1,368 AI Failures and What They Teach About Evals — Vishal Srivastava & Tanmay Sah, New York Life

Vishal Srivastava and Tanmay Sah studied 1,368 incidents from a public AI incident database to ask what real failures say about how AI systems should be tested. They sort failures across three areas (technical strength, deployment controls and business impact) and 13 sub-areas that cover about 98% of incidents. Weak deployment controls, such as human review and monitoring, are the most closely tied to harm. They show how to rank use cases by how much is at stake, why the evaluator itself needs checking, and a build-break-fix loop that tries to prove a system wrong instead of right, applied to a real airline chatbot that gave a customer the wrong refund policy. Chapters: - 00:00 Intro - 00:41 AI failures are rising fast - 01:35 Three dimensions of failure - 03:34 13 sub-dimensions - 06:09 What covers 98% of failures - 07:12 Which failures cause the most harm - 09:43 Ranking use cases by materiality - 10:43 Breaking risk into four parts - 11:40 What to test in an agentic system - 13:14 Can you trust the evaluator? - 13:48 Confirmation bias and failure-driven evals - 15:20 Build, break, fix - 17:03 A real chatbot failure - 18:30 Recap Vishal Srivastava & Tanmay Sah, New York Life: https://www.newyorklife.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!