AI chatbot for your website: is it worth it for a UK SME?
By Kieran Mullings, BEng Mechanical Engineering (UWE Bristol), technical co-founder of FlowGenix Automations Ltd.
An AI chatbot is worth adding to a UK SME website when visitors repeat the same questions and your pages already hold the answers. The bot also needs a handoff to a person or a booking form. A chatbot that guesses beyond your content creates legal and reputational risk, as Air Canada and DPD both learned in 2024.
FlowGenix Automations is an audit-led AI and automation consultancy for UK SMEs.
Key takeaways
- In Moffatt v Air Canada (2024), a Canadian tribunal held the airline liable for its chatbot's wrong advice and rejected the claim that the bot was a separate entity.
- DPD disabled the AI element of its UK customer chatbot in January 2024 after a system update let a customer make it swear and criticise the firm.
- Intercom prices its Fin AI agent at $0.99 per resolution, while a custom chatbot grounded on your own pages is billed mainly by model tokens.
- A website chatbot earns its keep by capturing and qualifying enquiries out of hours, not by replacing your contact form.
What does an AI website chatbot do?
An AI website chatbot answers visitors' questions in plain language using a large language model, then points them to the right page, product or booking form. A well-built chatbot answers only from your own approved content. A poorly built one answers from the model's general knowledge, which is where wrong prices and invented policies come from.
The technique that keeps answers grounded is retrieval. Your pages, FAQs and price lists are split into passages and stored in a vector database such as Pinecone. When a visitor asks a question, the chatbot retrieves the closest passages and the model answers from those alone. Every answer can cite the page it came from.
Every FlowGenix website build includes this pattern: a chatbot that answers questions and points visitors to the right page or product. Booking forms confirm instantly by email, and an SMS reaches the owner's phone. The chatbot is one part of a website built to capture and confirm enquiries, not a bolt-on widget. The full list of FlowGenix services shows how websites connect to the rest of the stack.
Jobs a website chatbot handles well:
- Opening hours, service areas, delivery times and prices from your price list
- Matching a visitor's problem to the right service page
- Collecting name, contact details and job details for a quote
- Booking a call or appointment through a live calendar
- Handing a conversation to a person when the question falls outside scope
Key facts
- Retrieval grounds chatbot answers in your own pages rather than general knowledge.
- Pinecone stores website passages for fast similarity search.
- FlowGenix website builds include a chatbot that routes visitors to the right page.
What are the risks of putting an AI chatbot on your website?
The main risks are wrong answers that customers rely on, embarrassing outputs that spread on social media, and personal data handled without proper notice. Businesses cannot disown what their chatbot says. A Canadian tribunal held Air Canada liable in February 2024 after its chatbot gave a grieving customer wrong refund advice.
In Moffatt v Air Canada, the British Columbia Civil Resolution Tribunal rejected Air Canada's argument that its chatbot was a separate legal entity. The tribunal described the chatbot as part of Air Canada's website and ordered the airline to pay C$650.88 in damages, plus interest and fees. The ruling is Canadian and not binding in the UK, but the logic is simple: a chatbot speaks for the business.
DPD supplied the UK cautionary tale. In January 2024, a system update caused its customer service chatbot to swear at a customer and write a poem criticising the company. DPD disabled the AI element. FlowGenix tests every chatbot with deliberate attempts to push it off-script before launch, because DPD's failure came from an update rather than the original build.
Key facts
- The British Columbia Civil Resolution Tribunal held Air Canada liable for its chatbot's advice.
- DPD disabled its chatbot's AI element after a January 2024 system update.
- UK GDPR applies to personal data collected through a website chatbot.
How much does a website chatbot cost?
A website chatbot costs anywhere from a free tier on a basic widget to several hundred pounds a month, depending on the pricing model. SaaS AI agents often charge per resolved conversation, such as Intercom's Fin at $0.99 per resolution. A custom chatbot built on your own content is billed mainly by language model tokens.
Per-resolution pricing is simple to understand but scales with success. A site resolving 300 conversations a month with Fin pays about $297 in resolution fees before any seat costs. Per-resolution pricing suits a business where each conversation is worth far more. For a small business with low-value queries, the bill can outgrow the benefit.
Token-based pricing scales with conversation length instead. Estimate it before building. Multiply monthly conversations by messages per conversation and by tokens per message, including the retrieved passages. For 300 conversations of six messages at about 3,000 tokens each, the total is 5.4 million tokens a month. Multiply that by the model's published per-million-token rate.
FlowGenix builds on the Anthropic API and sizes the model to the job, because a small model handles most FAQ traffic at a fraction of a large model's price.
| Option | Pricing model | Answers from your content | Hands off to a person | Typical fit |
|---|---|---|---|---|
| Rule-based widget | Free to low monthly fee | Only scripted answers | Usually | Very simple FAQs |
| SaaS AI agent, such as Intercom Fin | Per resolution plus seats | Yes, from your help centre | Yes | Busy support teams |
| Custom chatbot on your own pages | Model tokens plus hosting | Yes, retrieval-grounded | Yes, by design | SMEs wanting control and low running cost |
| Live chat with staff | Staff time | Yes | Already a person | Only during staffed hours |
Key facts
- Intercom charges $0.99 for each conversation its Fin AI agent resolves.
- Custom chatbot running cost scales with tokens per conversation.
- FlowGenix sizes the language model to the chatbot's actual workload.
How do you stop a chatbot making things up?
Stop a chatbot inventing answers by restricting it to retrieved passages from your own content. Instruct it to say "I don't know" outside that content, and route anything uncertain to a person. Then test it with hostile questions before launch and after every update. Grounding plus testing prevents most of the failures that make the news.
