All business knowledge in one place, ready for AI
A company’s documents, data and processes in one place an AI can answer from, with the source on every answer and each person seeing only what their role allows. Proven on our platform; live for a client in November.
After translating documents when they’re added. Test set.
Unchanged by the translation work.
19 on-topic questions answered, 3 off-topic ones flagged as gaps.
Company knowledge lives in three places: documents (manuals, catalogues, conditions), live records (orders, quotes, invoices, leads) and processes that sit in people's heads. An assistant that sees only one of them gives generic answers, or confident wrong ones.
We hit this first on our own platform. We wanted one assistant that answers across finance, projects and support with the real numbers, and that never shows a person data their role doesn't allow.
The same need came up in a client's sales system: staff, distributors and customers asking about products, quotes and orders, in several languages, with answers that have to hold up.
Step by step.
- 01
Documents and live data are handled differently. Approved documents are indexed so they can be searched by meaning. Live records, like orders or invoices, are never copied: they are read from your systems when a question needs them, so the numbers are current.
- 02
Access follows roles. Each person only gets answers from what their role is allowed to see, and that is decided on our side, never by the person asking.
- 03
Documents are reviewed and approved before they are indexed, and translated into the site’s languages, so a question in Spanish finds a document written in English or French. Answers quote the original.
- 04
Every answer cites its source. A second model checks that the sources actually support the answer; when they don’t, the question is logged as a gap instead of answered.
- 05
Actions are proposals. The assistant can suggest a follow-up or save a quote, but a person confirms it, and it can be undone.
Built with Company documents, approved and indexed · Search by meaning, in any language · Live data read from your systems · Access by role · Sources cited on every answer
Search scores alone don’t tell you whether the system knows the answer. On the client’s documents, some off-topic questions scored higher than on-topic ones. A second model that judges relevance fixed it.
Simpler search won. Searching by meaning alone did as well as a more complex mix on every core question, so we kept it simple.
Translating every question as it came in was slow. Translating the documents once, when they’re added, made every language work with almost no extra wait.
Copying live data into a search index goes stale. Reading it from your systems when needed, with each person’s own permissions, keeps answers current.
What we decided.
Adopted. Proven on our own platform first, where it has answered from our documents and records since April 2026. Then built into a client’s sales system, going live in November.
Once a company’s knowledge is in one place an AI can use, many tasks stop depending on who remembers what. For your business, that can mean:
- Customer support that answers from your own documentation, with the source on every answer.
- Quotes and follow-ups drafted from your catalogue, conditions and the client’s history.
- Onboarding: new staff ask how something is done and get the process document.
- One search across documents in several languages.
- Every answer limited to what that person’s role may see.
- It answers only from approved sources. If a document isn't there, it logs a gap instead of answering.
- It never carries out actions by itself. A person confirms them, and they can be undone.
- Access is decided from each person’s role, on our side, never by the person asking.
- Conversation data is purged on a fixed retention schedule.
Take the client build live, then measure in production: answer confidence, logged gaps and how often proposed actions are confirmed. The gap log tells the client which documents to write next.
How do you stop an AI assistant from making things up about our business?
It answers only from retrieved, approved sources and cites each one. A second model grades whether those sources support an answer; when they don't, the question is logged as a gap and not answered.
Can customers and staff use the same AI without seeing each other's data?
Yes. Each person only gets answers from what their role is allowed to see: a customer sees customer documents and their own orders; staff see what their role allows.
Does it work if our documents are in different languages?
Yes. Each approved document is translated at indexing and the translations are searchable. In testing, every cross-language question found the right document first.
Do we need to move our data into a new system?
No. Documents are indexed for search. Live records stay where they are, in your ERP or CRM, and are read when a question needs them.