The judgment reviewBoard oversight
Every AI framework watches the machine. None asks if you could still tell it was wrong.
The productivity case is sound and it arrives now. Your AI programme is probably well governed as a machine — inventoried, validated, monitored.
Take it to a meeting
Four questions a director can put to management, and what the fluent answer leaves out. No engagement required.
Judgment is deposited by repetition
The drafts, the rounds, the reconciliations — until a person can feel something is off before they can say why.
The machine absorbs that work first
The junior pass is the cheapest thing to automate. It is also the thing most worth having done.
The saving books this quarter
What it draws down appears years later, on no number at all. That stock is judgment capital.
What a board should ask about AI and judgment
01 / The gate — ask it of the benefits paper, in the meeting
“This shows the hours removed. Which of those hours were somebody’s training?”
Does the paper that approved this deployment name the work being displaced, and record who assessed what that work was forming?
- A good answer
- The approval paper names the displaced work and records who assessed what it was forming. Some boards can produce this today.
- A fluent non-answer
- A benefits case, a headcount saving, and reskilling somewhere in the transformation plan.
Owner Business case signer · Hosted by Investment gate
Before you act on any of it
What is actually being lost
Some of what the machine absorbs was forming judgment. Some was coordination overhead that built up because coordination was expensive. No instrument an HR function runs today separates them.
The layer Was that layer forming judgment, or navigating a bureaucracy?
Collins and Evans separated these in 2002: the expertise to do the work, and the expertise to know who can. The second is real, and it is the part cheap coordination replaces. Citigroup told its shareholders it was going from thirteen management layers to a median of eight.
One question separates them: did the work come back marked? Work that returns honest, structured feedback builds judgment. Work whose feedback is mush builds only confidence.
You are not being asked to adopt a new risk
Most of these duties already exist. What is new is that nobody measures whether they are being met.
AI Act Article 14
High-risk systems must be designed so overseers stay aware of the tendency to over-rely on outputs. Automation bias, named in the regulation as a design requirement.
Indicative and jurisdiction-specific — confirm against the current consolidated text before it goes in a board paperWhere institutions stand
Rung one is the absence of an instrument
No owner, no measurement, no line. Most boards are here.
An owner, a committee, one indicator in front of the board.
A threshold set in advance that has fired, with an action recorded.
An independent check, and a drill passed within twelve months.
Formation running faster than before. Nobody evidences this yet.
If you want to know where you stand
Ten questions. No email.
The framework sets out every instrument behind these four seats, and the board papers carry the questions in a form you can circulate before a meeting.