Situational judgement is the third of the six orchestration capabilities. When AI acts with some autonomy, the key human contribution is knowing when to step in.
The boundary keeps moving. Stanford's 2026 AI Index found AI agents' success on a benchmark of real-world tasks rose from 20% to 77.3% in a year.
Tasks that needed close supervision last year may not need it now. New tasks may need more.
Three intervention decisions
Delegate? Is this task safe to hand to AI with light oversight? Weigh complexity, stakes, novelty and the tool's track record.
Intervene? The AI is running and something unexpected happens. Raisch and Krakowski warn of the risks when people hand over too much and stop engaging.
Escalate? The situation exceeds both the AI's authority and yours. Know who to refer it to and what they need to know.
There is a vigilance paradox. The more reliable a system becomes, the harder it is to stay alert for rare failures. Aviation safety research has long documented this.
Judgement grows from experience
Gary Klein's research on naturalistic decision-making shows experts recognise patterns. They notice when something does not fit, often before they can say why.
Agrawal, Gans and Goldfarb argue that as AI supplies predictions, human judgement becomes more valuable.
That judgement needs practice. Ferdman argues in AI & Society that AI deskilling is a structural problem. If AI does all the routine work, people lose the experience that builds pattern recognition.
Review cases together. Discuss the calls people made and what they noticed.
Build this capability in your organisation
Growth Performance designs bespoke programmes that turn insight into measurable behaviour change.
The Exception Triage Framework
Four categories help people decide what to do:
- Green: routine variance. Slightly off but acceptable. Correct it, log it and move on.
- Amber: significant deviation. Materially wrong or outside normal limits. Intervene, correct and record it for review.
- Red: critical exception. A risk to people, compliance, reputation or safety. Override at once, escalate and review formally.
- Grey: uncertain. Unsure whether to act. Pause, seek input and favour human oversight.
Organisations should set who may override AI at each stage and what gets recorded. Feed exceptions back into how the system is designed.
Pre-mortems for AI workflows
Before launching an AI-supported workflow, imagine it has failed. Work backwards to the likely causes.
Ask three questions. What input might the AI misread? When might it be confidently wrong? Where would an error go unnoticed longest?
Designing checkpoints into the work is part of Adaptation.
Try this
Pick your most important AI-supported workflow. Run a 20-minute pre-mortem with the team. Add one checkpoint where the likeliest failure would surface.
Then take the free Decision Trace Readiness Diagnostic.
