AI for customer service in small businesses: what actually works
Easy Insight Team ·
AI works in small-business customer service when it handles the repetitive, well-documented half of your inbox — order status, opening hours, returns policy, "have you got my form?" — and hands everything else to a person immediately. It fails when it is asked to cover the judgement half. The deciding factor is your documentation, not the model you pick.
Are small businesses actually doing this?
Some are, and fewer than the noise suggests. Around 35% of UK businesses with 10 or more employees reported using at least one AI technology in June 2026, up from roughly 12% in late 2023, alongside 28% of businesses with fewer than 10 employees (ONS, published July 2026).
The more useful number is the depth one: the average adopting business uses about 1.6 AI technologies, up only slightly from 1.4 since late 2023, and only around 10% report extensive use. Adoption has spread sideways rather than deepened. Most businesses have a tool switched on somewhere; very few have changed how the work actually happens.
What does AI do well in customer service?
Four things, reliably:
Answering questions you have already answered in writing. Where a stable, documented answer exists — delivery times, refund windows, what your service includes — a modern AI agent finds it and phrases it well.
Triage and routing. Reading an enquiry, classifying it and putting it in front of the right person is unglamorous and genuinely valuable, especially with a shared inbox and no rota.
Out-of-hours cover. A reply at 22:40 that resolves a simple question, or takes proper details and sets an expectation, beats silence until Monday.
Drafting, not sending. The lowest-risk deployment is an AI that drafts the reply and a human who presses send. You keep most of the speed gain and lose almost none of the quality.
Where does it fall over?
On the cases that need judgement, discretion or a policy exception — which is precisely where a small business earns its reputation.
The instructive public example is Klarna. In 2024 it credited its OpenAI-built assistant with the work of about 700 agents. By May 2025 it was rehiring human agents, with chief executive Sebastian Siemiatkowski conceding that the focus on efficiency and cost had produced lower quality (Forbes, May 2025). It now runs a hybrid model: AI on the routine, humans on anything emotional, contested or multi-step.
If a business with Klarna's engineering budget could not automate past that line, a twelve-person firm will not either. The lesson is not "don't use AI" — it is that the escalation path is the product. An agent with a fast, no-arguments route to a human is useful; one built to prevent that handover will cost you customers quietly, and you will not see it in the deflection rate.
What does AI customer service cost?
Three pricing models dominate, suiting different volumes. All figures are list prices as of September 2026, exclude VAT, and use the currency the vendor publishes.
| Model | How it is priced | Indicative cost (Sept 2026) | Suits |
|---|---|---|---|
| Bundled with your helpdesk | Included in the plan, capped | Tidio's free plan includes 50 one-off Lyro AI conversations | Testing whether AI can answer your top ten questions |
| Per AI conversation | Volume tiers | Tidio Lyro from $32.50/month for 50 conversations | Low, predictable volumes |
| Per resolution | You pay when it resolves | Intercom Fin at $0.99 per resolution, seats from $29/seat/month | Volumes that swing month to month |
| Build your own agent | Consumption credits | Microsoft Copilot Studio pre-purchase pack at £153.80/month for 25,000 Copilot Credits | Microsoft 365 businesses wanting internal and customer-facing agents |
Two warnings about the arithmetic. Per-resolution pricing looks cheap until volume grows, and "resolution" is defined by the vendor, not by you — read that definition before committing. Credit-based pricing is hard to forecast, because one complex agent turn can consume many credits; set a budget cap on day one.
The bigger cost is rarely the licence. It is the fortnight someone spends writing down answers that live in people's heads. Budget for that, or the tool will underperform and you will blame the tool.
What do you need in place before it works?
One thing, mostly: written answers you trust.
Most failure modes trace back to the same root: the business asked a model to answer from a knowledge base that was thin, contradictory or out of date. Public reporting on Klarna's problems describes exactly this pattern, with confident answers about fees and policies that had changed between versions. An AI agent does not know which of your two conflicting refund policies is the live one. It will pick one, fluently.
