Draft for reference — this content must be reviewed by legal counsel before official publication.
Last updated: Draft — not yet in effect
1. Accurate and grounded
Agents answer from approved knowledge and data, with citations. When evidence is insufficient, they say so and hand over to a person rather than guess.
2. People stay in control
High-impact decisions such as credit approvals or claims always include human review.
3. Transparency
Customers know they are talking to AI. Every conversation and agent action is recorded for traceability.
4. Fairness
We evaluate agents across customer groups, accents and regions to detect and reduce bias.
5. Tested before release
Every change is automatically tested against real conversations and adversarial scenarios. Versions below quality thresholds are not released.
6. Continuous monitoring
Quality, safety and customer sentiment are monitored in real time, with alerts and incident procedures.
7. Data protection
Customer data is never used to train shared models. Sensitive information is redacted before reaching any model.