Private and on-premises AI · 2 min read

Running an AI agent on company servers: production architecture

Running a model locally is only one component of an on-premises agent; knowledge, tools, identities, audit evidence and safe failure behaviour complete the system.

Draw the real boundary

Locate the interface, model, documents, search index, tools, logs and backups, and identify every telemetry, update or external API path. Assign owners for the data, workflow and operation.

Document update, telemetry and identity dependencies alongside model hosting; an external call changes the claimed operating boundary.

Separate prediction from authority

Let the model propose a response or tool call, while deterministic code checks the user, permission, fields and current state. Never substitute a broad administrator credential for this control.

A deterministic policy layer should reject unauthorised tools and malformed arguments before they reach any business system.

Build inspectable layers

Use an authenticated gateway, pinned model service, permission-aware retrieval, narrow tool adapters, event records and infrastructure monitoring. Add write actions individually with confirmation and replay protection.

Version model, retrieval index, prompts and tool contracts together so an incident can be reproduced against the released configuration.

Accept a workflow, not a demo

Test normal, ambiguous, forbidden and adversarial cases, unavailable dependencies and cross-user data attempts. Begin with one team and retain evidence sufficient to reconstruct each action.

Acceptance should cover unavailable services, hostile documents, cross-user requests and repeated operations, not only correct answers.

Evidence and sources

Continue with the underlying material

pommeDeTerre

Need an estimate for your project?

Tell us about the project. We will break it into stages and explain the budget drivers.

View pricing