NORBERT BIEDRZYCKI
5 min readBusiness, SocietyRead this in Polish

Where will judgment come from in ten years?

“We know more than we can tell.” Michael Polanyi

The companies that have gone furthest in working with AI agents now talk about one resource that cannot be bought in tokens. They call it different things: taste, judgment, intuition. And these very companies, not entirely consciously, are sawing off the branch on which that resource grows.

The last resource

Jacob Sansbury, founder of the startup Spawn, described the moment when 700 items were waiting in the queue for his decision. The agents were working faster than he could respond. His solution was rather unusual: he gathered two years of his earlier decisions and built an “AI Judge” from them, an agent that rules on his behalf. Sansbury calls it transferring his taste into the system.

Anthropic, in its October report on how AI has changed work inside the company, writes that managers are increasingly hiring for autonomy and judgment, and less and less for narrow specialisation and speed of execution.

Even Cisco, a company with a forty-year history, entrusted the rewrite of 98% of its legacy code not to its most senior architect but to a twenty-seven-year-old who had earlier founded and sold a startup himself. CEO Jeetu Patel did not buy his knowledge of Cisco’s code. He bought his instinct, developed at his own risk.

All three stories lead to the same conclusion. Execution is becoming cheap. Judgment is becoming expensive.

A paradox few talk about

The problem is that judgment does not come out of nowhere. Michael Polanyi wrote about tacit knowledge, the things we can do but cannot fully describe. The doctor who looks at a patient and knows something is wrong before seeing the test results. The engineer who senses that an architecture won’t hold, even though all the tests pass. This knowledge cannot be read in a textbook. It is acquired through years of doing things with your own hands and getting them wrong.

Yet it is precisely those years that are disappearing. At Anthropic, the share of juniors among newly hired technical staff fell from 27% to 9%. The authors of the report soberly note that protecting the jobs of current employees will not help those nobody hired. Stanford researchers showed as early as August 2025 that in occupations most exposed to AI, employment of people aged 22 to 25 fell by a relative 13%. In Poland, according to No Fluff Jobs’ report on the IT job market, offers open to people at the start of their careers make up a few percent, whereas only a few years ago they were a significant part of the market.

Part of that decline is, of course, a correction after the pandemic, when companies over-hired. But the direction is clear. The work on which juniors learned their trade is being taken over by agents.

Three roads to judgment

When I think about where judgment came from in the people I have worked with over the years, I see three sources.

Repetition. Hundreds of similar tasks, after which you start to see patterns. This is the first thing agents take over, because repetition is exactly what they do best.

Mistakes. An error that cost something, and the memory of that cost. Agents make mistakes too, but it is not a human who pays for them or remembers them. The lesson stays in the log, not in someone’s head.

Accountability. The moment you put your name to a decision and then have to explain it to the board, the client or the team. This is the only source agents are not taking away, for now. And perhaps it is exactly where we should start designing a new development path.

A new path

The Cisco story offers an interesting hint. Krishnamurthy did not acquire judgment by climbing the corporate ladder. He acquired it by running his own company, that is, by taking responsibility early and at his own expense.

The Anthropic report shows something similar from the other side. Non-technical employees who previously did not write code at all went from two programming sessions a month to several dozen. Work increasingly consists of managing agents, reviewing their output and deciding what makes sense.

So perhaps a young person today should not start by writing simple code that an agent will write anyway, but by running a small team of agents under the eye of someone experienced. With real, if limited, accountability from day one. That requires a conscious decision from companies, because in the short term it is always cheaper not to hire anyone.

What even the best memory won’t replace

At Sophona we build AI Operators with context memory. An Operator knows how the organisation made similar decisions before, what the exceptions were and when a case must be handed to a human. In a sense it stores company knowledge that until now lived only in the heads of a few experienced people.

But we should be honest about the limit. An Operator’s memory stores how the organisation decided. It does not store how a person learned to decide. Sansbury’s AI Judge works because Sansbury spent years making decisions himself. His successor, if he never has to decide, will have nothing to pass on to the next system.

The dark company

I increasingly hear about the vision of a “dark company”, by analogy with the dark factory, where robots work with the lights off because they don’t need to see. An organisation in which agents do everything and people only look in from time to time.

It is worth remembering that even in the darkest factory someone had to design it, someone decides what it should produce, and someone answers for it when the line stops. Those people did not come out of nowhere. Once, they stood at the machines.

If AI-native companies want to have people in ten years whom they can trust with unlimited tokens, they must consciously build places today where those people make mistakes, learn and take responsibility. Judgment is the only resource that cannot be bought in tokens. It has to be raised.

Read next

  1. Who signs the agent's mandate?

    Every company has a policy that says who can sign an invoice for 50,000 zloty. Almost none has a policy that says what an agent deploying changes to production a hundred times an hour is allowed to do. In recent weeks two stories have appeared that are worth putting side by side.

  2. Everyone says that AI is replacing consulting. I think we are missing the main issue.LinkedIn

    In recent months, the debate about consulting has grown. AI is changing everything. Consultants could lose their jobs. Strategy work might be automated. Slides could be made in seconds without any human help. I understand the excitement. AI is truly impressive.

  3. The twilight of the leaders

    Over the past few decades, the group has become the corporate model for accomplishing work, but the perception of leadership has not changed much; they have been (and are still) seen as powerful individuals who form and direct those groups. Even as late as the