Deploying AI without your data ever leaving: a guide for security owners

I've seen more AI projects stopped by one security question than by any technical problem. And it's always the same question: where does our data go? This post answers it the way you answer a CISO - with architecture, not slogans.

The principle: AI comes to the data, not the other way around

The common approach ships your data out - to a vendor's cloud, to someone else's model. The right approach for organizations is the reverse: the agent runs inside the organization's network, on-prem or in your VPC, and the data never crosses the perimeter.

How it's actually built

• Your keys. The model connection uses your own enterprise API account - or an open-source model running fully in-network. You choose the provider, and you can switch.
• Read-only by default. The agent reads from systems; writing requires human approval. A critical action without a human signature simply doesn't exist.
• Least privilege. The agent sees only what its process needs. Not "access to the whole ERP" - the relevant tables.
• Full audit log. Every action is recorded: what the agent read, what it proposed, who approved. When the audit comes - it's all there.

Security isn't the blocker - it's the feature

In my conversations, the moment management relaxes is when they understand this architecture exists. Suddenly you can connect the agent precisely to the sensitive systems - because that's where the biggest value is, and that's where nobody dared to touch.

References

• Claude Enterprise - an example of an enterprise API with data commitments
• Our security page - the full architecture