USAI security
AI that only knows what you’re allowed to know
For a US AI-security company recognised by a leading industry analyst, we built the guard that checks every question against who is asking — before the AI ever sees the data.
- Client
- A US AI-security company recognised by a leading industry analyst
- Sector
- AI security · US
- Dates
- 2024–26
- Headline result
- 68% fewer wrong answers from the AI assistant
The problem
Companies want an AI assistant that answers from their own documents. The risk is obvious: ask it the wrong question and it may happily read out a salary file, a contract, or a customer record to someone who should never see it.
What we built
A guard that sits in front of the AI. Before any document reaches the model, it checks who is asking and what they are already allowed to open — using the company’s existing permissions — and keeps a record of exactly what was shown to whom.
What changed
The assistant answers from the right documents for the right person — and gives fewer wrong answers because it isn’t reading things it shouldn’t.
How it works
- 1
Identify
every question is tied to the person asking it.
- 2
Filter
documents are checked against that person’s existing permissions before the AI sees them.
- 3
Answer
the AI works only from what the person is allowed to read.
- 4
Record
a running log shows which data was used for which answer, ready for an auditor.
Results
What it delivers.
From the first version to the live system
Permissions applied before the AI reads anything
Which data answered which question, for whom
“If the AI can see everything, it will eventually say everything. The fix is to decide what it sees before it reads.”
Have a problem
like this one?
In two weeks, a Discovery Sprint shows you what superintelligence would do with it.