Five modes, not two
The question is usually posed as "cloud or local". In practice there are five steps, and the middle ones are where most organisations land:
- Local-only. Every request is processed in your environment. No prompt content goes to external providers.
- Local-preferred. Sensitive requests stay local; non-critical ones may — if policy allows — use the cloud masked.
- Cloud-preferred. Leading cloud models, but every prompt is inspected before transmission; sensitive content is blocked or masked per policy.
- Cloud-only with governance. For a fast start: a security and policy layer over approved cloud models, with visibility and a log.
- Air-gapped. AI use with no outbound connection. Processing, inspection, enforcement and logging with no external model access.
The point about the order
You do not have to start at the end. A pragmatic start at step 3 or 4 with a log beats a block that gets routed around — and moving up is a policy change, not a new project.
Where the software runs
BRANE is software. You decide where it runs: on a certified appliance, on your hardware to spec, or on a virtual platform to spec. On-premise is the first and strictest variant, not the only one — and in every variant the same assurance holds: no prompt crosses the boundary in cleartext.
The honest trade-off. Fully local means you provide the compute. If you do not want to, take a mixed mode and accept that non-critical requests leave masked.