Imagine asking an AI agent to schedule a meeting. You have an oncology appointment at 10 am; the organizer has childcare issues. The agent finds a good time, but reveals your medical appointment when explaining why. You never consented to share that. This is contextual privacy for enterprise AI: reasoning about what's appropriate to reveal. Current models cannot reliably judge which details should stay private versus justify a decision. The PiSAs benchmark measures how well models handle this reasoning gap. Spoiler alert: There's a lot of work to do, and I'll continue to schedule my meetings for the time being. Presented by Valentina Zantedeschi, on September 17, 2026. Paper: https://arxiv.org/abs/2607.05318 Dataset: https://huggingface.co/datasets/ServiceNow/PiSAs Code: https://github.com/servicenow/pisas More work from ServiceNow AI Research: https://www.servicenow.com/research/