133. Beyond the Blunt Instrument: Why “It Hallucinated” Isn't an Answer

What happens when an AI agent fails, and how can enterprises prove they saw it coming? On this episode of TSIA’s TECHtonic, Thomas Lah takes on one of the most important questions facing organizations moving AI from experimentation into the enterprise: How do you prove that AI is actually delivering the outcomes you promised? Drawing on TSIA’s Five Proofs of Outcome-Based Revenue, Thomas and guest Sekhar Sarukkai unpack why proof of performance and telemetry are becoming essential to proving business value. Sarukkai, a serial entrepreneur who previously founded Skyhigh Networks and Securent, now leads Chatsee.ai, which recently raised $6.5 million to build what he calls a failure intelligence layer for AI agents.

The conversation takes a revealing look at what really goes wrong when AI agents enter the real world. After analyzing 10,000 enterprise agent failures, Chatsee identified 157 distinct failure categories, and found that hallucinations account for less than 10% of actual failures. Instead, enterprises are facing bigger and often invisible challenges around resolution, escalation, silent execution, and the growing gap between pre-deployment controls and runtime governance. Thomas and Sekhar also unpack the hidden economics of AI failure, including how a seemingly minor error can quietly spread through downstream systems for weeks. Sekhar introduces a framework for measuring direct loss, propagation, detection delay, and reversibility, while making the case for shared accountability across enterprises, AI platforms, and integrators.

If your organization is serious about moving AI agents into production, and proving the value they deliver, this is a conversation you’ll want to hear.