People make commitments in Slack all day. “I’ll send it Friday.” “We’ll fix that before launch.” Then the thread scrolls away and the commitment goes with it. A SaaS team wanted an AI layer that catches those promises as they happen and tracks them to completion. The risk with a feature like this is obvious once you look at real conversations: sarcasm reads like a promise, questions look like commitments, and people commit on behalf of other people. So we started small: a paid discovery phase instead of a big contract.

Discovery produced the thing most AI features never get: an evaluation set built before the production code. I collected and categorized 46 edge cases covering the ways detection goes wrong, and every iteration was measured against them. Ten days after discovery started, the production MVP was live with 87% detection accuracy against the evaluation set. Every prediction is traced, and the failure paths are explicit, so the system degrades in a way the team designed, not one they discover from angry users.