Field report 3 · AI governance
Better too careful: classifying data before every AI call
How we decide, before calling a language model, whether an input is confidential and may leave the organisation.
- Status
- In trial Design in trial, not yet in production.
- Published
- Publisher
- SIMO GmbH, Aschaffenburg, Germany
Starting questionBefore calling a language model, how do we decide whether an input is confidential and may leave the organisation?
What we trialled
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 towards “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.
Learnings
- 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