So the prerequisite work is: export a month of enquiries, group them into themes, and write a canonical answer for the top ten. If you cannot reach a single agreed answer for one, that is not an AI problem — it is a business decision nobody has made, and automating it broadcasts the confusion faster. It is the same discipline that gets a business out of spreadsheet chaos: the tidying is the project, the tool is the easy part.
What are the UK compliance basics?
If your chatbot processes personal data — and a customer service chatbot almost always does — UK GDPR applies. Three things to settle before launch:
A data protection impact assessment. The ICO's guidance on AI and data protection states that in the vast majority of cases the use of AI involves processing likely to result in a high risk to individuals' rights and freedoms, which triggers the legal requirement to carry out a DPIA.
Transparency. Tell people they are talking to an AI at the point they start, with a link to fuller detail. It is also good manners, and customers notice when it is missing.
Retention. Decide how long conversation logs are kept and where, and make sure the vendor's default matches your privacy notice. Check whether your data trains the vendor's models, and turn that off if you have not told customers about it.
How would you start in a fortnight?
Week one: export the last month of enquiries, group them, and write canonical answers to the top ten. Decide which five an AI may answer alone and which five it always hands over.
Week two: switch on a free or entry tier, load only those five answers, and put it on one channel — usually the website. Set the handover rule so any complaint, refund request or unclear question goes straight to a person, with the conversation history attached.
Then measure two numbers, not one. Deflection rate tells you what the tool saved. The rate at which customers ask for a human straight after an AI reply tells you what it cost. If the second climbs, narrow the scope rather than tuning the prompt.
That is the whole method, deliberately small: the returns come from the documentation work and the escalation design, not the sophistication of the model. For the wider view of where automation pays back across a business, our AI automation service page covers process selection, and the AI practice overview sets out how we scope this kind of work.
Frequently asked questions
Can AI replace our customer service team?
No — not in a small business, and the evidence from larger ones is discouraging. Klarna publicly credited its AI assistant with the work of around 700 agents in 2024, then began rehiring human agents in May 2025 after its chief executive said the cost-driven approach had produced lower quality. Treat AI as the first filter on your inbox, not the whole team.
How much does AI customer service cost for a small business?
As of September 2026, entry-level AI agents start at around $32.50 a month for 50 AI conversations (Tidio's Lyro), while resolution-based pricing such as Intercom's Fin is listed at $0.99 per resolution. Building your own agent on Microsoft Copilot Studio starts at £153.80 a month for a 25,000 Copilot Credit pack. All figures exclude VAT and change often — check the vendor's page before you budget.
Do we need a DPIA for a customer service chatbot?
Often, yes. The ICO's guidance on AI and data protection states that in the vast majority of cases the use of AI involves processing likely to result in a high risk to individuals' rights and freedoms, which triggers the legal requirement to carry out a data protection impact assessment. Settle your lawful basis, your retention period for conversation logs and your transparency notice at the same time.
How many UK businesses are actually using AI?
Around 35% of UK businesses with 10 or more employees reported using at least one AI technology in June 2026, alongside 28% of businesses with fewer than 10 employees, according to the Office for National Statistics (published July 2026). Use is still shallow: the average adopting business runs about 1.6 AI technologies, up from roughly 1.4 in late 2023.
What should we automate first?
The questions you answer most often and can already answer in writing without checking with a colleague. Export a month of enquiries, group them, and take the top five that have a documented, stable answer. If the answer only lives in someone's head, write it down before you automate it.
Thinking about where AI fits your operation, not just your inbox? Our AI strategy work starts with the same question: what do you already know well enough to automate? If the answer turns out to be a bespoke assistant grounded in your own data, that is custom AI development.
Easy Insight is a UK consultancy for AI, web, apps and data — senior specialists only, no juniors.
Next step
Wondering where AI would actually pay in your business?
An AI strategy and readiness review (from £2,500) tells you where AI pays, where it doesn't, and what to leave alone. The price is fixed in writing before we start, and the advice comes from a team that runs AI in its own products.