Permission-aware knowledge assistant
Answers questions from internal documents, using only what the person asking is allowed to see, with a source for every claim.
The problem
Company knowledge is spread across wikis, drives and tickets, and an assistant over all of it creates a new risk: an answer that quietly includes something the asker should not see. Many internal assistants either block too much or leak.
Work it takes off your team
- Searching across many tools for one answer
- Asking colleagues questions the documents already answer
- Checking whether an answer is current
How it works
Ingestion with access rules
Each document keeps the access rules of the system it came from, so permissions travel with the content.
Retrieval filtered by the asker
Search results are filtered by who is asking before the model sees any text.
Answers with sources
Every claim links to a passage. When nothing relevant is found, the assistant says it does not know.
A feedback and evaluation loop
Real questions with expected answers run on every change, and thumbs-down feedback feeds the next test set.
Who needs this
The teams that feel this problem every week, and the industries where it shows up most.
Teams that feel it
- Support and customer success teams
- Sales engineering and pre-sales
- HR and IT helpdesks answering the same policy questions
- Operations and field teams with manuals
Industries where it fits
Is this only for large companies?
No. The same design scales down. What changes is how much is connected and how many people are involved.
Enterprise
Large document estates with strict access rules. Permission-aware retrieval is what makes an assistant safe to roll out widely.
Mid-size teams
Teams stop asking the same colleagues the same questions, and new joiners find answers on day one.
Small and micro businesses
A small team with a shared drive and a few hundred documents can run a hosted assistant over it, with simple access groups, as a lighter version of the same design.
How it plugs in
It connects to the tools you already use, through narrow, logged integrations. Nothing is replaced.
Who runs it day to day
A knowledge or IT owner. Document owners keep content current.
What you can expect
Time freed is the headline, but it is not the only result.
Less time searching and asking
People get a sourced answer in one place instead of searching several tools or interrupting a colleague.
Faster onboarding
New joiners can answer their own questions from day one.
Answers you can check
Every claim carries a source, so fewer wrong answers are acted on.
Knowledge gaps become visible
Questions the assistant cannot answer show which documents are missing or out of date.
Estimate the hours it could free up
Move the sliders to match your own team. The starting values are illustrative, only there to show the method.
Estimated time freed
9.0hours a week
432 hours a year, about 54.0 working days
Formula: volume × share handled × (minutes today − review minutes), over 48 working weeks. This is a planning aid, not a promise. Real results depend on your data, your process and what the evaluation shows during the build.
Decisions that mattered
Filter before the model
Text the person cannot access never reaches the model, so it cannot appear in an answer.
Honest about gaps
A clear 'I do not know' beats a confident wrong answer in a company setting.
Fresh by design
Ingestion runs as a repeatable job, so edited and deleted documents are reflected.
Built with: Python, a vector and keyword index with per-document access metadata, connectors to wiki and drive style sources, citation checks and an evaluation set.
This is a reference build on synthetic data. It shows how we would approach the problem and is not a client project. It makes no claims about results.
More builds
Let us open a conversation.
Have an agent or an automation in mind, or a collaboration to explore? Tell us what you want to build. We reply to every message with an honest view of fit and next steps.