The question that stops serious teams
When a B2B team evaluates putting AI into their support, the first objection is almost never “does it answer well?”. It’s “what is this thing going to be able to see and do with my customer data?”. And they’re right to ask: an AI that ignores internal permissions isn’t an improvement, it’s a security breach with good copywriting.
The generic copilot that reads “everything there is” can leak to a junior agent information from an account that isn’t theirs, or execute an action nobody authorized. In B2B, that’s not a detail: it’s the reason many AI pilots get cancelled.
RBAC: the AI shouldn’t see more than the person
The principle is simple and should be non-negotiable: the AI operates within the same permissions as the person using it, not above them.
That is role-based access control (RBAC) applied to AI as well. If an agent can’t see a certain account’s data, the assistant shouldn’t use it to reply on their behalf either. If a role can’t execute a certain action, the AI can’t execute it “for convenience”. The user’s permission is the ceiling, always.
In practice, that means three rules:
- Scoped reading: the AI only accesses the context the user has the right to see.
- Scoped action: the AI only proposes actions the user’s role could perform.
- No silent escalation: the AI never expands its own permissions to complete a task.
Per-account context, isolated
Beyond the user’s role, there is isolation between accounts. In a multi-tenant environment, one customer’s data can’t leak into another’s context, not even through the AI. Elevatia works with context isolated per organization and EU cloud data residency — something we detail in B2B SaaS on European cloud and GDPR.
When the assistant helps on a case, it does so with the context of that account and nothing else. There is no “model that has seen everything” mixing information between customers.
Human approval: control over actions
Seeing data with permissions is half. The other half is what the AI does with it. Here Elevatia’s principle is that the AI proposes, the human decides: actions with impact go through approval before they are executed.
That turns the AI into an accelerator, not an autonomous risk. The agent reviews the proposed response or action, sees where the context comes from, and approves or corrects. We develop this model in detail in AI in support with human approval.
Traceability: who, what and when
Security without a log is blind trust. Everything the AI proposes and everything a human approves is logged: what context was used, what action was proposed, who authorized it and when. That traceability serves to audit, to meet customer requirements and to improve the system with real data instead of hunches.
What to review when evaluating AI in your support
If you are weighing an AI solution, these are the questions that separate the serious from the risky:
- Does the AI respect the user’s permissions, or does it see more than they do?
- Is the context isolated per account in multi-tenant?
- Do actions with impact require human approval?
- Is there an auditable record of what it sees and does?
- Where does the data reside?
If any answer is “no” or “unclear”, the risk is yours, not the vendor’s.
Try it with your own rules
The best way to validate security is to test it with your roles and your real accounts in an isolated environment. That’s what the pilot is for: request the free pilot (~4 weeks, no cost) or book a 15-minute demo.
info@elevatia.io — we reply in English.