AI for small business: where to start, what it costs and what to skip

What is AI for small business?

AI for small business means software that reads and writes on your behalf: it pulls figures off a receipt, sorts a messy inbox or drafts a reply in your tone. Automation moves the result into Xero, a calendar or a spreadsheet. For most owners, the practical use is one dull, repeated job that stops eating the week.

You do not need a strategy document to start, and nothing here involves a robot running your firm. The aim is smaller: get a few hours back and make fewer slips.

The Office for National Statistics runs a voluntary survey of UK businesses every fortnight. In its September 2026 results, 30% of businesses said they used at least one type of AI technology.

For businesses with fewer than 10 employees the figure was 29%, and for those with 10 or more it was 38%. That is self-reported, so treat it as a rough picture rather than an exact count.

Using AI is not the same as getting value from it. A free chatbot used now and then counts in those figures. The harder part is fitting it into how your business runs, and that is where this guide spends its time. If you prefer to see one sector in detail, the AI for accountants guide goes further, and the guides hub lists the rest as they publish.

How can AI be used for small business?

Use it on jobs that repeat, involve a lot of reading or typing, and have a clear way to check the result. Replying to enquiries, reading invoices and receipts, chasing late payments and writing weekly summaries fit best. Judgement calls, money decisions and legal questions stay with a person.

Different sectors lean on different rows of that table. The sector pages show how it looks for accountants first, then trades, solicitors and salons. The AI automation services pages cover the four kinds of job I build most often. Letting agents, dental practices and financial advisers have their own pages too: letting agents, dental practices and financial advisers. If missed calls are the problem, the AI receptionist answers them and takes messages.

What does a worked example look like?

Take a small accountancy practice. Clients send receipts by WhatsApp, email and post, and someone loses most of Monday keying them into Xero and chasing the people who have sent nothing. That is a typical fit: the steps are the same every time, the input is easy to read and every posting can be checked.

Illustrative, not a client result.

Notice what stays with a person. Anything unclear waits in a short review list, and the team checks what lands in the accounts. The saving comes from not typing and not chasing, not from removing anyone's judgement.

A practice that size is a bigger job than a plain enquiry set-up for a sole trader. What you pay depends on how many systems are involved and how many odd cases there are.

What should you automate first?

Start with the job that costs you the most evenings and has the least risk if it goes slightly wrong. One job, not five. You learn how it behaves, you trust it, and the first project pays for the second.

Four questions help you pick:

• Does it happen at least weekly, in roughly the same way each time?
• Does it take you more than an hour a week?
• Is the input something a person could read in seconds, such as an email, a form or an invoice?
• If it gets one wrong, can you spot it and fix it before it reaches a customer?

If you answer yes to all four, it is a good first candidate. If the answer to the last one is no, keep a person in the loop, or leave that job for later.

Be wary of the exciting job. A chatbot on your website sounds modern, but the unglamorous ones usually pay back faster: a reply to every enquiry within minutes, invoices that go out the day the job is done, a reminder that chases for you.

Where do you start, step by step?

You do not need to commit to anything big. This order keeps the cost and the risk low.

1. Track one week. Note every repeated job, how long it took and what it needed (an email, a form, a login). Do not fix anything yet.
2. Pick one job. Use the four questions above. Choose the one that is frequent, boring and checkable.
3. Check what you already own. Your accounts package, email and calendar may already have AI features. Try those before paying for anything new.
4. Try it by hand on real examples. Put ten real, non-sensitive examples through an assistant such as Claude or ChatGPT and see how many it gets right. If it is not close, no amount of plumbing will rescue it.
5. Decide how it will run. Either it is built in your own automation account, or someone hosts and looks after it for you. The next section explains both.
6. Keep a person on the output at first. Let it draft, and approve each one for a few weeks. Loosen that only when the results have earned it.

If you would rather have the ranking done for you, that is what an AI audit is for. If you already know the job, you can go straight to a build.

Should you use the AI you already own or build something custom?

Try what you already own first. If Xero, Microsoft 365 or Google Workspace already does the job well enough, a custom build is money spent for little gain. Build only for the gaps between your tools.

Built-in features are cheap and need no set-up. They are also generic: they work inside one product and rarely reach across to the next. A built-in assistant can summarise an email in Microsoft 365 or Google Workspace, but it will not read an invoice that arrives there and post it to Xero for the right client.

