When Machines Do the Work, Judgment Becomes the Job

Anthropic, the company behind the Claude AI models, has published an unusual study of its own workforce. In How employment is changing at Anthropic, four of its economists use internal data on hiring, output, calendars and employee surveys to show what work looks like inside an organization built around AI agents. They argue it could be an early signal of where the wider economy is heading.

Some of the findings are striking. The share of junior hires in Anthropic's technical departments fell from 27% to 9%. In an April survey, 20% of employees described their role as shifting from doing the work to managing AI agents. And in interview scorecards, hiring managers increasingly commented on candidates' autonomy and judgment, while discussion of domain expertise shrank.

It is worth remembering that this is one company, at the frontier, studying itself. But the direction it describes matches what I hear from leaders in very different industries.

The Work That Remains Is Personal

Most coverage of reports like this focuses on the jobs that disappear. I'm more interested in what the hiring managers started looking for.

When AI can write the code, draft the contract and log the sales call, execution stops being the main thing that separates one professional from another. What's left is judgment: deciding what matters, what to trust, what to stop and what to defend. The report even describes human attention and judgment as the bottleneck the whole system strains against.

Judgment isn't a technical skill you can download. It is shaped by who you are: your values, your experiences, the problems you have learned to see that others miss. Two people with the same tools will make different calls because they are different people. As execution becomes abundant, that difference becomes the asset. I made a related case in AI Skills Are Becoming a Commodity. Knowing the Work Never Will.

Where Judgment Gets Earned

Here is the tension the report names honestly. Its authors note that in some roles, employees are increasingly removed from the hands-on work that originally built their skills. And the junior positions where people traditionally learned their craft are shrinking.

That should worry every leader, because judgment has to come from somewhere. People develop it through repetitions: making calls, getting some wrong, seeing the consequences and adjusting. Junior roles have always been where those first reps happened. If organizations hire for judgment while removing the places where judgment is formed, they will eventually run short of the very thing they are paying for.

This is where I think conviction belongs in the conversation. I describe it as the earned capacity to act on what you believe before the outcome is certain. The word that matters is earned. A professional who supervises a dozen AI agents still has to decide which output to trust and which to send back, often without full information. That is a conviction skill. It grows only when people are given real decisions and the room to own them.

Leading Teams That Lead Agents

The report offers a few clues about how to do this well. Anthropic says it built an early warning system to spot roles at risk, prioritized moving people internally as their work changed, and found employees moving across departments far more often than a year earlier. It also found that meeting time went up, not down, because the hardest problems still required people to debate and decide together.

That last finding is easy to overlook. Even in the most automated workplace in the study, human conversation became more important. Agents added to the work. They didn't replace the need for people to understand each other.

For leaders, a few questions are worth asking this week.

Where in your organization do early career people still get to make real decisions and learn from the results? If AI has absorbed those tasks, what has replaced them?

When you hire for judgment, what do you actually look for? Can you describe it, or are you still screening for credentials and hoping judgment comes along?

As your people shift from doing the work to directing it, are you helping them understand their own strengths, so they know which calls they are best suited to make?

And for individual professionals: which decisions in your work could only you have made, because of what you've lived and what you notice? Those are worth naming, because they are the part of your value no tool replicates. I wrote about protecting those reps in When AI Does Your Thinking, You Lose the Reps That Build Conviction.

Anthropic's report is candid that none of its practices will prevent disruption, and that protecting current employees does nothing for people who are never hired. That honesty is useful. The future it sketches isn't one where people matter less. It's one where the parts of people that can't be standardized matter more, and where leaders must deliberately create the conditions for those parts to grow.

When machines take on the doing, the deciding becomes the job, and good decisions come from people who know who they are.

Want to explore these ideas further? Learn more about my work on leadership, identity and conviction at www.theglennllopis.com.