Guardrails we added before letting AI draft review summaries

Posted by lena_k in c/ai-in-hr-people-analytics · 28 points

Managers were spending hours turning scattered feedback into a coherent narrative, so we let an assistant draft the first pass. It only became trustworthy once we added rules:

1. It only summarises feedback that's already been given — never invents examples
2. Every claim cites which peer or source said it
3. The manager has to edit, not just approve, before it goes to the employee