Boris Khazin, AI governance and digital risk advisor and author of 'The Robopsychologist's Problem: Governing AI in the Room not the Machine', joins Bob to explore why AI that finds the most efficient path to a goal demands a new approach to governance. Drawing on behavioral economics and decades in financial services, Boris explains why oversight must shift from one-time audits to continuous monitoring, and why controls belong in the environment around AI rather than inside it. They discuss shadow AI, the limits of compliance training in an era of cloned voices, biased training data, and where humans should sit in the loop. Boris closes with three things every risk and compliance leader should do tomorrow morning.
Keywords
Boris Khazin, The Robopsychologist's Problem, AI governance, governance risk and compliance, behavioral economics, game theory, agentic AI, continuous monitoring, guardrails, shadow AI, shadow IT, AI literacy, voice cloning, human in the loop, flash crash, training data bias, incident disclosure, AI regulation, value at risk, Isaac Asimov
Takeaways
Governing AI agents can't be a one-time audit; it requires continuous assessment and monitoring.
AI optimizes for the task it's given without the moral and social pressures that restrain people, so "cheating" is a logical outcome, not intent.
Effective controls live in the environment, not the model: limit access, block development capabilities, and make any exception temporary and approved.
Shadow AI is a signal. Ask why people go around sanctioned tools and give them a non-punitive way to request alternatives.
Training reduces risk but won't stop a convincing cloned voice; processes must let junior staff verify and say no without consequences.
Removing human checkpoints eliminates the pause that once caught errors, so place humans where judgment matters, not on hundreds of routine approvals.
Translating AI risk into value at risk lets leaders make real business decisions instead of reacting to color-coded scores.
Quotes
"Because its goal is to achieve the task. If you're turning it off, it cannot achieve the task. Turning it off is bad."
"Requiring people to disclose events in no way is it gonna nerf the development of AI or anything like that."
"Because look, if two out of a thousand people want a tool, totally understandable. If four hundred out of a thousand people want the tool, something is wrong."
"Make sure your agent doesn't have development capabilities. Because even though you put guardrails on it, it can create an agent that doesn't have those guardrails."
"Cybersecurity, especially around the AI companies, is designed to protect from outside coming in."
"Look, we're gonna sit down and we're gonna figure out what is the value at risk if you don't close this risk. And then it becomes a business decision."
Chapters
00:01 Welcome and introductions
00:21 From behavioral economics to AI governance
04:00 How regulated industries stack up on AI readiness
06:01 Why agents demand continuous governance
09:12 Behavioral economics and AI that cheats the test
15:55 Incident disclosure and the case for regulation
17:27 Inside The Robopsychologist's Problem
21:48 Guardrails, escalation, and the limits of human review
24:41 Shadow AI and asking why people go around tools
30:33 AI literacy, cloned voices, and verification
36:33 Biased training data in regulated industries
42:00 Agent limits, human in the loop, and restoring the pause
48:56 Why AI incidents go unreported
56:38 Value at risk versus red, yellow, and green scores
59:24 Three steps for compliance leaders tomorrow
Boris Khazin: https://www.linkedin.com/in/boris-khazin-0391931/
The Robopsychologist’s Problem: https://www.amazon.com/dp/B0HKD1632C
For AI readiness advisory work and marketing inquiries:
Bob Pulver: https://linkedin.com/in/bobpulver
Elevate Your AIQ: https://elevateyouraiq.com
Substack: https://elevateyouraiq.substack.com
