AI for business: where it earns its place and where it doesn't

This guide to AI for business is for the owner or director of a UK small or mid-sized business who wants to know where AI pays its way today, where it doesn't, what data protection law expects, and how to try it without risking anything that matters.

We design, build and run AI in software, including in products of our own. Most of this guide is about buying tools off the shelf, the right answer for most businesses most of the time. The last part is about AI automation built into the systems you already run, which only makes sense for certain jobs.

How far UK businesses have got

The Office for National Statistics' analysis of AI in UK businesses from 2023 to 2026, published in July 2026, found that around 35% of businesses with 10 or more employees reported using at least one AI technology in late June 2026, up from around 12% in autumn 2023. Use rises with size: 28% of businesses with fewer than 10 employees reported it, against 49% of those with 250 or more.

Most use is shallow. Of businesses with 10 or more employees that use AI, only 10% say they use it extensively, and the most common purpose is improving how the business runs. If you have not started, you are not far behind; the useful question is which jobs to start with.

AI for small business: jobs worth paying for

AI tools are good with language: reading messy text, summarising and sorting it, and writing a first draft. They earn their place where staff spend hours on that kind of work and a person can check the result quickly. Each job below has a tool you can buy this week, priced from the vendor's own page.

Two cautions. Fin counts an outcome when a customer confirms their issue is resolved, when they don't ask for more help after Fin replies, or when Fin completes a workflow you set up, including one that ends in a handoff to a person. Work out what that means for your volume before you sign. And the office tools are only as safe as your file permissions: Microsoft says Copilot only surfaces data to which each user has at least view permission, so a folder shared too widely becomes a folder Copilot can quote from. Tidy the sharing first.

Where AI doesn't earn its place

The same tools can be wrong while sounding certain. In June 2025 the High Court, in Ayinde v London Borough of Haringey, said freely available tools such as ChatGPT "are not capable of conducting reliable legal research", and added: "They may cite sources that do not exist." Microsoft's own Copilot documentation says generative AI responses "aren't guaranteed to be 100% factual". So keep AI away from these jobs, or keep a person firmly in charge of them:

• Facts that go out unchecked. Prices, tax positions, contract terms, anything a customer or a court will rely on.
• Decisions about people. Hiring, credit, pay and discipline carry legal duties, covered below.
• Jobs a rule already does. VAT sums, stock lookups and due dates belong in ordinary software, which gives the same answer every time.
• Rare jobs. If something happens twice a month, setting AI up and checking it costs more than it saves.
• Places where a wrong answer is worse than none. When we built the AI that reads job ads for our own job board, we decided it should never estimate a salary an ad doesn't state. A missing salary is better than a wrong one shown as an employer's pay.

The data protection basics

UK GDPR applies to AI like anything else that handles personal data. The Information Commissioner's Office (ICO) publishes guidance on AI and data protection, which it says is under review after the Data (Use and Access) Act 2025. The basics for a smaller business:

• Use business plans, not personal accounts. Google's Workspace pricing page says company data is "not used to train models". Anthropic's pricing page lists model training on Claude Team as "none by default", but on its individual plans as opt-out. Microsoft says Copilot prompts and responses aren't used to train its foundation models.
• Have a contract with the provider. The ICO's guidance on accountability says a provider with no purpose of its own for your data, acting only on your instructions, is likely to be your processor, and Article 28 of UK GDPR says processing by a processor "shall be governed by a contract or other legal act". Check the provider's business terms include one.
• Expect to do a DPIA. The same guidance says that "in the vast majority of cases" using AI involves processing likely to result in a high risk, which triggers the legal requirement for a data protection impact assessment. If you decide yours isn't high risk, record how you decided.
• Take care with decisions about people. The Data (Use and Access) Act 2025, whose data protection provisions are now all in force, widened the lawful bases for significant decisions made solely by automated means. The ICO adds "so long as you continue to apply appropriate safeguards", and the change doesn't cover special category data such as health information. The simpler course is to let AI sort and draft, and let a person decide.
• Tell people and send less. If AI handles customers' or staff's personal data, your privacy notice should say so, and nobody should paste in more personal data than the job needs.

This is a starting point, not legal advice. The ICO's guidance has the detail.

How to start: one process, a person checking, measured

1. Pick one job. Choose work that happens often, is mostly reading or writing, and where a mistake is cheap to catch. Summarising enquiries or drafting first replies for a person to send is a better first job than anything customers see directly. If you're not sure where the time goes, map the process first.
2. Measure it now. Count how many a week, how long each takes and how often one goes wrong. Without a before, you can't tell whether the tool helped.
3. Run it with a person checking every result. For the first few weeks, someone reviews every output before it is used and notes each correction.
4. Measure again and decide. If corrections are rare and the time saved is real, move to checking samples and anything the tool was unsure of. If not, stop. A tool whose answers all need rewriting saves nothing.
5. Write down the rules. Which tools, which data, who checks, and who to tell when it gets something wrong.

When to build AI around your own systems

Off-the-shelf tools work inside their own apps, when a person asks. When the work lives in your own systems, arrives all day and has to be right every time, it can be worth building AI into the process itself, so it runs on every document or message as it arrives (our guide to AI automation inside a business process covers how). The signs it is time to integrate AI into your business systems rather than buy another tool:

• The same job runs hundreds of times a week, and staff copy results from an AI tool into another system.
• The data the job needs sits in your own database, order system or CRM, where no off-the-shelf tool can see it.
• A wrong answer costs money or trust, so you need tests, checks and a record of every decision.
• An outage would stop the work, so you need a fallback.

4dayweek.io, the job board we built and run, is an example. It collects jobs from more than 1,300 company job boards, and three AI requests read each ad: one for the role, level and skills, one for the location text alone, and one for the salary. The location reader never sees the company name, so it can't guess the head office. The salary reader takes base pay only, and a guard drops implausible figures before they are stored. Two AI providers cover for each other, an alert fires if a quarter or more of requests fall back to the second, and tests run against the real models before any change to a prompt, model or provider goes live. None of that is possible in a chat window. The 4dayweek.io case study has the detail.

For most small businesses, building is not the first step. Buy the tool, prove the job is worth doing, and build only when the volume, the data or the cost of a mistake calls for it. Our bespoke software examples show the same build-or-buy test applied more widely.

If you want a hand

If a job has outgrown the tools, or you want a second opinion before you buy, start a project and tell us what the work is.