Grounding alone is not enough, because the model can still blend retrieved facts with general knowledge. A strict system prompt forbids answers without a supporting passage. Every answer should link to the page it came from, so the visitor can check it. Prices and policies should come from one source page that you keep current.
Testing is where most builds cut corners. DPD's chatbot failed after an update, not on launch day. FlowGenix keeps a test script of off-topic, rude and manipulative questions and reruns it whenever the model, prompt or content changes. Logs get reviewed weekly in the first month.
How to build a chatbot that stays on script:
- Collect the pages, FAQs and price lists the chatbot is allowed to use.
- Split the content into passages and store them in a vector database.
- Instruct the model to answer only from retrieved passages and cite the source page.
- Define a refusal message and a handoff route for out-of-scope questions.
- Test with off-topic, rude and manipulative prompts before launch.
- Review conversation logs weekly for the first month, then monthly.
- Rerun the full test script after any model, prompt or content change.
Key facts
- Retrieval-grounded chatbots answer from approved passages only.
- DPD's chatbot failure followed a system update, not the original launch.
- FlowGenix reruns a hostile-prompt test script after every chatbot change.
Does a chatbot help your SEO or AI search visibility?
A chatbot does not directly improve your Google rankings or your chances of being cited by ChatGPT, because search crawlers rarely read conversations inside a chat widget. The content you write to feed the chatbot does help, once you publish it as crawlable pages. Clear FAQs and service pages serve both the bot and search engines.
Most chatbot widgets load conversations through JavaScript after the page renders. Googlebot and AI crawlers such as GPTBot and ClaudeBot index the page, not the private conversations. Hiding your best answers inside a chatbot therefore removes them from search. Publish the answers as pages first, then let the chatbot retrieve from those pages. The guide on how to get your business mentioned in AI answers explains what answer engines do read.
The questions visitors ask the chatbot are useful research data. The logs show what customers actually want to know, in their own words. FlowGenix turns recurring chatbot questions into new FAQ entries and service page sections, which feeds the search recovery work on the same site.
Key facts
- Search crawlers index page content, not private chatbot conversations.
- Chatbot logs reveal the exact questions customers type.
- FlowGenix turns recurring chatbot questions into crawlable FAQ content.
Chatbot, contact form or live chat: which should a small business use?
Most small businesses should keep their contact form and add a chatbot that answers routine questions and collects enquiry details out of hours. Live chat with staff suits only businesses that can staff it reliably. A chatbot with a human handoff covers the gap between those options at the lowest running cost.
Book a free 30-minute audit to see where your website loses enquiries and whether a chatbot would recover them.
Response speed is the deciding factor. A contact form submitted at 9pm waits until morning. A chatbot can answer the question, qualify the job and book a slot immediately. FlowGenix website builds pair the chatbot with an instant confirmation email and an SMS alert to the owner, so the enquiry stays warm.
| Factor | Contact form | Live chat with staff | AI chatbot with handoff |
|---|---|---|---|
| Available out of hours | Yes, reply waits | No | Yes |
| Answers questions instantly | No | Yes, when staffed | Yes, from your content |
| Qualifies the enquiry | Only with set fields | Yes | Yes |
| Running cost | Near zero | Staff time | Tokens or per-resolution fee |
| Risk of wrong answers | None | Low | Managed by grounding and testing |
Key facts
- A contact form submitted at night waits until the next working day.
- Live chat only works during staffed hours.
- A chatbot with handoff answers and qualifies enquiries at any hour.
Frequently asked questions
Can a website chatbot book appointments?
Yes, if your calendar or booking system has an API. The chatbot checks live availability, offers slots and confirms the booking during the conversation. Google Calendar, Microsoft Outlook and most booking platforms support this. A confirmation email and an SMS alert to the business owner should fire the moment the booking lands, so nothing waits unread.
Do I need to tell visitors they are talking to a chatbot?
Yes. Label the chatbot clearly as an automated assistant and link to your privacy notice from the chat window. UK GDPR requires you to explain how you use personal data collected in the conversation. Clear labelling also manages expectations, so visitors know when to ask for a person instead of arguing with the bot.
Will an AI chatbot work on a Lovable, Framer or WordPress site?
Yes. Most chatbots load through a small script tag, which works on Lovable, Framer, WordPress and most other site builders. The chatbot's knowledge base lives outside the website, so the platform choice does not limit what it can answer. Check that the script does not slow page loading, because speed affects both visitors and rankings.
What happens to the data visitors type into a chatbot?
The conversation passes to the chatbot provider and the language model provider, then usually into your CRM or inbox. Check each provider's data retention terms and whether they train on your data. Business API terms from Anthropic and OpenAI do not use customer data for training by default. Record the processors in your privacy notice.
How long does it take to set up an AI chatbot?
A chatbot grounded on an existing website with good FAQ pages can go live within a few weeks, including testing. Sites with thin content take longer, because the content has to be written first. The writing is never wasted, since the same pages help search visibility and answer engine citations as well.
Sources
- British Columbia Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149
- OECD AI Incidents Monitor, DPD AI chatbot disabled after swearing and criticism
- Fin by Intercom, Pricing
- Information Commissioner's Office, Guidance on AI and data protection
Book a free 30-minute audit and we will show you which questions your website should be answering on its own.
About the author
Kieran Mullings is technical co-founder of FlowGenix Automations Ltd (company number 16555743). He builds the voice agents, n8n workflows and AI search foundations behind FlowGenix client systems. He holds a BEng in Mechanical Engineering from UWE Bristol and leads the technical side of every FlowGenix audit. More at About FlowGenix.
Find out what your website should answer on its own.
Thirty minutes on a call. We look at where your website loses enquiries and whether a chatbot would recover them. You keep the roadmap either way.