Agents, RAG and automation
Put your knowledge to work.
Assistants that answer from your documents and cite their sources. Automation that handles repetitive steps, with a person at the critical decisions.
WHO IT IS FOR
If the answers already exist, but finding them takes time
For companies whose knowledge is locked in documents — technical manuals, product data sheets, procedures, circulars — and where every answer requires a call to the person who “knows where it is”. And for processes that move between three tools and two people, then stop as soon as someone is away.
USE CASES
Grounded answers, processes that close the loop
A technical catalogue you can query
The sales team asks in natural language — “which model can handle these conditions?” — and receives an answer with a reference to the right data sheet. Not a plausible summary: the source, ready to open and verify.
Customer service with the procedure to hand
For recurring questions, the assistant prepares an answer citing the current internal procedure. The operator sees the source, checks it and decides whether to send: the system drafts, the person sends.
Procedures that answer new starters
Onboarding, quality, safety: a new colleague asks, the system answers from the current document and cites the chapter and version. Repetitive questions stop interrupting the people doing the work.
A process that completes itself where it can
An incoming email is read, the data extracted and a draft action prepared; a person confirms, the system records it in the management system and notifies the relevant people. From trigger to completion, with one human touch — the one that decides.
HOW IT WORKS
Built not to invent
The assistant answers only from the documents you provide: outside them, it says it does not know and passes the question to a person. Every answer cites the source — document, section, version — verifiable in one click: hallucination prevention by design, not by promise.
The source is compulsoryTransparency note
“Cite the source” is not a statement of intent: it is a design constraint. An answer that cannot be supported by the indexed documents is not returned — the system says it does not know and escalates to a person. An honest “I do not know” is better than an elegant but incorrect answer.§
The automations keep a record of every step: what came in, what the system decided, what the person confirmed. And everything can be switched off: the AI we sell is the AI you know how to disable.
The route follows the method: audit, pilot on one real process measured before and after, production with governance. In the lab, you can already try the principle with preloaded questions: answers checked by hand, with real sources. Free search will arrive with the endpoint.
You can try the lab searchDeexma Labs project
In the lab, you can try the search with real sources in preloaded-question mode. The answers are written and checked by hand, not generated by a model; free search will arrive with the endpoint.§
HONESTY
What we do not promise
An assistant that knows everything
It only knows what is in your documents. If the answer is not there, it says it does not know: that is the correct behaviour, not a limitation to hide.
Plausible sources
Citation is compulsory by design. An answer without a source is not returned: we prefer a gap to a well-presented invention.
No maintenance
Documents change and procedures are updated: the knowledge base must be maintained, and that is part of governance. Anyone promising otherwise is selling you a prototype.
A human out of the loop
Where errors are costly, confirmation remains with a person. Always: it is not a pricing tier, it is how we work.
Where does your process get stuck?
The audit takes one week and maps where an agent will pay off and where it will not — including when the honest answer is “nowhere”.
Talk to us