# 1,368 AI failures point to deployment, not models _New York Life researchers coded 1,368 AI failures and found deployment controls, not model strength, best predict harm._ **Published:** 2026-10-22 **Source:** https://www.startuphub.ai/ai-news/technology/2026/1-368-ai-failures-point-to-deployment-not-models --- [AI Engineer](https://www.youtube.com/watch?v=mMw3arwQsL0) hosted New York Life researchers Vishal Srivastava and Tanmay Sah with a dataset of 1,368 AI failures and a blunt claim: evals that only test the model miss where harm actually happens. I think they are right, and their numbers make it hard to defend frequency-based testing any longer. No attacker needed here, the affected systems are ordinary enterprise and agentic deployments that fail through bad controls, not a remote exploit. Srivastava said the team coded every incident against a 13-dimension framework built around three questions: how strong is the system technically, what deployment controls contain errors, and what business context absorbs the impact. Six sub-dimensions cover technical strength from data quality and system design to predictive validity, fairness, failure modes and explainability, four cover deployment including decision authority, human mediation, exposure and monitoring, and three cover business impact including financial, legal and scale. Using those dimensions they can cover almost 98% of failures, with an average of six sub-dimensions present per incident. Not all dimensions matter equally. Srivastava showed harm association versus frequency and argued deployment controls dominate risk. Human mediation and monitoring that catches harm before it spreads sat high on harm association, while scale and system design appeared often but correlated less with realized harm. Grouped together, deployment controls were more critically associated with harm than business impact, which in turn outranked pure technical failure. Sah then turned that lesson into how to test agentic systems. He decomposed risk into four parts, how often the model fails, whether that failure reaches a real decision, whether it goes uncaught, and how severe the consequences are. The same model creates different risk depending on authority, review and monitoring, so evals must cover retrievers, tool failures, permission boundaries and recovery, not just held-out accuracy. Srivastava tied the method to cognitive science, citing Peter Wason and confirmation bias, and the finding that 79% of subjects failed a simple 2-4-6 rule test by trying to confirm rather than falsify, and he argued evals must be failure-driven, a build-break-fix loop that maps where claims break and routes high-materiality use cases away from uncertain regions. They made it concrete with an airline chatbot. A bereaved customer asked about a discounted fare, the bot said buy now and claim the discount within 90 days, the customer acted on it, the airline refused because policy did not allow it, and a tribunal held the airline responsible. Sah said a proper system-level test would have started from the requirement that advice must match approved policy, challenged the bot with refund-after-travel questions, added a policy check or human escalation when it failed, and retested to see if the control caught errors before they reached the customer. That posture looks different from how reliability is usually sold. [incident.io](/startups/incident-io) pitches an on-call platform that unifies incident response with agentic root cause analysis, while a Zenodo replication package for the Embodiment Gap study bundles NHTSA incident reports and openFDA surgical system counts to study real-world harms. Both treat incidents as data to learn from, but Srivastava and Sah push further, to prioritize eval spend by materiality rather than test volume. To be fair, a materiality scale is company-specific and their 98% coverage is retrospective, so it will not predict every new failure. If a founder asked me what to do Monday, I would say stop buying another model benchmark and prove your human review and monitoring actually catch mistakes before customers do. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.