Most small firms use AI beside their systems, not inside them, and that's where the hours go

By Jason Sibley

Plenty of small firms have tried AI. Fewer have put it inside the software they run on. A government survey shows the gap and, I think, where a small team could find capacity without hiring.

What the government survey found

The government's science and technology department (DSIT) published the UK Business Data Survey 2026. Fieldwork ran from October 2025 to January 2026. It covers businesses handling digitised data: 86% of UK businesses, according to DSIT.

The main figures, with their bases:

• Of businesses handling digitised data, 41% used AI-based technologies. That included 51% of small businesses and 41% of micro businesses; for sole traders, it was 40%. DSIT notes that other estimates vary and adoption is hard to measure consistently.
• The most common reasons were researching information (28%, for example instead of a search engine) and summarising in-house information or drafting reports and correspondence (21%).
• Among AI users, 21% said their tools were integrated into existing business systems. That included 31% of small businesses and 27% of micro businesses; for sole traders, it was 18%.
• Integration varied by sector: 39% in information and communication. Construction stood at 14%, manufacturing at 12%.

DSIT says integration "may be limited", possibly reflecting differences in firms' capability and resources.

Among AI users, firms with integrated AI were more likely to analyse data (62% against 30%) and collect data (55% against 37%). DSIT says this "may suggest an association". It doesn't establish cause. Firms already keeping tidy data in proper systems may be the ones getting AI working inside them. The survey can't tell us whether integration made the difference.

Beside versus inside

The survey's examples include Microsoft Copilot within Microsoft 365 and AI embedded in CRM or finance systems. It also mentions AI within workflow or productivity platforms.

Beside means opening a chatbot, copying text from an email or spreadsheet, then copying the answer back. The survey's top two uses (research and drafting) fit that pattern, though it doesn't say how people do it.

This is my reasoning, not the survey's. That approach carries a copy-paste tax: gathering context and moving results by hand. The AI lacks the wider customer record or invoice history. Someone must track versions and decide what's safe to paste, as discussed in shadow AI in small teams.

Inside the tool, context is already there and output lands where work happens. That's where I think capacity gains come from. Neither I nor the survey can give you hours saved. But you can count the extra steps in your own week.

Four questions before you switch anything on

• What's already in what you pay for? Check the provider's pages. Some plans include AI; others charge extra. Some don't offer it.
• What data can it see? Check permissions and sharing settings. An assistant inside Microsoft 365 or your CRM uses data in that system. The Microsoft 365 Copilot settings check walks through this.
• Who checks the output? Name a person. AI drafts can sound confident and still be wrong.
• How will you know it saved time? Time one task now, then check again after a few weeks.

A 30-minute look at what you already have

Step 1: List what you pay for (10 minutes)

Use bank statements or your accounts package to list software subscriptions, including the plan name and users. A partial list is useful.

Step 2: Ask an AI to sort the list (10 minutes)

Use this prompt for features to check and questions for providers. Keep confidential information out, including customer names and passwords.

I run a small business in the UK. Below is the software we pay for, with plan names where known.

[Paste your list here: software name, plan, who uses it, what we mainly use it for]

For each tool, do two things:
1. Name the kinds of AI feature I should look for in that type of software, such as drafting or summarising.
2. Write questions for the provider or settings checks. Is the feature included in our plan? What data can it see, and where is that data processed?

Do not guess or invent feature names or details of prices and plans. If unsure whether a feature exists or is on our plan, write "UNKNOWN, check with the provider". Mark missing information "[OWNER TO FILL]" rather than assuming. Use plain British English and keep it short enough to read in one sitting.

Step 3: Check the answers against the real source (10 minutes)

Treat the output as a to-do list. Pick the two tools your team uses most and verify claims against provider help pages or admin settings.

For accounting software, read the guide to AI for bookkeeping and invoices. For a CRM, try the HubSpot AI tiny pipeline guide.

For Google's email and documents, there's a guide to Gemini in Workspace and Gmail.

Choose a single task to try and name who checks the result.

Write the rule down

DSIT found that 17% of AI-using businesses had a policy or guidelines: 5% had a formal written policy and 12% an informal one (base: 1,870 AI-using firms). Most therefore reported no policy or guidelines. That matters more once AI can see customer and finance data.

Keep the policy short: name approved tools and say what must never be pasted or shared. Specify who checks output. Our simple AI policy for small teams offers a starting point. This isn't legal advice. If rules touch client contracts or personal data, consult a solicitor.

If you'd like company on this

At My AI Apprentice we run AI sessions for small businesses, including working through what's already in the tools you pay for. No pressure either way.