AI for UK SMEs: agents for customers, augmentation for staff

AI pays in a UK SME in two places. The first is in front of your customers: agents and automation that finish their journey for them, and new ways to explore what you sell. The second is beside your staff: tools that save them time and check the quality of their work. It rarely pays as a strategy, a chatbot for its own sake or a fix for messy data.

Most SMEs have already done the easy half. Staff are drafting emails with a chatbot. Far fewer have asked what AI could do for the people who pay them.

Is AI a “pivot or die” moment for your business?

For some businesses, yes. A rival who lets customers get a quote, a booking or an answer in two minutes makes a three-day email chain look expensive.

“If you’re an SME and you don’t have a technical founder, you should absolutely be somewhat terrified.”

But AI isn’t a silver bullet either. It isn’t crypto, and it isn’t magic. The truth sits in the middle: real disruption, applied one job at a time. Point the fear at one real problem, not at the whole business.

How many UK businesses use AI in 2026?

A lot more than two years ago, but mostly in shallow ways. Three surveys are worth knowing, and they measure different things.

BICS is the ONS Business Insights and Conditions Survey, and its latest wave ran from 21 September to 4 October 2026.

The ONS figure is the official one. It breaks down by size like this:

• Fewer than 10 employees: 29% use AI.
• 10 or more employees: 38%.
• 250 or more employees: 51%.

The surveys ask different questions of different samples, so the numbers differ. They point the same way. Adoption has roughly tripled since 2023, and the gap between small and large firms remains.

The more useful finding is about depth. Among ONS respondents with 10 or more staff, the most common use was text generation with large language models, at 21%. An ONS analysis from July 2026 found adopters used an average of 1.6 AI technologies, barely up from 1.4 in 2023. The BCC found only about one in ten SMEs had moved on to bespoke AI systems.

So the typical UK firm has a chatbot licence and nothing more. The DSIT research is older, but its finding still holds. Most firms not using AI simply hadn’t found a job worth giving it. The rest of this guide is about finding that job, on both sides of the business.

Where does AI pay? Two halves

Start with whichever half has the bigger number attached. For many SMEs that is now the customer side, because staff tools are cheap to buy off the shelf.

For customers: save them time and finish the job

Customers don’t want to talk to your AI. They want the thing done. The test for every customer-facing idea is simple: does it get them to the end of their task faster than today?

Agents that complete a flow end to end

An agent is a language model that can take actions, not just answer questions. It can look up an order, check a calendar, fill in a form or call your own systems. The useful ones finish a flow, rather than handing over at the first awkward step.

Good candidates are flows your team currently finishes by hand:

• Booking and rescheduling. The customer says what they need and when. The agent checks availability, books it and confirms.
• Quoting. The agent asks the questions your sales team always asks, then produces a quote within limits you set.
• Onboarding. It collects documents, checks they’re complete and chases what’s missing, before a person ever looks.
• Order and case status. “Where’s my delivery?” answered from your live data, not a template.

The rule is permissions. Give an agent exactly the actions it needs, log every one, and set a limit above which a person approves. A refund agent that can issue £20 credits is useful. One that can issue any amount is a liability.

Automation the customer feels

Not every customer win needs a conversation. Sometimes AI simply removes a step.

On a customer call with an HR-tech client, we learned that one of their customers paid a full-time salary to translate Spanish reports. Their own sales team had said translation was too complex to add. Machine translation is now good enough for most reports, so we shipped an auto-translate button within 24 hours. It went out behind a feature flag, a switch that shows it to a few users first, then rolled out more widely.

That is the shape to look for. Find the work your customers do around your product, and do it for them.

New ways for customers to explore

The second customer opportunity is the interface itself. Menus and filters assume the customer knows what they’re looking for. Many don’t.

• Conversational search. The customer describes the problem in their own words. The system finds the right products, articles or answers, and shows where each came from.
• Guided configuration. Instead of a 40-field form, the customer answers a few plain questions. The system builds a configuration, quote or plan, and explains each choice.
• Comparison and explanation. “What’s the difference between these two plans for a team of 12?” answered from your own pricing and terms.

Under the hood this usually uses retrieval-augmented generation (RAG). The system finds the relevant passages in your content first, then writes an answer from them. Our guide to getting your data ready for AI explains it properly.

AI can also see and hear. Pristine Condition trains warehouse and aviation staff to lift safely. After we rebuilt their old software as a mobile app, usage grew by around 9,000%. The AI came next: it assesses a person’s lifting technique through an ordinary phone camera, in real time, with no internet connection. The user gets feedback in the moment, and the employer gets evidence rather than a self-reported yes.

Running costs matter more here than for staff tools, because every customer conversation costs money. We’re publishing a separate guide to what an AI chatbot costs to run.

