Answering Assessment Questions
How to get an accurate classification.
The classification is only as good as your answers. A few rules make the difference.
Answer for reality, not intent
Classify the system as deployed today. If the model was built for internal analytics but sales started using it to score leads from EU customers, classify the lead-scoring reality.
Involve the right people
- The engineer knows what data goes in and what the model actually outputs.
- The product owner knows who relies on the output and for what decision.
- Answering alone from one perspective is the most common source of misclassification.
When you're unsure
Choose the more conservative answer. If you genuinely can't decide whether the output "materially influences" a decision about a person, assume it does — then let the Art. 6(3) follow-up questions test the exception properly.
One system at a time
Don't classify "our platform". Classify the recommender, the fraud model, and the support chatbot separately — they will land at different risk levels with different obligations.
Re-run after changes
A model swap, a new user group, or a new data source can change the classification. Re-run the assessment after any substantial modification — see Keeping Documentation Updated.