against a build) is a 90-minute call that ends with a written list of up to three automations worth building, each with a fixed price.
What does a good first month look like?
Week 1: choose and set up. Pick the job, the owner and about five people who do it. Note roughly how long the job takes now. Approve the tool and send round the paste rule.
Weeks 2 and 3: use it on real work, not demos. A ten-minute check-in each week: what saved time, what came back wrong, what someone worked out that others should copy.
Week 4: decide. Compare the time with your week 1 note, list the mistakes the checks caught, and keep it, change it or drop it. If you keep it, write a paragraph on how it's done so the next person doesn't have to work it out.
Then pick job two. The best sign the first one worked is that the people in the pilot would complain if you took it away.
Where can I get help with the first step?
That's what my AI Advisor hour is for: £195 for an hour-long call and a written summary afterwards, covering which tools are worth paying for, whether your data is ready, what to keep away from AI, and where it would save real time. That includes the cases where the honest answer is to wait, or to use a plain automation. If you want someone to ask every month while you roll it out, the monthly advisor starts at £350. Not sure you need either? A free 20-minute intro call will tell you.
Common questions
How do I start using AI in my small business?
Pick one job that happens every week, is mostly reading and writing, and is already checked by a person, such as meeting notes or proposal drafts. Write down what information it needs and a one-line rule for what staff may paste in. Run it for a month with about five people on a business AI plan you already have, then measure and decide.
How can AI help a business with 20 to 200 staff?
Mostly by taking first drafts and summaries off people's plates: meeting notes, proposals, long email threads and month-end commentary. The person still checks and sends the result. Decisions about people, unchecked figures and regulated processes are better left until the basics are working.
Do we need an AI policy before we start?
You need a rule before you start, and it can be one sentence: company work goes into the approved business tool on a work account, and client personal data, payroll, passwords, unreleased results and NDA material never go into any AI tool. A one-page policy can follow within weeks. DSIT's June 2026 survey found only 17% of small and 22% of medium AI-using businesses had a formal, written AI policy.
Which AI tool should a small business use first?
Usually the one you already pay for. Microsoft 365 business plans include Copilot Chat at no extra cost, and Google Workspace includes Gemini. Buy paid seats only for the people who'll use them most, after a short pilot shows who that is.
How much does it cost to start using AI?
It can start at nothing extra on Microsoft 365, because Copilot Chat is included. A paid Microsoft 365 Copilot Business seat is £16.10 per user per month paid yearly, ex VAT, on Microsoft's UK page as of October 2026, with a £13.80 first-year offer on annual commitments made by 31 December 2026.
How do we know if AI is actually saving time?
Note how long the job takes before the pilot, then compare after a month of real use. Count the mistakes the checks caught as well as the time saved. If the people using it would object to losing it, it's working; if they've quietly stopped, try a different job.