Judgment under uncertainty is the ability to make a sound decision when the facts are incomplete, ambiguous or in conflict. It means weighing what you know, what you do not, the cost of being wrong in each direction, and whether waiting or asking would change the answer. It is the core of most review work around AI agents.
Why does judgment under uncertainty matter in AI agent work?
Agents handle the cases where the facts are clear. What reaches a person is what remains: the claim with a photo that does not quite match, the customer whose income evidence is three months old, the request that is probably genuine but arrived by an unusual route. Rules help, but rarely settle these on their own. Someone has to decide, and decide well.
Employers find this hard to hire for, because it does not show up in a list of qualifications. The case is made in judgment is the AI skill employers cannot find. That makes it valuable to anyone who can show it.
How does it work in practice?
Good judgment under uncertainty tends to follow a short sequence:
- Name the uncertainty. What exactly is unknown? "I cannot tell whether this payslip is current" is workable; "something feels off" is not, yet.
- Weigh the costs of error. What happens if you approve and are wrong, and if you refuse and are wrong? They are rarely equal.
- Ask whether more information is available, and at what cost. A phone call that takes five minutes is usually worth it on a large decision.
- Decide, escalate or hold. If the decision is yours and the evidence is good enough for the stakes, decide. If not, route it, which is where escalation judgment comes in.
- Record the reason, including what you did not know.
What is an example?
A lending operations analyst we will call Theo reviews an agent's recommendation to approve a personal loan of 8,000. The agent's affordability check passed, but the customer's bank statements show regular payments to a lender that do not appear on the credit file. That could be a loan not yet reported, or a payment for something else. The cost of approving wrongly is a customer in debt they cannot afford; the cost of refusing wrongly is a lost customer and a complaint. Theo holds the application and asks the customer about the payments. The answer, a recent car loan, changes the affordability result. Lending employers look for exactly this habit, as set out in which AI skills consumer lending employers value most.
How is it different from escalation judgment or under-trust?
Escalation judgment is one part of judgment under uncertainty: deciding whether the call is yours to make. Judgment under uncertainty is wider: it includes making the call yourself when it is yours.
Under-trust is a failure of judgment under uncertainty: rejecting or redoing an agent's work by default, because uncertainty feels uncomfortable, even when the evidence supports the draft. It is explained in what under-trust in AI is. Good judgment avoids both over-trust and under-trust by looking at the evidence for each case.
Which terms are related?
- Risk appetite: how much uncertainty a business is willing to accept for a given action.
- Decision rationale: the written reason, which should state what was uncertain.
- AI literacy: the baseline understanding of AI that the EU AI Act asks businesses to ensure in staff who deal with it, explained in who needs AI literacy under the EU AI Act.
For the full set of skills employers test, read AI skills for job seekers.
Questions people ask
How do I show judgment under uncertainty in an interview?
Pick a case where the facts were incomplete. Say what was unknown, how you weighed the cost of being wrong each way, what you did to reduce the uncertainty, and what you decided. Interviewers care more about the reasoning than the outcome.
Is it bad to be uncertain when reviewing an agent's work?
No. Uncertainty is normal; acting as if it is not there is the problem. Naming it, and recording it in your reason, is part of doing the job well.