The board approving AI without knowing what it was approving
- Problem
- The company had eleven AI initiatives in production and no way to answer, in a board meeting, who was accountable for each. Every approval was debated on projected return; none on what would happen if the model got it wrong with a real customer.
- Involvement
- Introduced a simple classification criterion based on consequence, not technology, as a prerequisite for any approval. Each system began reaching the board with a named owner, an explicit failure scenario, and a contestation path for the affected customer.
- Impact
- Two of the eleven initiatives were suspended for having no identifiable owner. The remaining nine moved to quarterly review with human-override data. Average approval time went down, because the agenda arrived ready.
The starting point wasn't a technology diagnostic but a read of the minutes from six prior meetings. Not one contained a question about consequence, only about return and timeline.
Consequence-based classification split the systems into three bands: no direct effect on people, indirect effect, and direct effect on a customer's or employee's life. Only the third band required board approval; the rest moved down to the executive committee. That unclogged the agenda and raised rigor where it mattered.
Initial resistance came from the technology team, which read the change as distrust. It reversed once it became clear that the named owner also gained formal veto authority, which in practice increased the team's power, not the opposite.
The metric that most changed the conversation was the human-override rate. On one system it had been zero for fourteen months. Not because the system was good: because nobody knew they could override it.