Open your IT systems to AI without handing over control.
Reading through the screen, reading from the database, writing to the database: each type of access carries its own level of risk and its own guardrails. LeBrain enforces all of them the same way, through the Hub, and keeps an audit trail.
The bot sees what a user sees, with that user's permissions. The application's own controls stay in place.
Dedicated bot account, action logData is read outside the application's controls, and a heavy query could slow down production.
Read-only, predefined queries, anonymization before the modelThe application's business rules are bypassed; one error could change a record with no check in place.
Approval before write, full log, rollbackThe database should not be ruled out: it opens up in stages, reads first, then writes under control.
LeBrain opens up the database, writes included, behind 5 guardrails.
Every time an agent touches a database, it goes through the same chain of controls, enforced by the Hub. You get the speed of direct access with the control of application-level access: permissions, audit trail and approvals stay in your hands.
1 · Least privilege. A dedicated service account, read-only by default; writes are limited to approved tables.
2 · Predefined queries. Skills run parameterized, pre-approved queries, never free-form SQL written by a model.
3 · Anonymization. Sensitive data is masked before any call to a model.
4 · Approval before write. Rule-based checks and human approval come before any change to a record.
5 · Audit log and rollback. Every read and every write is logged, replayable and reversible.
| Least privilege | Predefined queries | Anonymization | Approval before write | Audit log and rollback | |
|---|---|---|---|---|---|
| Read | applied | applied | applied | not required | applied |
| Write | applied | applied | applied | mandatory | applied |
Read: 4 guardrails. Write: all 5.
The model interprets. Automation executes, the same way every time.
Nothing that touches your systems ever depends on a model's answer: the agent prepares the parameters, clear governance checks them, and tested automation applies them. The same rule holds across all 6 modules.
No lock-in to any model provider.
One access point for every model
LLMs, SLMs, OCR and vision models all go through the same gateway: quotas, routing, and a log of every call. Switching models leaves everything else unchanged.
Bring your own model
The Hub works with the model of your choice, including through your own AI subscription.
Sensitive data stays masked
A dedicated agent detects and pseudonymizes personal data before any call to a model.
Everything can be replayed
Every call, every read and every write is logged, along with who requested it, and can be replayed.
Who can use which model
Permissions are granted per team and per agent; Governance gives you the full picture.
What we deliver is yours
Applications, code, documentation and requirements baselines are handed over to you; they run without us.
What CIOs and CISOs ask us.
Can an agent write directly to our databases?
Only behind the 5 guardrails: a least-privilege service account, predefined queries, anonymization, approval before write, and an audit log with rollback. Access is read-only by default, and writes are limited to the tables you approve.
Can a model generate and run its own queries?
No. Skills run parameterized queries that are approved in advance. The model picks the operation and prepares its parameters; it never writes SQL that gets executed as is.
Which models do you use?
The ones you choose. The gateway supports language models, lightweight models and vision models; you can switch without rebuilding your agents.
Is our data sent to model providers?
Sensitive data is masked before any call to a model. You choose the model, and therefore where processing happens; we define this with you from day one.
What happens if an agent makes a mistake?
An agent can only do what clear governance allows. Writes require approval, every action is logged and reversible, and exceptions are surfaced in Governance.
Where are the delivered applications deployed?
On your infrastructure. The code, documentation and deployment scripts are handed over in your repository.
Security backed by the company behind LeBrain.
LeBrain is developed by Novelis, whose information security management system is ISO 27001 certified and which holds ExpertCyber, the French government-backed cybersecurity label.
Let's talk about your first use case.
A 30-minute intro call to talk about your needs.