Generative AI matured into a feature. Agentic AI is becoming an operating model. The question for boards has changed — and the failure modes have changed with it.
The board's question
“Where are we using AI?”
Failure mode
A hallucinated answer in a chat window. Embarrassing. Recoverable.
The board's question
“What can we connect AI to?”
Failure mode
A wrong API call inside a sandbox. Contained. Auditable after the fact.
The board's question
“What is AI deciding for us today, and who is accountable?”
Failure mode
An agent loop that drains an account, mis-routes a freight order, or breaches a regulator's rules — at machine speed.
The structural change: we are moving from systems that recommend to systems that commit transactions. Every architectural assumption built for the first model — request/response auth, human-mediated approval, end-of-day reconciliation — has to be revisited.