TLDR: Box CEO Aaron Levie says CEOs can fall into “AI psychosis” because they lack exposure to the last mile of trial and error behind enterprise AI. He argues leaders should use AI heavily to grasp real agent work and governance.
Key Takeaways:
- Levie challenges the CEO distance from operations, where most AI value comes after prototypes meet messy customer environments.
- He links “AI psychosis” to inflated beliefs from polished chatbot outputs and says agents need 10 or 20 downstream steps for results.
- His push for CEOs to use AI aims to close the gap between AI promise and the governance, cost, and long tail work that determines adoption.
Levie is basically saying the problem is not intelligence, it is commuting. If you never walk the last mile where systems break, you will trust the happy path.
Levie is basically saying the problem is not intelligence, it is commuting. If you never walk the last mile where systems break, you will trust the happy path.
Q&A
If executives only see polished AI outputs, what decision traps tend to follow?
They can underfund the hidden work: data readiness, integration, evaluation, monitoring, and governance, then blame the tool when real performance fails.
How might “AI psychosis” show up in enterprise roadmaps, beyond hype meetings?
Teams may set targets based on demo metrics, ignore failure modes, and delay controls for cost, access, and hallucination risk until after deployments.
Why does Levie argue CEOs should use AI more, instead of just hiring experts?
Because leaders still decide priorities and risk tolerance; hands on use helps them ask better questions about downstream steps and constraints.
What does UCSF’s “AI associated psychosis” imply for how companies should frame AI safety internally?
It suggests training and escalation paths should treat user harm as plausible, not theoretical, and that organizations need clear guidance for vulnerable users.
If enterprise AI adoption slows from cost and governance, what would “successful agents” look like in practice?
More measurable workflows with defined scope, human approvals, tracked costs, and ongoing evaluation, rather than broad autonomy promises from prototypes.
No comments yet. Be the first to share your thoughts!