Home › Could it leak information between departments?
Could it leak?
Keeping departments and clients apart inside one system
The short answer
When one AI assistant serves several departments, or several clients, something has to keep their documents apart.
There are three ways to build that wall: separate systems, separate sections inside one system, or labels and filters on shared storage. The stronger the wall, the more it costs. The cheaper the wall, the more it depends on every piece of software getting every request right.
Think of a building with several tenants
You can give each tenant its own building. You can give them separate floors in one building, each with its own locked door. Or you can put everyone in one open-plan office and trust the colored labels on the filing cabinets.
All three can work. But in the open-plan office, one person reading the wrong label is all it takes.
Three ways to build the wall
The assistant keeps your documents in a searchable catalog. Engineers call the people or groups who must be kept apart tenants: departments, subsidiaries or clients.
Separate systems
Each tenant gets its own catalog on its own service. Microsoft describes this for its Azure search service as offering the "highest level of data and workload separation" (Microsoft Learn).
Gives you: the strongest wall. A mistake in one system can't reach another tenant's documents.
Gives up: money and effort. Microsoft notes the cost per tenant is higher, and every update has to be made many times.
Separate sections
One system, with a walled-off section for each tenant. Pinecone, one search database, says its sections are "stored separately", which reduces "the risk of application bugs that could search the wrong tenant's data" (Pinecone docs).
Gives you: a real wall at a lower cost, and easy deletion of one tenant's data.
Gives up: the software still has to pick the right section for each request.
Labels and filters
Everything sits together, and every document carries a label such as "HR" or "Client A". Each search filters by the label.
Gives you: the cheapest and most flexible option, and it suits people who belong to several groups.
Gives up: the wall is only a filter. Microsoft says this model "offers less data and performance isolation than the other models" (Microsoft Learn).
Step through it
One question, three kinds of wall
- A lawyer working for Client A asks the assistant a question.
- Depending on how the system is built, the question passes through one of three kinds of wall.
- When everything works, all three return only Client A's files.
- But with labels, one search that leaves the filter off brings back Client B's files too.
- Separate sections and separate systems put a real wall in the way, so the same mistake is far less likely to reach another client.
How a well-built system does it
- The wall matches the risk. Where mixing would be a legal problem, such as two clients of one law firm, or a deal team and everyone else, the stronger walls are worth their cost. Labels are fine for ordinary departmental tidiness.
- The system decides who is asking, not the request. Microsoft's guidance is that a trusted part of the system must work out the tenant from the person's login and "apply the corresponding filter to every query", and never accept a tenant name sent by the user's device (Microsoft Learn).
- The database enforces it, where it can. Some databases can apply the label rules themselves, so a forgetful piece of software can't skip them. PostgreSQL calls this row security (PostgreSQL manual).
- It extends walls you already have. Many firms already run "ethical walls" in their email and file sharing. Microsoft's version, Information Barriers, exists so that "a lawyer's data obtained from one client can't be accessed by a lawyer at the same firm who represents a different client" (Microsoft Learn). The assistant should respect the same walls.
- Inside each wall, the usual checks still apply. Keeping departments apart doesn't replace checking who may see each document. Why the check must happen before the assistant reads anything.
How each wall can fail
- A forgotten filter. With labels, one search that leaves the filter off reaches everything.
- Built-in exceptions. Even database-enforced rules have them. PostgreSQL's row security is off until someone sets it up, and powerful administrator accounts "always bypass" it (PostgreSQL manual).
- Shared parts behind the wall. In a shared catalog, Microsoft notes other tenants' documents "can affect relevance scores": the wall stops people reading each other's files, but not every trace of them (Microsoft Learn).
- Walls that don't cover everything. Microsoft's Information Barriers, for example, "can't restrict communication and collaboration between groups and users in email messages" (Microsoft Learn). Ask what each wall does and doesn't cover.
It has happened
In June 2025, Asana took its new AI connector offline for nearly two weeks after finding a flaw that could expose one organization's information to other organizations' users. Asana said it was "a logic flaw" and "not the result of a hack", and told BleepingComputer that roughly 1,000 customers were affected (BleepingComputer, The Register).
Questions to ask your IT team or vendor
- Which of the three walls do you use between our departments, and between us and your other customers?
- Where does the system decide which department or client someone belongs to, and can a user change that?
- Is the wall enforced by the database itself, or only by the application? Which accounts can get around it?
- Does the assistant respect the ethical walls or information barriers we already have?
- Can we see it tested: someone in one department asking for another department's documents, and getting nothing?
Other names you may hear
- Tenant, multi-tenancy
- One system serving several groups that must be kept apart.
- Tenant isolation
- Whatever keeps those groups apart.
- Namespace, partition, collection, shard
- Names for a separate section inside one system.
- Metadata filtering
- Labelling every document and filtering every search by the label.
- Row-level security
- A database enforcing the label rules itself.
- Ethical wall, information barrier, Chinese wall
- A rule that two groups in one organization must not share information.
- RAG
- Retrieval-augmented generation: the assistant looks things up in your documents, then writes the answer. How it works, at EnterpriseLLM.ai.