The judgment shortage

Every plan for adopting AI starts with the same quiet assumption: the work is a fixed pile, machines will take some of it, and people will be left with less to do. It is the assumption underneath the layoff headlines and underneath most boardroom AI strategies.

Then you watch what actually happens when a team automates in earnest, and the assumption falls apart. Talk to anyone genuinely running AI inside their work, not piloting it but running it, and you hear the same report: they have never had more to do. The pile was not fixed. Automating it made it bigger.

This is not a temporary glitch that the next model release will iron out. It is a structural effect, and businesses that understand it will make very different, and much better, decisions about AI than businesses that do not.

Cheap production, expensive selection

When something becomes drastically cheaper, people do not consume the same amount and pocket the savings. They consume enormously more. It happened with computing, with photography, with publishing. It is now happening with the production of knowledge work itself.

A model can produce a draft, a report, a design variation, a working prototype, a plausible answer to a support ticket, in seconds and at negligible cost. So people produce many. Ten options where there used to be one. A first version of everything, immediately. Volume stops being the constraint anywhere a model can operate.

But every one of those cheap artifacts creates an expensive question: is it right, is it good, and should we act on it? A draft demands an editor. A generated pull request demands a reviewer who understands the system it touches. Ten design options demand someone with the taste to pick one and the authority to kill nine. A confident answer demands someone who can tell confident from correct.

Production got automated. Selection did not. And selection, judging and verifying and choosing and deciding what happens next, is the part that was always scarce. The bottleneck in knowledge work has not disappeared. It has migrated from making things to judging things. Most organisations are still staffed, structured and measured for the old bottleneck.

Why the machines do not just judge, too

The obvious rebuttal: models keep improving, so surely judgment gets automated next.

Some of it does. The mechanical layer, the checks that can be written down, does it compile, does it match the style guide, does it contradict the policy document, is being automated already, and should be. But two parts of judgment sit beyond that layer, and they are not incidental. They are structural.

First: a model’s output is shaped by what the world has already done. Ask it for a proposal and you get the shape of everyone’s proposal. Competence, instantly, at the market average. When your competitors use the same tools, average is free for them too, and free for everyone is worth nothing to anyone. The value concentrates in whatever is specific: your constraints, your customers, this quarter, the thing your industry has not figured out yet. Specificity is exactly what cannot be in the training data, because it is happening now, to you.

Second: judgment is accountability, and accountability does not automate. Someone signs off. Someone’s name is on the decision when the customer is angry or the number is wrong. You can delegate the drafting to a machine. You cannot delegate the owning. Every workflow that includes an irreversible step, money moves, the message sends, the product ships, has a human-shaped slot in it. Not because machines are too weak, but because responsibility is a relationship between people.

So the equilibrium is not machines doing everything. It is machines producing at enormous volume, and humans deciding, with better instruments and higher stakes, what is true, what is good and what happens next. More automation means more output crossing more judgment points, which means judgment becomes the thing you plan around.

What this means if you run a business

Almost every failed AI initiative we are watching in the market fails the same way: the company bought production capacity and never rebuilt the judgment capacity around it. Licences were purchased. Output went up. Nothing measurable improved, because drafts piled up in front of the same reviewers, the same approval chains, the same one experienced person who was already the bottleneck before the machines arrived.

If the bottleneck has moved, the redesign has to move with it. Practically:

Find the judgment points before you automate anything. Trace the workflow and mark where a human verifies, chooses or approves. Automation upstream of a judgment point multiplies the load on it. If you do not strengthen that point with clearer standards, better instrumentation and explicit ownership, you have not sped up the process. You have moved the queue.

Make the standards explicit. The reason review is slow in most companies is that what counts as good lives in one person’s head. Writing it down was optional when output was scarce. At machine volume, unwritten standards are a denial-of-service attack on your best people. What can be made explicit can be partly automated. What stays tacit is precisely where your senior people should spend their day.

Measure the decision, not the draft. Output metrics go up automatically now. They are vanity by default. The numbers that matter sit at the judgment points: how fast a decision gets made, how often it is right, how far a mistake travels before it is caught.

Buy less, redesign more. A tool purchase changes the cost of production. Only redesign changes where the judgment sits, who holds it, and how it is equipped. That is the actual work of AI transformation, and it is organisational at least as much as it is technical.

The shortage is the opportunity

None of this is bad news. A shortage of judgment is another way of saying that human judgment has never been worth more per unit. The companies that treat AI as a headcount story will spend the next few years drowning their best people in cheap output. The ones that treat it as a redesign story, machines on volume, people on judgment, standards written down, decisions instrumented, will move at a pace that looks unreasonable from outside.

The work does not end after automation. It concentrates. Right where it always mattered most.