Leader-Level Judgement and AI

Leader-Level Judgement, three stacked layers: Context-aware (top), Accountable (middle), Durable (base). A parallel column labelled Model recommendation shows the AI output entering at the middle — the leader’s call passes through all three layers; the model’s recommendation has none.

​

Accountable. The leader signs, owns what follows, and faces the people affected. AI does not sign anything. The model produces a recommendation; the recommendation has no skin in the consequence. A leader who outsources the call to AI’s framing has not transferred accountability, only the thinking. The accountability stays where it was, but the basis for defending the call has been hollowed out.

​

Durable. A call worth making will be tested. The first test usually arrives within days, often as a contrary AI synthesis prompted slightly differently. Durability is the habit of distinguishing real new information from a re-shuffled framing of the same information, so the call survives the second-day pushback that would otherwise reverse it on no new evidence. Not stubbornness; the temperament that holds across the test period and updates only on what warrants an update.

​

A judgement missing any of the three is something weaker: without context it is generic, without accountability it is advice, without durability it is a mood.

​

Genesis

Stacks Michael Polanyi (The Tacit Dimension, 1966) on what the leader knows but cannot fully tell; Donald Schön (The Reflective Practitioner, 1983) on reflection-in-action as the live faculty of skilled professionals; Phil Tetlock (Superforecasting, 2015) on the temperament that holds framings provisionally, updates honestly, and resists both over-claim and silence; and Andrew Likierman (Judgement at Work, Profile Books, 2025) on judgement as a learnable process with six elements: knowledge and experience, awareness, trust, feelings and beliefs, choice, delivery. Likierman’s reframe, process, not trait, situates the three components above as the AI-era specialisation of his framework: a teachable process, where the human remains in control. Likierman also extends judgement-as-discipline beyond the moment of choice to the auditable evaluation of judgement quality given what was known at the time, not on whether the outcome was good, the basis on which durability can be assessed honestly.

​

Why it matters now

Five forces pull the leader to outsource judgement to AI. The output is well-formed, raising the cognitive cost of disagreement. The model can argue both sides equally well, leaving the leader without ground to stand on. Defending an independent call is socially expensive in ways following the analysis is not. The framing borrows the leader’s authority once internalised, so delegation slides into abdication without the leader noticing. And AI fluency manufactures overconfidence, the bias Likierman, citing Kahneman (Thinking, Fast and Slow, 2011), names as the most dangerous in leaders, because it suppresses knowledge-gathering, awareness, trust, and the consideration of alternatives at once. None of the five is weakness; they are built into how the tool presents output and how organisations evaluate decisions.

​

The moves

Surface the context AI cannot see before reading the model’s recommendation: two paragraphs of tacit knowledge committed to writing, so the comparison is honest. Sign before you publish: a one-line statement of the call in the leader’s own words, written before the document goes to anyone else; if the model’s recommendation does not match it, the question is which is closer to what the leader actually believes. Hold the call across the test period: write down at the point of decision what kind of new information would warrant a change; updates that pass the list are real updates; updates that do not are noise.

​

How ChangeSchool applies it with executives

We run senior cohorts through the Pre-AI Note exercise: each leader brings a current consequential decision and writes the tacit-knowledge paragraph before AI is opened on the question. The before-and-after comparison reliably surfaces context that would otherwise have been retrofitted to the model’s framing.

​

The discipline

A pre-AI note on consequential calls: ten minutes with a notebook before opening AI on any signed-and-stood-behind decision. A signed-call log kept weekly, two lines per call recording the basis on which it was signed; if every entry reads because the analysis recommended it, the model is making the calls. A post-test review once a quarter, looking back at calls signed three to six months earlier; the only way the leader finds out, honestly, what their judgement is currently worth before a high-stakes decision tests it the hard way.

​

​

​

‘The model produces the recommendation; the leader produces the call.’ 

 

 

Viren Lall, Managing Director,

ChangeSchool LDN (2026).

virenlall.com/leader-judgement-ai

​

​

​