An AI support agent that actually knows your product,
and says so when it does not
Most support bots fail the same way: they guess an answer that sounds confident and is wrong. We build agents that answer only from a knowledge base your team can edit directly, and that hand off the moment a question goes outside it.
What it is and who needs it
An AI support agent answers customer questions directly from your own documentation, and says it does not know rather than inventing a confident wrong answer. It is for any business fielding the same questions repeatedly through chat, booking details, policy questions, status checks, where a person is currently typing the same answer by hand dozens of times a day. It is not meant to replace a support team handling genuinely hard cases; it is meant to absorb the repetitive ones so the team can focus on what actually needs a human.
What is inside
The knowledge base is the core of the build, structured so your team can open an admin panel and edit an answer directly, no code, no waiting on a developer. The agent runs on Claude or GPT with instructions to answer only from that knowledge base and nothing else, so it cannot wander into guesses about your refund policy or your opening hours. A clear hand-off rule routes anything outside the knowledge base straight to a human, with the full conversation attached so nobody has to ask the customer to repeat themselves. Every channel you run gets wired into the same brain and the same knowledge base, so an answer updated once is correct everywhere.
How we build it
We start by reading whatever documentation already exists, policy pages, FAQ documents, previous support transcripts, and structure it into a knowledge base the agent can answer from reliably. Where documentation is thin, we write it with your team before the agent goes live, since a support agent is only as good as what it is allowed to say. The first channel launches with hand-off rules tested against real questions pulled from your support history. Once it is stable, we connect the remaining channels and build the admin panel for ongoing edits. The first month runs with weekly review of unanswered questions, feeding directly back into the knowledge base.
What to watch
The main failure mode is a knowledge base that falls out of date while the agent keeps answering from it confidently. Routing every unanswered question to a visible log, and actually reviewing that log, is what catches this before a customer gets a wrong policy answer. A second risk is scope creep: letting the agent attempt an increasing range of questions without the knowledge base keeping pace invites exactly the confident-wrong-answer problem this whole approach is built to avoid. Keep the hand-off rule strict even after the agent seems reliable; loosening it is how trust erodes slowly and then all at once.
Timeline and price
| Option | Price | What it covers | Timeline |
|---|---|---|---|
| MVP | from $2,000 | One channel, knowledge base from existing docs, hand-off to a human | 3 to 4 weeks |
| Production | from $5,000 | Multi-channel, editable admin panel, alert routing, dialogue dashboard | 5 to 7 weeks |
| Full control (handover-ready) | from $8,500 | Everything in Production, plus a full handover package: architecture docs, test suite, admin access audit, and a walkthrough so your own team or another vendor can run it without us | 7 to 8 weeks |
Running cost on top of the build is usually $15 to $60 a month in model calls, depending on question volume.
What you own at the end
You own the knowledge base, the dialogue history, the admin panel and the full source code, deployed on infrastructure in your name. Editing an answer never requires us or any developer once handover is complete, and a written document explains the whole system for anyone who takes it over later.
Related
Pairs with the AI knowledge base assistant when the same knowledge base also needs to serve your internal team, and the AI sales agent product for the conversations that should end in a sale instead of an answer. For the underlying retrieval architecture, see RAG search product. See the AI agents service page for the full range of agent builds we run. Real build: the visa consulting centre AI support bots case study and the staffing agency recruitment bot case study. Fielding the same five questions all day? Get in touch and send us what your team currently answers by hand.
FAQ
How much does an AI support agent cost?
From $2,000 for a single-channel agent with a knowledge base built from your existing documentation. A multi-channel build with an editable admin panel and alert routing runs $5,000 to $8,000.
How long does it take?
Three to four weeks to launch the first channel once your documentation is organized. If the documentation does not exist yet, writing it with your team adds one to two weeks before the agent goes live.
What is the stack?
Python and FastAPI, Claude or GPT for the model, PostgreSQL for the knowledge base and dialogue history, and the native API of each channel. The admin panel for editing answers is usually a small Next.js or Flask interface.
Who owns the agent and the data?
You. The knowledge base, the dialogue history and the code are yours, hosted on your own server or VPS. We hand over admin access on day one, not after a separate request.
What does it do when it cannot answer?
It says so and hands off to a human, rather than inventing a plausible-sounding answer. Every unanswered question is logged so your team can decide whether to add it to the knowledge base.