Problem
An organization using several AI models needs a consistent rule for where each request goes. Ad-hoc choices leak sensitive data or waste money.
Flow
- Classify the request and attached context by sensitivity (public, internal, restricted) and by difficulty.
- Check policy: which deployment targets are permitted for this workspace and class?
- Keep restricted work inside the organization's environment.
- Mask names, numbers and identifiers before any permitted external call; unmask the response internally.
- Escalate to a stronger model only when confidence is low and policy allows.
- Log the route taken and why.
Policy table
| Data class | Private model | Cloud model |
|---|---|---|
| Public | Yes | Yes |
| Internal | Yes | If the workspace allows, with masking where required |
| Restricted | Yes | Never |
Evaluation
A router is only as good as its tests. We define a test set per organization — including Bangla, English and mixed-language questions — and measure quality by route, the rate of policy violations (target: zero) and latency. No figures are published before they are measured.
Trade-offs
Masking can remove context a model needs. The policy therefore lets an owner decide, per workspace, which routes are worth it.