AI for business: a practical guide for UK small businesses

AI for business today mostly means AI assistants like Microsoft Copilot, ChatGPT and Claude - tools that draft emails, summarise documents, chase invoices and prepare reports, with a person checking what comes out. For a UK small business the first job is getting the AI your staff already use onto an approved business plan. The second is connecting it to your own systems, so its answers are about your business rather than businesses in general.

It's written for owners and directors of UK firms with roughly 10 to 100 staff who don't have time to become AI experts.

Probably. In research for Microsoft UK by Censuswide (2,003 UK employees, published 13 October 2025), "71% of UK employees have used unapproved consumer AI tools at work, and 51% continue to do so every week". And they're using them for pretty much the jobs you'd expect - 49% for drafting workplace communications, 40% for reports and presentations, and 22% for finance-related tasks.

This usually isn't bad intent - people just want to work faster. The problem is where the data goes:

• Business data ends up in personal accounts. A quote or a customer complaint pasted into someone's free ChatGPT account sits outside your business, under that person's own terms.
• Personal plans can train on it. On free and personal plans, whether conversations are used to improve the provider's models comes down to a setting each person controls, not you (details per provider below).
• There's no record. You can't see what was shared or by whom, and you can't delete it when someone leaves.
• It's still your responsibility. The ICO's guidance on AI and data protection is clear that "overall accountability for data protection compliance lies with the controller, ie your organisation". If someone on your team puts customer personal data into a tool you haven't checked, it's still your business's problem under UK GDPR.

That isn't a reason to ban AI (which tends to just push it further out of sight). It's a reason to give people an approved tool that's at least as good as the one they're using now.

For most SMEs this is a handful of decisions rather than a project:

1. Pick one approved assistant on a business plan. You get business data terms and central admin, and for the three below, no training on your content by default.
2. Write a one-page staff AI policy. Which tool is approved, what it's for, that a person checks anything before it reaches a customer or your accounts, and who to ask if they're not sure.
3. Decide what data can go in. A simple traffic light works well. Public information is fine anywhere. Internal documents go in the approved tool only. Customer personal data, staff records and anything under NDA only go where you've checked the terms cover it, or not at all.
4. No personal accounts for work. Ask staff to move work over to the approved tool and delete work chats from their personal accounts.

The jobs that pay back quickest all work the same way - the AI prepares it and a person approves it.

• Drafting emails and replies in your tone, from a few bullet points.
• Summarising long email threads, meeting notes and contracts.
• Chasing overdue invoices. Connected to Xero, it can list what's overdue and draft a chaser for each customer, for someone to check and send.
• Reporting. First-draft commentary on this month's figures.
• Quoting. Drafting a quote from your price list and the customer's request, as long as it knows how you price (more on that below).

Where it falls down is anything that needs facts it hasn't been given. It'll answer just as confidently about your margins whether it can see the numbers or not.

Microsoft Copilot - the natural fit if you run on Microsoft 365. It works where your team already is (Outlook, Teams, Word, Excel and SharePoint), and your IT admin manages it alongside the rest of Microsoft 365. Microsoft has renamed Microsoft 365 Copilot to Microsoft Copilot, so you'll see both names.

• Enterprise data protection. For work accounts, Microsoft says Copilot is covered by its Data Protection Addendum, respects your permissions, and doesn't use prompts, responses or the Microsoft 365 data it accesses to train foundation models.
• You can start for free. Copilot Chat is included with eligible Microsoft 365 subscriptions and gets the same data protection when staff sign in with their work account.

Now the limits. Microsoft's own Copilot Chat FAQ says that without a paid licence, Copilot Chat can't access your company's shared data, each person's own mail and files, or data from connected systems - staff can only upload files by hand. The paid licence, Microsoft 365 Copilot Business (firms up to 300 users), is £16.10 per user a month paid yearly, or £19.32 paid monthly, ex VAT, on top of your Microsoft 365 plan. Microsoft is running a first-year price of £13.80 on annual billing until 31 December 2026. And because Copilot sees whatever each user can see, tidy SharePoint permissions first, or the salary spreadsheet in the wrong folder becomes easy to find.

