Privacy-Preserving Architecture for Enterprise AI
Companies run more and more of their work through large language models. Contracts and provider safeguards lower the risk, but you are still trusting a provider to honor what it promised about your data.
Most model routers choose a model by quality, latency or cost. I looked at whether the sensitivity of the data could decide instead. Public work can go to any approved model. Confidential work is restricted to open-weight models an organization can run on its own servers.
The thesis asks two things. How an architecture can enforce that rule rather than leave it written down as a policy, and what changes when confidential work is limited to open-weight models. To answer them I built a proof of concept for an enterprise AI assistant that keeps public, organization-managed and user-scoped information apart, reviewed it against its requirements, traced a demonstration, and ran a pilot at Siili Solutions.
The findings hold for the paths that were evaluated. Self-hosted inference and confidential documents stayed design and discussion topics rather than measured results.