Hiring for judgment when everyone can use the tools

For most of the last decade, hiring an engineer or an analyst meant testing whether they could do the thing. Now the thing is assisted by a tool anyone can reach. The floor has risen for everyone at once, which means the test that used to separate candidates separates almost no one. The ground under the skill is moving fast: the World Economic Forum expects 39% of core skills to change by 2030, and labour-market analytics firm Lightcast finds, on global postings data, that 56% of AI-related jobs now sit outside IT, so the capability has spread far beyond the specialists who used to hold it. Two people can produce the same output. Only one of them knows when the output is wrong.

That’s the shift our clients are absorbing this year. The req still reads like a skills list, but the skills on it are increasingly table stakes. What decides whether a hire compounds or slowly costs you is whether the person has judgment about where the tool belongs, and that’s precisely the thing a skills-based screen was never built to catch. Analyst Josh Bersin frames the winner of this shift as the ‘superworker’, someone AI makes more valuable, and at senior level that added value is judgment, not throughput.

We see it in the funnel. A shortlist of 6 candidates now clears the take-home exercise at a rate that would have been unheard of 3 years ago, because the take-home is the kind of task the tool is best at. The exercise still runs; it just no longer discriminates. Hiring managers read the passing scores as reassurance and then discover, a quarter into the role, that competent output and sound judgment were never the same measurement.

The reason judgment is hard to hire for right now is mechanical. When a model generates the work, the visible quality of that work converges. The strong candidate and the weak one both submit something that reads well, compiles, passes the obvious checks. The variance that used to sit in the artifact has moved somewhere you can’t see it: into the decisions about what to accept, what to rewrite, and what to throw out.

So the artifact stops carrying information. A polished deliverable told you something when producing it was hard; it tells you almost nothing when producing it is cheap. What still carries information is the reasoning behind the deliverable: where the candidate trusted the tool, where they overrode it, and whether they can explain the difference. A generic assessment never asks for that reasoning, which is why it now filters for the wrong thing while feeling rigorous.

It helps to be precise about what judgment even is here. In The Jazz of Physics, the cosmologist and saxophonist Stephon Alexander makes the point that the freest improviser is usually the most rigorously trained; the soloist can depart from the rules because the rules have become so deep they have stopped being visible. Hiring judgment is that same faculty. It reads as instinct and it’s actually earned, the residue of having made the call before and lived with what came back. That’s why you can’t interview for it by asking someone to perform a task. They will simply perform it well. You get at it by working backwards through decisions they have owned. Where did they let a model run unattended, and where did they insist on a human check? What did they get wrong once, and what rule did they build so it wouldn’t happen again? The specificity of those answers is the signal.

The candidates who have it talk about trade-offs. They can describe the case where automating would have been faster and worse, and they can name the cost they were protecting against. The candidates who don’t have it default to capability, what the tool can do, because they’ve never had to own the consequence of it doing the wrong thing. The strongest single indicator we find is a track record of choosing not to automate something and being able to explain why. That decision only comes from someone who has seen the downside up close.

Here’s where the hiring plan goes wrong before the interview even starts. The most common way teams budget for AI is as a discount: the model makes each person faster, so the same output needs fewer of them, so the line goes down. It’s a clean story and it’s expensive to discover late. The model changes what the job is, and the new job wants a different person.

When the routine build is assisted, the value lives in everything around it: specifying the problem precisely enough for the model to help, reviewing what it produces with real scrutiny, and judging how the piece fits a system that has to actually work. That’s a more senior competency. Demand for AI-fluency has risen roughly sevenfold in two years on McKinsey’s (US-centric) reading, and Gartner ranks ‘shaping work in the human-machine era’ among CHROs’ top priorities for 2026, both pointing at the same senior seat. Teams that budgeted AI as a productivity discount missed the seniority shift underneath, and staffed an assisted function as if it were an unassisted one, heavy on producers and light on the judgment layer.

The second-order effect makes the miss worse. When a model can generate plausible work quickly, the danger is too much output that looks right and isn’t. The bottleneck moves from producing to reviewing, and a weak reviewer in front of a fast model is a liability that scales. The 2026 State of AI at Work report (Udacity with Accenture) calls the answer a ‘human premium’, skilled people paired with AI outperform automation alone, which at the hiring layer means weighting the review seat. The cost lands in the budget as a saving and shows up in the field later, where it’s hardest to trace back. Costing the shift as a headcount reduction hollows out the exact seniority the assisted model depends on.

The seniority shift becomes concrete the moment the work meets consequence. In a regulated setting, a pharma GCC running an AI-assisted pilot, say, the tool can almost always do the step. The real question is whether the resulting step can be defended when an auditor asks who was accountable, what was checked, and how you would know if the model was wrong. A workflow that can’t answer those questions passes the demo and fails the audit, later, when the cost is far higher.

Every pilot that survives has one person who owns that line: senior enough to decide which steps the tool may touch, which it may only assist, and which it stays out of entirely, and credible enough that quality and regulatory functions accept the call. This is the judgment call that decides whether a pilot survives audit, and who owns it is a staffing decision, not a technical one. It’s a distinct seat that sits between the data scientist who built the model and the QA generalist, and the centres that skip it are the ones scaling a fast pilot that won’t survive contact with an inspector.

The practical move is to build your assessment around a real decision with a real cost, not a sandbox exercise. Give the candidate an ambiguous situation where the tool would help and also mislead, and watch where they place the human. What matters is whether they can reason about where trust belongs and hold the line under pressure to ship.

For the senior seat, weight consequence carried over tools mastered. Someone who has stood behind a call in front of an auditor, or owned a failure that had real downstream cost, understands the boundary in a way no tool training conveys. The fluency can be built on top of that instinct; the instinct rarely gets built in a quarter. Hire the boundary-owner before scale.

Run the process this way and it rewards the thing that’s genuinely scarce. In a market where everyone on the shortlist can use the tool, that scarcity is the edge worth hiring for, and a generic screen filters it out while telling you it did its job.