I’m a CTO. AI still needs your human expertise | Madhu Shashanka

AI is increasingly getting good at doing some tasks, but it can never replace human workers’ judgment

Autonomous cars, which may be the safest we have built, keep getting easily avoidable parking tickets. While Waymos crash far less often than human drivers over tens of millions of miles, they’re still terrible at performing basic tasks – keeping out of disabled parking spots and tow-away zones, or steering clear of a church garage during Sunday service, according to the Wall Street Journal.

That is the state of AI today: impressively capable at something complex, yet embarrassingly clumsy at something easy. And it has implications for what work will remain human and what will be automated by AI.

People expect AI to lead to fewer jobs in 34 of the 37 countries surveyed by the Pew Research Center. In wealthy countries like the US, Australia and South Korea, about seven in 10 expect job losses over the next 20 years. And across countries, concern among young people is rising. While fears focus on job numbers, the bigger change will be what jobs become.

The nature of human work will move towards judgment. Experts refer to AI as “jagged”, superhuman at certain tasks but weak at others, with no clear boundary between the two. AI rapidly improves where work can be automatically checked and examples are endless. But it’s shaky where no automated test can determine whether the answer is right. Even if AI is correct a majority of the time on a task, it often doesn’t tell you when it’s inaccurate. If you’ve used AI, you’ve probably experienced how it can be confidently wrong.

Some of these errors will solve themselves as models get better. But jaggedness is a symptom of how these models learn. I see this often as a technology leader building with these systems.

A smarter model still needs a right answer to learn from. When one gap is closed, it shows up somewhere else no one thought to check. But many business decisions don’t immediately have a clear right answer. Whether a choice is good depends on a future no one can see and data no one has. That’s where judgment comes in. So AI will keep mastering tasks that have checkable results but struggle where a person has to decide what a good outcome looks like.

As a result, a new type of work is starting to appear. Someone tests whether the output of AI is correct based on real work, not just on benchmarks. Someone sets limits on what AI is allowed to do before it runs, so the worst mistakes are avoided. Someone runs and manages several agents the way a shift manager runs a crew. Picture an airport control tower. Radar and software track flight paths, with humans judging the few that matter to avoid collisions. No one is manually watching every screen. That would never scale.

Setting AI’s limits is harder than it sounds. In a 2026 study, after an agent was told to protect a secret, it destroyed its own email server, deleting every single email. It was given a goal, without any limits on what it could do. The researchers described it as the right values paired with terrible judgment. Deciding what is off limits and what a mistake would cost takes someone who knows the work and will answer for the result.

Here is the good news for people worried about their own jobs. The people best suited for doing this new type of work aren’t computer programmers. They are the ones who know how the real work is done. Catching what a confident AI left out or got wrong takes someone who knows the complete answer. That knowledge is uncheckable, unwritten and takes the form of judgment honed over years of doing the task and watching it break.

Ford recently learned the value of human expertise. In recent years, it leaned hard on automating its quality control, cutting thousands of jobs. But earlier this year, it had to rehire 350 veteran engineers to oversee the process and hunt for failure points before a part reached the factory floor. They brought the judgment of watching parts fail over years. With humans at the helm, Ford’s quality ranking then hit first among mainstream brands for the first time since 2010.

But we are making a serious mistake with the jobs we’re eliminating due to AI: the entry-level, routine ones. That routine work is how judgment gets built over time. Junior employees work under the supervision of senior experts until their experience turns into instinct. AI eliminates the friction related to learning how to do the job. But the demand for judgment did not go away. We are just destroying the path to developing it.

The real question is not whether machines will take over our work. It is whether we will have the people who can tell when the machines are wrong. Right now, we are automating the training ground where judgment is learned. If you have hard-earned expertise about anything, the next few years will only make you more valuable. And if you are just starting out, the old apprenticeship model will not hand it to you. You have to go and earn it, deliberately, on tasks the machines cannot yet judge.

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