Service 02
An assistant that knows your business.
A general model knows the internet and nothing about you. We ground assistants in your documentation, your policies and your product, so the answers are specific, cited and safe to put in front of a customer.
- Timeline
- 3–6 weeks
- Suited to
- Businesses answering the same questions repeatedly
Typically starts at
Priced in your currency and fixed in writing before work starts. Change the region in the navigation to see local pricing.
Start a ProjectSee what each package includes01The problem
Why most chatbots get switched off
They are bolted on, fed a scrape of the marketing site, and shipped without anyone checking what they say under pressure. Within a month support is fielding complaints about the bot on top of the original questions.
It invents answers
Asked something outside what it knows, a general model guesses fluently. In a support context a confident wrong answer is worse than no answer.
It cannot cite anything
Without a source the visitor cannot verify it and your team cannot correct it — so nobody trusts it, including you.
It traps people
No route to a human turns a small question into a lost customer and a bad review.
02Our approach
How we approach it
Retrieval first, generation second. The assistant answers from passages it actually found in your content, shows where each came from, and declines when the content does not cover the question.
Grounded in your content
Indexed from your docs, policies and product content, with a refresh path so the answers do not quietly go stale.
Citations on every answer
Each response links the source passage, so a visitor can check it and your team can fix the underlying page.
Allowed to say no
Tuned to decline rather than improvise. Out-of-scope questions get an honest answer and a route onward.
Handoff that keeps context
Escalation carries the full transcript to your inbox or helpdesk, so nobody has to repeat themselves.
Your documentation, policies and product content are indexed on a schedule. A visitor question triggers retrieval of the passages most likely to answer it. The assistant generates an answer only from those passages and cites each one. If the retrieved content does not cover the question, it declines rather than improvising, and offers escalation — which carries the full transcript to your inbox or helpdesk. Questions it could not answer are reported back as content gaps.
03Deliverables
Exactly what you receive.
No vague line items. This is the list that goes into the agreement.
Knowledge
- Content audit and source inventory
- Indexing pipeline with a scheduled refresh
- Retrieval tuning against real questions
- Coverage report showing what it cannot answer
Assistant
- Answer generation with inline citations
- Scope, tone and refusal behaviour
- Escalation to email, helpdesk or live chat
- Accessible interface, keyboard and screen-reader tested
- Conversation logging and quality review
Handover
- Analytics on questions, gaps and deflection
- Content recommendations from unanswered questions
- Editorial guide for keeping answers current
- Team walkthrough
04How it runs
Five stages, each with a clear exit.
- 01
Question mining
We start from what people actually ask — support tickets, search logs, sales calls — not from what the site happens to document.
- 02
Grounding
Content indexed and retrieval tuned until the right passage comes back for the real questions, measured before any answer is generated.
- 03
Behaviour
Scope, tone, refusals and escalation, tested against the awkward questions as well as the easy ones.
- 04
Interface
Designed into the site rather than dropped on top of it, and usable by keyboard and screen reader.
- 05
Handover
Launched with analytics, plus the loop that turns unanswered questions into better content.
05Related work
How this looks in practice.
06Questions
AI Assistants & Chatbots, answered.
Bundled widgets answer from a generic index and cannot act. This one is grounded in your content with citations, tuned against your real questions, and built to hand off cleanly. If a bundled widget genuinely covers your case, we will tell you — it is cheaper than hiring us.
It says so, and offers the route onward. We treat a wrong confident answer as a defect, and decline behaviour is part of the test suite rather than an afterthought.
Yes — checking an order, booking a slot, raising a ticket. Once it takes actions rather than answering questions it is an agent, so we scope it that way and build the tools and boundaries to match.
The index refreshes on a schedule from your existing content, so updating the source page updates the assistant. No separate bot knowledge base to maintain in parallel.
AI Assistants & Chatbots
Let us look at your project.
Send the brief — or just the problem. We will tell you what we would do and what it would cost.