The challenge
Some managers spent considerable time drafting performance evaluations. The organization wanted to make that work easier while keeping feedback tied to role expectations and giving people room to question ratings that might reflect bias.
How we put human oversight first
We started by defining what the AI could help with and where people had to make the decisions. The assistant could help organize and draft feedback and flag language or reasoning worth a closer look. It had no authority to assign or approve ratings.
- Use the organization’s existing standards. We configured the assistant around the review process, role expectations, core value behaviors, and guidance from bias training. Those materials gave the assistant context for drafting and flagging possible issues.
- Keep managers responsible for the evidence and feedback. Managers reviewed and revised the drafts and remained accountable for their evaluations. A flag was a prompt to revisit the evidence, not a finding of bias or a recommendation to change a rating.
- Preserve HR review and manager calibration. HR continued to review evaluations, and managers used their existing calibration process to compare how they applied rating criteria and discuss inconsistencies. People retained responsibility for final ratings.
What changed
Some managers spent less time drafting. In some cases, the assistant prompted managers to check whether recent events were getting too much weight. Managers could then look back across the full review period and decide whether the evidence supported their feedback.