• Use what you own when the job lives inside one tool and the built-in feature does it well.
• Use a general assistant such as Claude or OpenAI's models by hand when the job is occasional, such as drafting or summarising.
• Consider a custom build when information has to move between tools, the job repeats weekly or more, and a person would otherwise retype it.
• Consider a private set-up when the files are sensitive and you do not want them leaving your control.

A useful test is to ask who does the joining up today. If the answer is you, copying and pasting between two screens, that joining is the part worth automating.

What can AI not do?

AI cannot be right every time, and it will not tell you when it is wrong. These are the limits worth knowing before you spend anything.

• It can be confidently wrong. A figure, a date or a name can come out looking perfect and be incorrect.
• It does not know your business unless you tell it. It will guess at your prices, your hours and your policies if you let it.
• It copes badly with unusual cases. The tenth odd request in a hundred is the one that needs you.
• You still have to check it. The work shifts from typing to reviewing.
• It will not fix a muddled process. If nobody agrees how a job should be done, automating it makes the muddle faster.

There are also jobs where the right answer is not to use it. If you handle a dozen enquiries a month, a good template and a calendar link may do more than any automation. If a job needs a judgement call every time, it is not a candidate.

Who should wait? Anyone whose process is still changing every month, because you would be rebuilding as fast as you build. Also anyone with no one who can spot a wrong answer, or whose job is so rare that the time saved will never repay the cost. I would rather tell you that on a free call than sell you a build.

What does AI cost a small business in the UK?

There are two kinds of cost, building the thing and running it, and they are easy to mix up.

Building. My prices start from £495 for an AI audit, from £1,500 for a build and from £95 a month for support. These are starting points, not fixed prices, and the quote follows the call. The pricing page has the detail.

Running. The tools it runs on, such as an automation platform or an AI service, are paid for directly by you, so the accounts stay yours. What they cost depends on the tools the job needs, and I give you the expected figure before you agree to anything.

Here is what the main automation platforms charge, as published on their own pricing pages (checked 26 October 2026). Prices change, and they are quoted in each platform's own currency, so euros and dollars appear side by side.

AI usage itself, such as the charge for the language model that reads a document, is usually small for a business of this size, but it depends on volume. Ask for the expected monthly figure for your own job rather than relying on a general number.

Who builds it, and who looks after it?

There are two ways to run an automation, and I explain both on the first call so you can choose.

• Built in your account and handed over. I build it in your own n8n, Make or Zapier account and hand it over. You own it, and you can add optional support from £95 a month or look after it yourself.
• Hosted and managed by me. I run it for you on a monthly care plan.

The first suits an owner who likes to understand what is running. The second suits one who would rather not think about it.

It connects to the tools you probably already use. The ones I can connect are Xero, QuickBooks, Dext, Sage, FreeAgent, Karbon, Senta, Microsoft 365 and Google Workspace. If your tool is not on that list, ask before assuming it cannot be done.

Is it safe to use AI with customer and business data?

It can be, if you choose the right kind of account and share only what the job needs. Typing customer details into a tool whose terms you have not read is where most of the risk sits.

Consumer chatbots, free or paid, can use what you type to improve their models, depending on the settings. Business and API products usually work differently.

Anthropic, for example, says in its privacy centre that it will not use inputs or outputs from its commercial products for training by default, with exceptions such as feedback you choose to submit (page updated 18 August 2026). Check the current terms of whichever provider you use, because they change.

If you handle personal data, UK GDPR applies to how you use AI as it does to everything else. The Information Commissioner's Office publishes guidance on AI and data protection, though at the time of writing its page is under review because of the Data (Use and Access) Act. This is general information, not legal advice.

The private option costs more than a standard build, and the price depends on how many documents and people are involved. A private AI assistant starts from the same £1,500 as any build, and a written quote follows the call.

If your data is sensitive, it is worth the conversation. If it is not, a business plan with sensible habits is usually enough. Whichever route you take, I can write down which tools touch which data. That is not a compliance promise, because compliance depends on your own business and your regulator. The GDPR and AI guide goes through the rules in plain English, and AI for law firms covers the extra care that client confidentiality needs.

How do you know it is working?

Measure before and after, in hours and pounds, using your own numbers. If a job took three hours a week and now takes one, that is two hours back. If you cannot measure it, you cannot tell whether the cost was worth it.

Check three things for the first month. Is the output right? Is anyone having to redo it? Is it still running? A short weekly look at what it handled, and what it left for a person, is enough.

Be ready to turn it off. A good automation is one you can pause without breaking anything else. If it is not earning its keep after a fair trial, stopping is a sensible result, and it is cheaper to find out early.