How Much Does AI Automation Cost a Small Business? Real Numbers

Most articles on this question refuse to give numbers. Here are numbers.

For a UK small or mid-sized business in 2026, AI spend comes in three tiers: off-the-shelf assistants at roughly £15–£60 per user per month; connected automation — AI wired into your actual systems for one process — typically £5,000–£25,000 as a one-off project; and ongoing support to keep it working, usually a monthly retainer in the hundreds to low thousands depending on scope. Which tiers you need, and in what order, depends on what you're automating — but the shape of the spend is that consistent.

The rest of this guide is what's inside each number, what quietly inflates them, and the arithmetic for deciding whether any of it is worth paying.

Tier 1: off-the-shelf assistants (£15–£60 per user per month)

ChatGPT Team, Claude, and Microsoft Copilot all sit in this band. For that you get drafting, summarising, meeting notes, research assistance, and general document work — the tasks that live in a browser tab, disconnected from your systems.

For most SMEs this tier is worth having almost regardless of strategy: if each user saves an hour a week, the subscription pays for itself several times over at any realistic wage. The trap is stopping here and concluding AI "doesn't do much" — a disconnected assistant can only ever talk about your business, not act in it.

Two cost notes. First, buy the business tier, not the consumer one: the data protection terms are different, and for anything touching client information that difference matters. Second, count only seats that get used. The commonest waste in this tier is licences bought in an enthusiastic January and unopened by March.

Tier 2: connected automation (£5,000–£25,000 per process)

This is the step that changes the economics: AI connected to your CRM, accounting package, inbox, or practice system — with permissions, audit trails, and a human approval step where it matters — so a process runs instead of being talked about. Enquiry triage that updates the CRM itself. Invoice chasing that actually goes out. Recall lists that fill the diary.

What moves the price inside the range:

• Number of systems touched. One system, one direction is the bottom of the range. Three systems with data flowing both ways is the top.
• How the systems connect. Tools with existing MCP servers or good APIs are cheap to reach; legacy or niche software needs custom connector work. (Check the directory — if your tools are there, that's money off.)
• Data sensitivity. Health, financial, or legal data adds governance work — DPIAs, data processing agreements, tighter permissions. Real cost, non-optional.
• Approval workflow. "Draft it and a human approves" is cheaper than "act autonomously with guardrails", and is the right starting point anyway.

A fair rule of thumb: a single well-scoped process on mainstream tools lands £5,000–£10,000; the £15,000–£25,000 end is multi-system work or regulated data.

Tier 3: keeping it working (the part budgets forget)

Models change under you. APIs deprecate. The process itself drifts. An automation built in January and untouched thereafter degrades — sometimes visibly, sometimes silently, which is worse. Ongoing support — monitoring, fixing, and extending down your roadmap — is a real line item, typically a few hundred to a few thousand pounds a month depending on how much is running. Budget it from the start; an automation with no owner is a liability with a delay on it.

The hidden costs (and the hidden saving)

Three things inflate real-world spend beyond the invoice: your team's time for testing and feedback during the build (short but not zero); data cleanup, because automating a process exposes every inconsistency in the data feeding it; and tool sprawl — five overlapping £30/month subscriptions bought before anyone mapped what was actually needed.

Against that, the arithmetic on the other side. A task taking 5 hours a week at £20/hour fully loaded costs about £5,200 a year. Automate 80% of it and a £6,000 build pays back in roughly 18 months — before counting speed, consistency, or the work that person does instead. Run this sum per process, honestly, and the decision usually makes itself. Our guide to proving the ROI of AI automation turns this into a proper measurement approach.

When it's not worth it

Skip automation where volume is low (a task done monthly rarely repays a build), where the process is broken (automating chaos produces faster chaos — fix it first), and where judgement is the product (AI prepares, humans decide). And treat "we'll build everything at once" as the most expensive sentence in the field: one process, measured, then the next — the case for that sequencing is in where to start with AI automation.

Budgeting it properly

For a typical 10–50 person firm starting sensibly: Tier 1 seats for the people who write for a living, one Tier 2 build on the highest-value process, support from month one — first-year total somewhere between £10,000 and £30,000, with the payback visible inside the year if the process was chosen well.

Choosing it well is the actual trick, which is why we start with a map of the whole business rather than a tool list — and why the broader context in our full guide to AI for SMEs is worth reading before any invoice gets signed. The two-minute readiness assessment will tell you which tier you're actually ready for.