Permission-aware AI
Ask the same question, get a different answer, and both are right. Why permission-aware AI matters, and how GYCO applies roles, regional access and document clearance.
Picture two people in the same company asking GYCO the same question: “What was on-time delivery last month?”
One is a regional operations manager responsible for the warehouses in the south. The other is the national head of
operations. They get different numbers. The regional manager sees the figure for their own warehouses. The national head sees the whole network.
That isn’t a bug. Both answers are correct. Each is the answer within what that person is allowed to see.
The risk with “one AI for everyone
A chat assistant that can see everything is a shortcut around every access rule your organisation has. Someone who
can’t open the HR folder could simply ask about the disciplinary procedure. A regional team could ask about another region’s margins. A new joiner could ask what’s in the board minutes.
Access control in an AI tool can’t be an afterthought layered on top of the answers. It has to decide what the AI can
read in the first place.
Three layers of access in GYCO
Roles decide what you can do
Every user has a role, and every role has a permission matrix. It sets what that role can read, change or run in each
area of GYCO. GYCO comes with four default roles: viewer, analyst, manager and admin. Only administrators can add documents to the knowledge base.
Row-level access decides which data you see
A user can be given a designation, such as “operations manager, South region”
, that limits them to their own rows of
data. GYCO builds that limit into the query behind every question they ask. It isn’t a filter applied to the finished
answer: rows outside their scope never come back in the first place. The same limit applies wherever GYCO reads
business data on their behalf: chat answers, dashboard charts, insights and alerts.
Row-level access applies to the people you give a designation to, so set one for anyone whose view should be limited.
Clearance decides which documents you can read
Every document in GYCO carries a department and a clearance level from 1 to 5. A user sees a document only if their own department and clearance allow it. That check happens inside the search itself, before the AI sees a single
passage. People who add documents can’t give them a clearance above their own.
Invisible, not just locked
There’s a subtle point here that matters. When you’re not allowed to see something in GYCO, it doesn’t tell you the
item exists and you can’t open it. As far as you’re concerned, it isn’t there.
That’s deliberate. “Access denied” is itself information, because it tells you that a document or a record exists. GYCO treats it as not found instead.
The same rules, everywhere
Access rules only work if they hold on every path. In GYCO, the same roles, designations and clearances apply
whether you’re chatting, opening a dashboard, reading an insight or receiving an alert. Downloads follow the same
logic: an exported answer is available only to the person who produced it.
A record of who changed what
Permissions change over time. People join, move teams, take on a region and leave. GYCO’s audit log records:
- sign-ins;
- changes to users, roles, departments and settings;
- document uploads;
- changes to KPIs and alerts
Each entry shows who made the change, and what was changed, before and after. Administrators can filter the log
and export it.
When someone leaves, their account is deactivated rather than deleted, so the record of what they did stays intact.
Questions to settle before you switch it on
Permission-aware AI is only as good as the permissions you give it. These are the conversations worth having early
when you set up GYCO.
- Who needs a designation? Anyone whose job is limited to a region, a site or a client should have one before they start asking questions.
- What clearance does each set of documents need? Clearance is set in GYCO when documents are added, so agree the levels with the document owners first. Don’t assume the permissions on the source folder will carry across.
- Who administers it? Keep the admin role to the few people who need it. They’re the ones who can add documents and change who sees what.
- Who reviews the audit log, and how often? A log nobody reads protects nobody. Name an owner.
The point
The goal isn’t to make AI cautious. It’s to make it trustworthy in the same way your other systems are. Everyone gets
fast answers, and everyone gets the answers that are theirs to have.
