Grounded answers
Retrieval over your docs, help centre, PDFs and product pages — with citations and an honest "I don't have that" instead of a plausible invention.
Service 03
AI Chatbots on Websites
An assistant on your site that answers from your own documentation, in your visitors' languages, and tells you — in numbers — whether it saved your support team time or just moved the queue.
A website chatbot is the most visible AI project a company can ship, and also the easiest to get wrong. The usual outcome is a widget that gives confident wrong answers and quietly damages trust.
I build them the other way around: grounded strictly in content you control, visibly uncertain when it does not know, human handoff in one click, and an analytics layer that ties conversations to support-ticket volume and conversion.
Deliverables
Multilingual by design
English, Czech and Slovak from day one. Same knowledge base, language-appropriate answers, and per-language analytics so you can see where visitors need more content.
Lead capture & qualification
Structured qualification flows, routing to the right inbox or CRM, and clean data out the other end — not a wall of pasted conversation text.
Human handoff
Detects frustration and hard questions, transfers the whole transcript, and never makes a visitor repeat themselves to a person.
Tone & guardrails
Brand voice, refusal behaviour for out-of-scope questions, and hard rules about what the assistant will never promise.
Analytics & instrumentation
Resolution rate, deflection rate, unanswered-question reports, conversion attribution and a weekly list of content gaps your support team can act on.
How I build one
Pick the questions
We start from your real support tickets, not from what you assume visitors ask. The first release usually covers 30–50 questions, and it answers those extremely well.
Prepare the knowledge
Your content, cleaned and chunked. This step usually finds documentation problems first — and the chatbot project fixes your help centre as a side effect.
Ship the narrow version
One widget, one language at first, visible answers with sources, feedback buttons, and a hard-coded fallback to email or phone.
Instrument and tune
Log every unanswered question. Feed them back into content. Watch deflection against ticket volume, not just conversation counts.
Widen carefully
Add languages, add actions like booking a call, then evaluate the agentic paths — only where the conversation is genuinely transactional.
Outcome
- Support volume from informational questions drops measurably, and you can prove which ones.
- Visitors in Czech or Slovak get a real answer instead of an English-only widget.
- Leads arrive qualified, structured and already summarised for your sales team.
- A live list of the questions your content does not answer yet.
- No invented answers about pricing, policy or legal terms — those route to a human.
Typical stack
TypeScript · Node.js · React widget · pgvector / Qdrant · PostgreSQL · Redis streaming · Anthropic / OpenAI · Docker · privacy-first analytics
Frequently asked questions
Why is this not just the ChatGPT widget?
Because it cannot know your pricing, your warranty terms or your product's limitations, and it will not be accountable for what it tells a customer. Grounding, refusal behaviour, human handoff and your own analytics are the entire value.
Which languages do you support?
English, Czech and Slovak natively, with a multilingual knowledge base so answers are not machine-translated on the fly. Other languages are possible, but I would rather tell you where quality degrades.
Can it take actions, not just answer?
Yes, with limits. Booking a call, creating a ticket, checking order status or filling a form are all reasonable. Anything financial or legally binding routes to a person.
What about GDPR?
No cookies for tracking by default, consent before personal data is stored, IP addresses not logged in the chat transcript, and processors listed in your privacy policy. I will also tell you which parts of a chatbot you cannot honestly make GDPR-compliant.
How do we know it works?
Baseline support-ticket volume before launch, then a monthly comparison plus resolution rate on the assistant's own conversations. If it is not moving those numbers in a quarter, I will say so.