Why we pulled the plug out of a perfect working AI Model
summary
Builds a classifier that assigns per-attendee no-show probabilities and surfaces ranked lists; on one campaign it outputs a 2000-person list with ~75% no-shows but timing kills actionability because most registrations occur weeks before events so there’s no lead time to convince attendees. Heavy GxP/CSV validation overhead and lack of process ownership prevent deployment, so the model works offline but fails to reach production for over a year; solution: simple A-to-Z POC, sandbox, and embed AI into the process from design.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 2:36Abandoning a working AI model
- 7:51Model design and promise
- 10:40Timing flaw in model
- 13:34Lessons from failed AI
- 17:28AI in regulated industries
- 20:39AI as foundation
- 22:27Conclusion
related talks
36mAvoiding 5 missteps that undermine your AI readiness and success
34mMachine Learning to know if you R coming
24mChallenges in building Machine Learning models in production
14mHow We Turned AI Hype into Real Value
23mHow to fail AI implementations
25mWhy Good Models Fail After Deployment
59mWhy Great Models Fail: Lessons From 9 Years of Deploying ML Models
44mAI Engineering Anti-patterns