AI for SMEs: a practical guide

Most writing about AI for business is aimed at enterprises with data teams and seven-figure budgets, or it is a vendor pitch wearing a how-to guide as a disguise. If you run a firm of five to two hundred people, neither is much use. You do not have a data team. You have a business to run, a stack of tools that mostly work, and a nagging sense that some of what your team does by hand every week should not be done by hand.

This guide is the missing middle: what AI genuinely does well for an SME today, what it costs, where to start, and what the law expects — based on what we see building and supporting this for firms across accounting, legal, healthcare, recruitment, and professional services.

What AI can actually do for an SME right now

The honest headline: today's AI is exceptional at language and document work, good at moving information between systems, and unreliable at anything requiring judgement it hasn't been given the context to make. That maps to four areas where SMEs get dependable value.

Administration and operations

Meeting notes and action points written up before you're back at your desk. Inboxes triaged and drafts prepared for the routine 70% of replies. First drafts of proposals, reports, and letters that a human edits rather than writes. This is the least glamorous category and consistently the fastest payback — here are five patterns we see saving teams 10+ hours a week.

Customer-facing work

Routine questions answered instantly and accurately from your own documentation; the tricky ones routed to a human with the context attached. Enquiries qualified and logged in the CRM before anyone picks up the phone. The pattern that works is AI handling the volume, people handling the exceptions.

Finance and back office

Bank reconciliation suggestions, invoice chasing sequences, expense categorisation, supplier document extraction. Because this work is rules-based and high-volume, it automates well — and because it touches money, it is exactly where governance matters most. Our guide to AI for accounting firms covers what's realistic here in detail.

Connecting AI to the tools you already use

The step that turns a chatbot into a colleague is giving it governed access to your real systems — your CRM, your accounting package, your project tracker. An open standard called the Model Context Protocol makes this practical without a bespoke integration project: browse the 160+ integrations in our MCP server directory, or start with how to connect AI to your business tools without writing code.

What it costs — real numbers

Three tiers, in practice. Off-the-shelf assistants (ChatGPT Team, Claude, Copilot) run £15–£60 per user per month and are worth having almost regardless. Connected automation — AI wired into one of your actual processes with permissions and audit trails — is typically a £5,000–£25,000 project depending on the systems involved. Ongoing support — keeping it working as models and APIs change, and working down the roadmap — is a monthly retainer. The expensive failure mode is none of these: it's six months of licence fees for tools nobody uses because they don't fit the process.

Where to start (and where not to)

Don't start with a tool. Start with a process — one that is high-volume (happens daily or weekly), rules-based (a competent temp could follow the instructions), and low blast radius (a mistake is annoying, not existential). Automate one step, measure the result, widen from there. We've written a fuller framework for picking the first process and for proving the ROI once it's running.

The reason we start every engagement with AI Mapping is that the highest-value process is rarely the one that first comes to mind — and the prioritised roadmap stops you automating a process that should have been fixed instead. Two minutes with the AI readiness assessment will tell you roughly where you stand.

The rules: UK GDPR, the EU AI Act, and governance

Two legal facts matter for UK SMEs. First, UK GDPR applies now: automated decisions about people — customers, candidates, staff — carry obligations today, and the ICO has been explicit that AI tools do not suspend them. Second, the EU AI Act reaches across the Channel: sell AI-powered services into the EU, or use AI whose output lands there, and it can apply to you. Its transparency duties have been live since 2 August 2026 — a chatbot must say it is a machine — while the high-risk obligations covering hiring and credit were pushed back to December 2027. Here's how to work out whether it reaches you.

Governance doesn't need to be heavy — it needs to exist: who approved the tool, what data it can see, who checks its output, and what happens when your vendor has a bad day. Our guide to governing AI automation without killing it covers the practical version; for boards and exec teams there's a dedicated briefing service.

Build, buy, or partner?

Buy where your need is generic (transcription, drafting). Build where the process is your competitive edge and the systems are yours. Partner when you need building done without hiring for it — which is the gap Crox exists to fill: we map, build, and support, in that order. The fuller decision framework is in Build, Buy, or Partner?

Guides for your industry

The general principles land differently in each trade, so we write them up one industry at a time: accounting firms, legal practices, recruitment agencies, estate agents, dental practices, GP practices, construction firms and trades, manufacturers, logistics and haulage, hospitality, and architects — with more on the way in the insights section.

Frequently asked questions