Where your data lives and who issues access

Neuraan's infrastructure runs on AWS, in the United States. For everything else it uses other enterprise providers: Microsoft Azure for the language model, MongoDB for the agent's information, Zilliz for search and Langfuse for logging every run. Neuraan has no servers of its own. Each provider applies its own protection and privacy policies.

The infrastructure is shared and the isolation is logical: each agent has its own key, configuration and files, separate from those of other agents. If your contract requires it, we set up a dedicated environment for your operation.

The language model runs in an enterprise Microsoft Azure tenant, under contract. It receives the content of each step the agent reasons through: instructions, messages and tool results. Your data isn't used to train models. That's why we use a paid enterprise service and not free models.

When your agent can't solve something, it logs it and suggests how to improve. That improvement applies to your agent, not to a language model.

Data is encrypted in transit and at rest.

Your team issues the access. In each system, the agent signs in with its own user, not that of someone on your team, and only has access to what you allow. Depending on the system, it connects via API, through files or directly on screen when there's no other way. Your team decides whether that user can only read or can also write.

You can cut off the agent's access at any time by deactivating its user in your system. The effect is immediate and doesn't depend on Neuraan.

You decide which actions need your sign-off. For those, the agent shows you what it's about to do and waits for your confirmation.

Your business rules always run the same way. The language model interprets messages and drafts replies; it doesn't decide your rules.

Each tool is validated before it runs and runs in a restricted environment isolated per agent: it can only use the operations it's allowed and has no access to anything outside its agent.

Every step the agent takes is logged. Access to that log within Neuraan is restricted, and we hand it over whenever you ask.

The dashboard requires a verification code at every sign-in. Sessions expire after 30 minutes and we monitor unusual access.

If an integration goes down in the middle of a case, the agent retries, notifies your team and records how far it got so it can pick up from there. The other cases in the batch carry on. Every change to the agent is versioned and can be rolled back to a previous version. An action already executed in one of your systems doesn't undo itself: it's corrected with another action, by the agent or by your team.

An executor, the Edge Agent, is installed on that machine. It carries no language model: it only opens an outbound connection to Neuraan, executes the allowed actions and returns the result. By default it asks for your confirmation before running commands.

We don't have certifications yet. Our architecture is being aligned with SOC 2, and controls from that framework are already in place. AWS, Azure and MongoDB Atlas have their own certifications; they aren't Neuraan's.