Two products that are easily confused

Outward customer dialogue and the inward assistant look alike on the surface — both are an input field with an answer. The requirements are opposites. Outside, what counts is that nothing false and nothing confidential is said; the bot may well do less. Inside, depth counts: staff need access to details no customer may see, and they notice immediately when answers stay superficial.

Solving both in one system usually means getting neither right.

Why every assistant must be able to stay silent

Language models are trained to appear helpful. Without explicit permission to know nothing, they fill gaps — convincingly phrased. The countermeasure is not prompt cosmetics but a rule in the answer path: if the evidence found is insufficient, the correct output is “I find nothing on that”, plus the route to a human.

An honest non-answer costs one click. An invented answer costs trust — and depending on the information, money too.

Permissions belong beneath the search, not above it

The most common design flaw in internal assistants: search first, filter afterwards. Confidential content then appears in intermediate processing and can bleed into the answer. The correct order is the reverse: the asker’s role defines the search space before the first hit exists. An assistant must never show more than the same person would be allowed to see in the source system.

Connecting messengers without walking into a dead end

On your website: full control over data and design, no third-party provider in between.
WhatsApp and similar: wide reach, but content passes through a third party — that belongs in the privacy policy and in the decision about which information is given there at all.
In every case: a clean handover path to a human, including the conversation history. Nothing frustrates more than having to explain everything again to the staff member.

How we support you

Separating customer dialogue from the internal assistant — each with its own source and permission concept.
Connection to your knowledge sources via a RAG system with source citation and a controlled non-answer.
Permission filtering before ranking, including negative tests with several roles.
Handover path to your team including the conversation history, so nobody has to ask twice.