A RAG AI agent that admits what it does not know
Queries the company knowledge, answers with sources, admits when it lacks enough information and hands off to a person.
AI chatbots that know when to stop talking
A chatbot answers; an agent acts. This page is about the first: an assistant that holds a conversation over your knowledge and escalates cleanly. If what you need is something that reads and writes in your systems, that is AI agent development.
Why most chatbots get switched off
It answers confidently and wrongly
Without grounding in real sources, a model fills gaps with plausible text. That is worse than no chatbot at all.
It traps people
No escalation path, or one that loses the conversation. The customer ends up writing an angry email anyway.
Nobody knows if it improved
Every prompt tweak is a guess because there is no evaluation set to compare against.
The bill surprises everyone
Model usage per conversation was never estimated against real volume.
Scope
What we implement
Grounded knowledge
Answers built from your documentation with the source cited, and an explicit "I do not have that" when retrieval comes back empty.
Tone and boundaries
What it may promise, what it must never state, and how it declines. Written down and tested, not left to the prompt.
Escalation that carries context
A person receives the conversation and the summary, not a bare notification.
Evaluation set
Real questions with expected answers, run on every change so you know whether it got better or worse.
The channel your customers use
Web, WhatsApp or in-product. Chosen last, after the task is defined.
Cost and observability
Usage per conversation, logs and alerts, estimated against your real volume.
Evidence
The demo shows the awkward cases
A valid answer with sources, a question it cannot support, and a handoff to a person.
What is the difference from an AI agent?
A chatbot converses: it answers, guides and escalates. An agent acts: it reads and writes in your systems, with bounded permissions and human approval on consequential actions. Many projects start as the first and grow into the second; we say which one your case needs rather than selling both.
Will it replace my support team?
No, and we would not sell it that way. It absorbs repeated questions and hands over the rest with context. We publish no deflection percentages: without a comparable baseline of yours they would be invented.
How long does a pilot take?
Weeks, not months, when the knowledge exists and the task is bounded. What stretches it is source documentation that has to be written or corrected first.
Related services
AI agents that are bounded, evaluated and hand off to a person
Agents connected to your sources and processes, with bounded actions, evaluation and human handoff.
API, ERP and CRM integrations that hold up when something fails
Synchronisation between systems with error handling, retries, duplicate control and monitoring.
Chatbot or agent?
Tell us what it should answer, from which sources, and whether it needs to do anything in your systems. That decides which one you need.