Chatbots have a credibility problem. The 2018 rules-based ones were dreadful, and a lot of people still picture those when they hear "AI chatbot". The modern version (LLM-powered, grounded in your own content) is different — but it's not magic, and how you deploy it makes more difference than which platform you pick.
What modern AI chatbots are good at
Three jobs they do genuinely well:
- Answering questions that are answered somewhere on your website. Pricing, hours, scope of services, returns policy, qualifications. The chatbot reads your published content and answers in conversational form.
- Qualifying leads. Walking a visitor through a structured intake — postcode, budget, timeline, sector — and producing a clean lead record by the end. This is where most of our chatbot deployments earn their keep.
- Capturing context for humans. When the chatbot can't answer, it collects what the customer was asking, captures their email, and hands the conversation off cleanly. The human then has a transcript to start from rather than a cold "how can I help?".
What they're bad at
Three jobs to never trust them with:
- Anything regulated. Medical advice, legal advice, financial advice, tax positions. Even if the model knows the right answer, the liability and compliance posture is wrong. Don't.
- Hard numbers and bookings under load. Don't have the chatbot promise availability or quote prices unless it's reading a live source of truth. Otherwise it will improvise and you'll have to honour what it said.
- Open-ended product recommendations on complex items. "Which workwear order should I place for my crew of 12?" is a conversation, not a chatbot prompt. Build the right tooling for that instead.
How to deploy one without embarrassing yourself
Three rules:
Constrain the scope explicitly.
The chatbot's system prompt should say "answer questions about [services], hours, prices, locations and how to get started. For anything else, take a name and email and tell them we'll be in touch." That sentence alone prevents 90% of embarrassing chatbot moments.
Ground every answer in your published content.
Don't let the model improvise from its general training. Modern setups use retrieval against your own pages so the chatbot quotes you, not Wikipedia.
Make the human handover obvious.
Customers should be able to ask for a human and get one (or get a clear "we'll reply by X time" if it's out of hours). The worst chatbot UX is the one that won't let you escape.
Sector notes
For accountancy firms, the chatbot's job is qualifying prospects — turnover band, sector, services needed — not giving tax answers. For tradespeople, it's capturing job specifics (postcode, scope, urgency) before someone calls back. For healthcare practices, it's strictly information about the practice, never clinical advice; even simple symptom triage is off-limits unless you have specific clinical sign-off.
Cost and timeline
A useful, constrained AI chatbot grounded in your own content is typically £1,500–£3,000 to deploy, with a small monthly running cost for the language model itself. It can be live in 5–10 working days assuming your website content is already structured. Bespoke versions that integrate with a CRM or booking system are scoped separately.
A note from the Bloomorbit studio
We deploy AI chatbots for UK small businesses with the constraints set sensibly from day one. Cardiff-based, working UK-wide. If you'd like an honest opinion on whether one is right for your business — or whether your current chatbot is making you look worse — we'll give it.