For staff: save time and raise quality

Augmentation means the person still does the job, with AI doing the slow parts. It suits work where judgement matters but most of the effort is reading, writing or sorting.

Saving time

The clearest wins are repetitive text and document work:

• Extraction. Reading forms, invoices, contracts or CVs and pulling the same fields out of each one.
• Drafting. First versions of replies, reports or proposals, for someone to check and send.
• Triage. Sorting what comes in: complaint, sales lead or invoice. Which support ticket needs a human today?
• Redaction. Removing personal details before a document is shared.

For Everyone’s Invited, a safeguarding charity, we built a system that strips names, dates and locations from survivors’ submissions automatically, before review. It handled the review of more than 50,000 submissions and saved the team around 12,500 hours of manual anonymisation.

Checking quality

The second staff use gets less attention. AI is a good second reader. It doesn’t replace the expert, but it catches what a tired person misses.

• Checking work against a standard. Does this proposal cover every point in the tender? Does this contract match our usual terms?
• Consistency. Is the same question answered the same way by every person on the team?
• Spotting gaps. Which records are missing a field that matters later?

Quality checks are safer than full automation, because a person still decides. They’re a good first project for regulated work.

Start staff tools with a licence

Most staff augmentation doesn’t need a build. Buy a business plan of ChatGPT, Claude or Microsoft 365 Copilot, not a personal account, and check its data terms. Use it yourself for a month, then give it to the people closest to the problem.

Watch what they keep coming back to. Build only when a task needs your own data, your own systems or checks a general tool can’t do.

Is plain automation enough?

Sometimes there’s no model involved at all. At The Home Help Company, the team spent too long on schedules, payroll and client messages by hand. We moved them from WhatsApp to Slack, digitised scheduling, and automated billing and PAYE.

The CEO got back 30 hours a month on payroll and invoicing. The business hired 20 new staff and took on 200 more clients. If a plain automation does the job, we’ll say so.

Where doesn’t AI pay?

This is the part most AI consultancies skip.

• A chatbot because competitors have one. If it can only answer questions and not finish anything, it adds a step.
• Agents with no limits. An agent that can refund, cancel or change prices without a ceiling will eventually do it wrongly.
• Processes nobody can describe. If three people do the job three ways, AI will automate the confusion. Fix the process first.
• Data that lives in someone’s head. AI can’t search knowledge that was never written down.
• Low-volume work. A task that takes two hours a month isn’t worth a build. A licence will do.

If the output touches personal data, read our AI and GDPR guide before you start.

How do you pick a first AI project?

Score each candidate against five questions. Our AI readiness scorecard does this for you in a few minutes.

Write the problem down in one sentence, with a number in it. “Customers wait two days for a quote” is a brief. “We should use AI” isn’t.

“I can’t emphasise enough that you have to start with a problem that you’re trying to solve.”

Don’t aim for perfect. The target is better than today, measured against that number.

What does AI cost for an SME?

There are three shapes, and they cost very different amounts.

• A licence. ChatGPT, Claude or Copilot, billed per person, per month. Good for staff augmentation. No build.
• A configured assistant. The same models set up over your own content, with your instructions and permissions. A short project, then platform fees.
• A custom build. An agent or interface connected to your own systems, with limits, logging and someone looking after it. Then hosting and model costs.

We compare the three, with current vendor prices, in off-the-shelf AI vs custom.

If AI is part of a new product, it fits our fixed-price LaunchPad. It costs £25k for two months, £60k for three or £100k for four, excluding VAT. Discovery sets the scope within the tier you pick, and you get your money back if the plan isn’t right. You own the code and IP.

For AI added to an existing product or operation, we price the build after we’ve seen your process and data. Anyone quoting a custom AI build before that is guessing.

How do you get started?

1. Have a free chat. We talk through the jobs you think are candidates, on either side. Sometimes we tell you a licence will do.
2. Count the problem. Volume, time and drop-off, plus where the data lives and what personal data is in it.
3. Try real examples. A sample of real cases shows whether the idea holds before anyone commits to a build.
4. Agree a plan and a price. What to build, what not to, the risks and the data protection steps.

If you already have something half-built, we also offer a free architectural review.

Who should you work with?

Look for three things in an AI partner. They should ask about your customers and process before your technology. They should raise data protection without being prompted. And they should show you something they built that’s running today.

Our founder was founding engineer at AI Music, which Apple acquired. We’re partners with Databricks and AWS, and we’re listed in The Manifest. We use AI in our own work, and senior engineers review and own everything that ships. Our AI + data service explains how we work.

If you’ve got a job in mind, for your customers or your team, talk to us. The first chat is free, and we’ll tell you honestly if it’s worth doing.