Interviewing for Judgement When Agents Do the Grunt Work

A candidate can produce a clean first draft of almost anything now, and so can everyone else. Watching someone produce one tells you very little about whether you want to hire them.

The skill worth testing is review. Give the candidate a piece of work that is fluent, well organised and wrong in a specific way, and watch whether they catch it, how long it takes, and what they do next. That single exercise tells you more than an hour of questions about how they would approach a problem.

Why review became the scarce part

When a draft took two days, most of the value was in producing it. When a draft takes minutes, the value moves to deciding whether it is right.

Confident wrong output is a particular hazard because it has none of the usual warning signs. Bad work by a tired person usually reads as bad work. A wrong answer from a capable tool reads as a good answer. The formatting is neat, the tone is assured, the structure is sensible, and the third assumption is false.

Someone with judgement checks the assumption. Someone without judgement checks the spelling.

The exercise

Prepare one artefact relevant to the role. A short technical design, a client email, a data summary, a policy note. A status update drafted by an agent works well for a manager role. Make it genuinely good in every respect except one, and make the flaw the kind that matters:

• A calculation that is arithmetically fine but uses the wrong base figure
• A recommendation that is reasonable but ignores a constraint stated in the brief, such as a plan that hands next week to someone already on approved leave
• A summary that describes a trend correctly and attributes it to the wrong cause
• A confident claim about a rule or a regulation that would need checking

Give them the brief and the artefact. Ask them to tell you whether they would send it, and why.

Then stay quiet. The pause is part of the test.

What to watch for

Do they check anything, or do they respond to the writing? Strong candidates go back to the brief within the first minute. Weaker ones start editing sentences.

Do they find the specific flaw or only sense that something is off? Both are worth something. "This feels wrong somewhere" is a real signal, but it is not the same as pointing at the line.

What do they do with uncertainty? The best answer to a claim they cannot verify is usually not to guess. It is to say what they would check and where. Someone who confidently rubber stamps an unverifiable claim will do exactly the same thing on your work.

Do they distinguish taste from correctness? A candidate who spends the whole exercise on tone and structure while a factual error sits untouched has told you what they will prioritise.

Do they say no? Some people will not say "I would not send this" to an interviewer. That is worth knowing before you hire them into a review role.

Follow up questions that earn their place

After the exercise, a few questions go deeper without turning into hypotheticals.

• "Tell me about a time you approved something that turned out to be wrong. What did you miss?" The useful part is whether they describe a process change or just say they were more careful afterwards.
• "What would you have to see to trust this without checking it?" This gets at whether they think about verification systematically or case by case.
• "How would you check this if you had five minutes instead of thirty?" Reviewing well under time pressure is the actual job. Everyone reviews well with unlimited time.
• "What is the worst thing that happens if this is wrong and we send it?" Judgement is partly about matching scrutiny to stakes.

Do not test for suspicion

There is a failure mode on the other side. Someone who distrusts every draft, rewrites everything from scratch and refuses to accept help is not exercising judgement either. They are just slow.

The behaviour you want is proportionate. Heavy checking where the cost of being wrong is high, light checking where it is not, and a clear sense of which is which. If a candidate spends twenty minutes verifying a low stakes internal note, that is a finding too.

Why this matters operationally

If agents are doing first drafts inside your business, someone is approving that output before it reaches a customer, a regulator or a board. That approval step is where quality actually lives, and it is a human job. Deciding which outputs need that signature is a separate piece of work.

SicherOne is built on that assumption. Agents work with full context because project management, HR and the agents sit on one set of records, and a human approves agent output before it ships. The approval is not a formality in the design, it is the point of it.

Which brings the hiring question round to something concrete. You are not looking for people who can produce work faster than a machine. You are looking for the people whose name you would be comfortable seeing on the approval.