AI governance
Better too careful: classifying data before every AI call
Starting questionBefore calling a language model, how do we decide whether an input is confidential and may leave the organization?
Result
Open
Design in testing: a local model classifies every input, as confidential when in doubt. That costs model quality, deliberately.
A local model classifies every input before it leaves the organization and treats it as confidential when in doubt. Only clearly uncritical borderline cases go on to central review. That direction is built in and cannot be relaxed by configuration.
What we tested
We designed a two-stage classifier. A local model labels every input. When in doubt, it marks the input as confidential and does not forward it to the central review. Only clearly harmless borderline cases go there.
The bias toward “confidential when in doubt” is built in and cannot be relaxed through configuration.
What it means for business architecture
Classification is the precondition for splitting requests between a local and an external model. Without it, every hybrid architecture remains a matter of trust.
Lessons learned
- Erring on the side of confidentiality costs model quality. We accept that deliberately.
- Two classifiers have to stay aligned, otherwise users see contradictory results.
- A classification that only runs centrally would send out exactly the data it is meant to protect.
- data classification
- hybrid architecture
- local AI
- confidentiality