Frontier model or open weights? A framework for regulated data

The default advice on AI projects is to use a hosted frontier model. For many projects that is right. For health, financial services, defence and public sector work it is often not available, because the data is constrained to a jurisdiction.

Start with the data, not the model

Split the workload by what it touches. Non-sensitive material can use any hosted model. De-identified operational data usually can, under a processor agreement. Patient records, payment credentials and classified material often cannot leave at all.

The choice should stay reversible

The most useful decision is refusing to let model choice harden into the codebase. A thin inference gateway in front of every model makes moving from hosted to self-hosted a configuration change rather than a rewrite.

The pattern we ship

For restricted workloads: a small self-hosted model on-premise for sensitive extraction and classification, with a hosted model only on de-identified or public material.