Technology
A knowledge assistant that answers questions from your documents can be built in an afternoon. A knowledge assistant you dare to use in a governmental setting cannot. Below is where the difference sits.
01
Permissions
“Can someone from social services reach HR documents through this thing?”
No, and that is the first thing set up, not the last. Nor the source reference: no pointer to a document they may not open. An authorisation test is part of the delivery.
Permissions
The same question, three roles, three answers
What the assistant should answer depends on what the person asking may open, and not only on what the documents say.
Question
“What does the latest appraisal of colleague De Vries say?”
Employee
May open
- Policies and regulations: may open
- Documents of the own department: may open
- Personnel files: may not open
Answer
This information is outside your access rights.
No source. No reference to the file. No confirmation that it exists.
Manager
May open
- Policies and regulations: may open
- Documents of all departments: may open
- Personnel files: may not open
Answer
This information is outside your access rights.
Being allowed to see more is not being allowed to see everything. An appraisal belongs with HR, not with the line.
HR
May open
- Policies and regulations: may open
- Documents of the own department: may open
- Personnel files: may open
Answer
The latest appraisal dates from March 2026 and was rated “good”.
Source: Personnel file De Vries, page 2
Illustration with fictitious data. The permissions come from your own environment; the assistant adds none and removes none. This is the design principle and it is tested with an authorisation test before anything goes into production.
- may open
- may not open
This is why most projects stall at a general knowledge base holding only public documents. The moment personnel files, objections, care files and executive papers have to enter the same system, the question changes from “can it answer” to “can anyone demonstrate that the separation holds”.
That is stricter than it sounds. The answer must contain no content from a document the user has no right to. The source reference must not point to it. And the system must not confirm that such a document exists either, because that is information in itself.
That is why we set this up first, separated per department and per corpus, and test it in your own environment before anything goes into production. An authorisation test is part of every delivery.
02
Provenance
Every answer points to the document and the passage it came from. Not for tidiness, but because a member of staff who decides something on the basis of that answer has to be able to show what the decision rests on.
03
Staleness
“What happens to a document we withdraw?”
It should disappear from the assistant along with it, and the period within which that happens is an agreement we record with you before anything goes into production. Ask this of every supplier you speak to.
Regulations are withdrawn, policy is revised, versions pile up. An assistant that quotes the old version because it happens to match better does more damage than it solves. Document management is therefore not an afterthought but part of the design.
04
Silence
“And what if it simply makes something up?”
Then you will see it, because a source is always attached and it opens with one click. You should not get an answer without a source, and that is one of the six measurement points before anything goes into production. That is not a guarantee that the model never goes wrong, it is the reason you notice before anyone acts on it.
The hardest behaviour to achieve is no answer at all. A language model is built to answer, and it will do so for questions whose answer is not in your documents. An assistant that says it does not know is worth more than one that is often wrong.
05
Fragmentation
Documents are rarely in one place. SharePoint, a network drive, email, a line-of-business application, each with its own structure, metadata and permissions. Bringing those together without losing the permissions is the bulk of the work, and the least visible part.
Open-weight models
Open-weight models are language models whose weights are freely available, so they can run locally instead of at a provider. For a government body or an organisation holding confidential documents that matters in two ways: no data goes to an external service, and for the core of the setup you do not depend on a supplier who changes his terms, his prices or his models. What you notice is nothing special: the same assistant, the same answers with sources, without internet.
Runs where you want
The same setup runs in your own Azure tenant or on your own hardware. We deliver that setup as a whole, not as a collection of parts for you to tie together. Where it runs is your choice, and that choice changes nothing about the five failure points above.
Microsoft certified, not Microsoft dependent. For organisations already working in Microsoft 365, their own Azure tenant is the logical place. For those who want no data leaving the building, the same setup runs without cloud services on their own hardware.
These are the five that stop projects in practice, and this is what we have worked on. The setup runs in your own Azure tenant or fully offline on your own hardware. Exactly how we approach it is something we show in a conversation, not on a page.
More questions we already knowWant to know whether this works for your organisation?
An exploratory conversation of one hour usually gives a clear picture. Get in touch and we will schedule it.
We show it in your own environment. Without internet too, on a laptop at your table.
besst.ai. Your documents answer questions. Only for those allowed to see them.