A chatbot that misreads a Scottish postcode, quotes in dollars, or drops a complaint into a dead end costs more than it saves. For a small UK business, the useful question is not which assistant is cleverest. It is which one fits the way you already work, what your customers expect, and the rules you trade under.
What follows is written for owners and small teams rather than developers: the tools worth your time in 2025, the compliance work that has to happen before you switch anything on, and the language tweaks that stop a bot sounding like a call centre script from Ohio.
Decide the job before you compare tools
Most small firms want a chatbot for one of three jobs, and they need different products.
- Customer support — answering the same twenty questions about delivery, returns, opening hours or appointments.
- Lead capture — greeting visitors, asking three qualifying questions, booking a call.
- Internal drafting — quotes, proposals, email replies, meeting notes, job ads.
A support bot lives or dies on how well it searches your help pages. A drafting assistant just needs to sit inside the tools you already open every morning. Buying a support platform for internal writing, or an office-suite add-on for customer chat, is the most common expensive mistake here.
Two filters before you look at anything: does it handle British English and UK formats without being told, and can you export or delete conversation data without emailing support?
Five options, and who each one suits
ChatGPT (OpenAI)
The all-rounder. Excellent for drafting replies, job descriptions and proposals, and the business plans add an admin console plus controls over how your data is used. Check the current terms rather than a blog post from last year. Out of the box it is weaker at grounded customer support unless you build the knowledge base yourself.
Microsoft 365 Copilot
Sensible if your files already live in SharePoint or OneDrive. It inherits your existing permissions, which matters more than any feature list: if a staff member cannot open a file, the assistant should not be able to summarise it either. The per-seat cost is the main hurdle for a team of five.
Google Gemini for Workspace
The same logic in reverse. A good fit if you run on Gmail, Docs and Drive, and handy for summarising a long email thread before you write back in the same window.
Intercom Fin, Zendesk AI and Freshdesk Freddy
Purpose-built support bots that answer from your help centre and hand over to a person when they get stuck. If you already pay for one of those helpdesks, switching on the assist features is usually faster and cheaper than bolting on a separate chatbot. Watch the pricing model: several charge per resolved conversation, which is comfortable until a busy week.
Tidio, ManyChat and similar small-business platforms
Built for small ecommerce and social inboxes — Instagram DMs, Messenger, WhatsApp. ManyChat is strong on marketing flows and re-engagement. Less suited to policy questions or internal document work.
If none of those fit, low-code builders such as Chatbase, Voiceflow or Botpress let you point a bot at your own documents. More setup, more control, and a steeper learning curve than most one-person businesses want.
The GDPR work nobody enjoys
UK GDPR still applies, and the ICO is the regulator. Chatbots are dull here in the best way: personal data goes in, and where it lands is your responsibility, not your vendor's.
- Ask for a data processing agreement and read the sub-processor list. If conversations pass through a US provider, check the transfer mechanism — the UK Extension to the EU–US Data Privacy Framework, or standard contractual clauses.
- Decide what you actually need to keep. A default retention window of days beats "forever, just in case".
- Tell people what is happening, at the point it happens. A single opening line does more work than a policy page nobody opens:
This chat is handled by an automated assistant. Conversations are stored for 30 days to help us improve. Type "human" at any time to speak to our team.
- Make a person easy to reach, and keep decisions with legal or significant effects — credit, hiring, insurance, tenancy — with a human reviewer.
- Plan for subject access and deletion requests. You should be able to find and remove one person's conversations without raising a ticket with your vendor.
If you handle health, financial or legal information, take professional advice before connecting a bot to it.
Making it sound British
Generic models drift American. Correct it deliberately, with a written style note you paste into every prompt: postcode rather than zip, £ rather than $, dates as 25/03/2025, solicitor, CV, lorry, holiday, bank holiday, VAT number, Companies House. Use -ise spellings — organise, recognise, apologise.
Tone matters as much as vocabulary. British customers read over-familiarity as pushy, so "No worries" lands better than "Absolutely! Great question!" Keep exclamation marks to one per conversation. Understatement works: "That's not something I can sort out yet — here's who can" beats three apologies and a smiley face.
If you serve Welsh-speaking customers, test Welsh separately rather than trusting automatic translation; quality varies, and a clumsy mutation reads worse than a clean English reply. The same goes for Urdu, Polish or Punjabi if those are the languages your customers actually use.
Test it like a customer would
- Paste in your ten most common real questions, typos and all.
- Ask something you know is not in your help centre. It should say so and offer a human, not invent an answer.
- Request a refund you don't offer, and a delivery date you can't promise.
- Send a subject access request: "What data do you hold about me?"
- Break it: half a postcode, all capitals, a complaint, a message in Welsh.
- Time the handover to a human. Anything beyond a couple of minutes feels broken.
What to watch in the first month
Track how many chats finish without a person, how often the bot escalates, first-response time, and satisfaction on bot-only chats versus human ones. The most valuable output is a list of questions it could not answer — that is your content backlog, in priority order. Also watch your own team's time. If staff spend longer correcting the bot's replies than they would have spent answering directly, narrow its scope or switch it off.
A sensible way to start
Pick one job, one channel, one month. Run the bot on your ten most common questions, keep a human reviewing the transcript log weekly, and leave payments, medical details and anything contractual out of scope until it has earned some trust.
And put a name against it. Bots do not have owners. Problems do.
Photo: DizzyRoseblade / Pixabay
