AI leaves a judgment gap that experienced workers must fill

Where AI falls short

Across 34 of the 37 countries in a Pew Research Center survey, people expect AI to reduce the number of jobs. In wealthy countries, including the United States, Australia and South Korea, roughly seven in ten expect employment to decline over the next 20 years. Anxiety is also rising among young people in many countries.

But the number of jobs is only part of the question. The nature of the work that remains matters too.

AI capabilities are uneven: a system can outperform people at some tasks and fail at others, without a clear boundary between the two. Models improve quickly when answers can be checked automatically and training examples are plentiful. They are less reliable when there is no automated way to verify the answer. And even when a system gets most answers right, it often gives no signal when it has made a mistake.

Some errors will disappear as models improve. But, Shashanka says, the unevenness is also tied to how these systems are trained. A smarter model still needs the right answer to learn from. Fix one gap, and another may surface—one nobody thought to check.

Many business decisions have no obvious right answer. They depend on an unknown future and on data that does not yet exist. AI can take on more work where outcomes are verifiable; where people must decide what counts as a good outcome, human judgment remains central.

The work around AI

New roles are emerging around systems that act on people’s behalf. Some specialists test AI in real-world conditions rather than relying only on benchmarks. Others set limits on what systems are allowed to do. A third group manages multiple AI agents, much like a shift supervisor coordinating a team.

The model is closer to air traffic control than to watching every screen by hand: radar and software track aircraft, while people assess situations that could lead to a collision. Human oversight is not a matter of monitoring every action. That would not scale.

Setting the limits is harder than it sounds. In a 2026 study, an AI agent was told to protect a secret. It responded by destroying its own email server and deleting all the messages. The system had a goal but no boundaries on acceptable actions. Researchers described the result as “correct values coupled with extremely poor judgment.”

Deciding what to forbid—and what might go wrong—requires someone who understands the work and is accountable for the outcome.

Expertise is not the same as coding

The people best equipped for these roles may not be programmers. They may be the people who know how a job works in practice. Catching what a confident system has missed or distorted requires knowing what a complete answer should contain. That knowledge is difficult to automate or capture in a set of instructions; it develops over years of seeing how work fails.

Ford recently offered a concrete example. After years of automating quality control and cutting thousands of jobs, the company rehired 350 experienced engineers early this year. Their job was to monitor the process and identify failure points before parts reached the factory. They brought years of experience observing breakdowns. With people leading the process, Ford ranked first among mass-market automotive brands in quality for the first time since 2010.

I think that makes the more consequential question less about whether AI can do a task than about who will know when it has done it badly. Companies can automate routine work and still need human judgment. But if they eliminate entry-level roles, they may also eliminate the route by which employees develop the experience that judgment depends on.

The disappearing path to judgment

Junior employees learn under experienced colleagues. Over time, repeated exposure to the work turns into intuition. AI can remove some of the difficulty involved in learning a job, but it does not remove the need for judgment. It may instead remove the work through which people acquire it.

Shashanka argues that hard-won experience will become more valuable in the coming years. Yet the traditional path—learning from senior colleagues while doing routine work—may no longer reliably provide it. New workers will need to build that experience deliberately, on tasks machines still cannot evaluate.

Madhu Shashanka is co-founder, chief scientist and chief technology officer of Concentric AI, a data security software company.