ChatGPT Business - when your team already uses ChatGPT. Moving people off personal accounts onto a Business plan is often the least disruptive fix. OpenAI says ChatGPT Business and Enterprise data isn't used for training by default. It's priced per user per month, so check OpenAI's pricing page for the current UK price.

Claude Team - for firms not on Microsoft 365, or who want more connections. In our own checks Claude had the widest range of ready-made connections of the three. Anthropic says it won't use inputs or outputs from its commercial products (Claude for Work and its API) to train its models by default. The Team plan is $20 per person a month paid yearly, or $25 monthly for a standard seat, priced in US dollars.

Business plans vs personal plans. On ChatGPT's personal plans, conversations can be used to improve OpenAI's models unless each person turns that off. On Claude's (Free, Pro, Max) it depends on each person's privacy settings. On a business plan, the business sets the rules.

In short:

• On Microsoft 365: Copilot Chat for everyone, paid Copilot licences for daily users.
• On Google Workspace, or wanting more connections: Claude Team or ChatGPT Business.
• Team already living in ChatGPT: ChatGPT Business.

Prices checked 6 October 2026. All three providers changed names and prices during 2026, so it's worth checking before you buy.

Out of the box, an AI assistant knows a lot about businesses in general and nothing about yours - how you price a job, which customers pay late, what your margin is on each product line. Ask it to price a quote or forecast next month's cash and it'll still give you an answer. Generic at best, wrong at worst.

Pasting information in every time is slow and puts more data into chats, and if two people paste in different versions of the same spreadsheet they'll get two different answers. In the government's AI adoption research (DSIT, 3,500 businesses), 29% cited AI being too complex or difficult to integrate and scale as a barrier. The assistant is the easy bit. Connecting it to everything else is where it gets hard.

The firms getting real value out of AI have systems that play well with it. In practice that means:

• Data in one place rather than scattered spreadsheets. If prices are in one spreadsheet, jobs in another and customers in a third, no assistant can reliably join them up.
• Connections to your real systems. Most assistants now connect to other software through a shared standard, the Model Context Protocol (MCP), which Anthropic created and handed to a Linux Foundation project in December 2025. Think of it as USB-C for AI - one plug shape, so any assistant can use any system that offers it.
• Sensible permissions. The assistant should see what the person asking is allowed to see, and nothing more.
• One shared source of truth. One agreed definition of margin, overdue or a won job, so everyone gets the same answer.

Ready-made connectors only get you so far. When we checked on 3 October 2026, Xero, HubSpot and Microsoft 365 had them in all three assistants. Sage Accounting, Sage 50, FreeAgent, Simpro, Tradify, Commusoft and Procore had none, and neither does the pricing spreadsheet on your shared drive. Each connector covers one system at a time, so a question like "which customers do we make least money on?" (which needs accounts, CRM and job records together) usually can't be answered.

Connections do bring one new risk. An assistant that can read your inbox, see business data and send messages can be tricked by a cleverly worded email into sending data out (Simon Willison calls this the "lethal trifecta"). Start read-only and keep a person approving anything that sends, posts or changes records.

This is the work we do. We build the systems and custom connections (MCP servers) that give AI the right context - your prices and rules, one store that joins Xero with your CRM, job system and key spreadsheets, and connections for systems that don't have any. They work with Copilot, ChatGPT or Claude, start read-only, and keep a person approving anything that changes data. For one UK specialist supplier, pricing a quote went from about 1 to 2 hours to about 5 minutes.

Builds are fixed-price from £3,000, with optional care plans from £250 a month, ex VAT (see pricing). Weighing up outside help? Have a read of what to ask an AI automation agency. Reporting still in Excel? Start with automating your Excel reports. Running a practice? See AI for accountants.

• Ask your team (without blame) which AI tools they're using for work.
• Choose one assistant on a business plan and write the one-page policy.
• Pick one job, like chasing overdue invoices, and connect one system read-only to try it on.

Get the free step-by-step guide to connecting ChatGPT, Claude or Copilot to your business systems → Free - just enter your name and email to open it.

Or if you'd rather talk it through, book a free 30-minute review. Show us one job your team does by hand and we'll tell you honestly whether AI, a custom connection or neither is the right fix.