Where Automation Runs Backward

In February, Jamie Dimon told investors that JPMorgan had already displaced some workers with AI and was trying to move them into other roles. A month later, he warned that this wave of disruption could move faster than past technological shifts. David Solomon, at Goldman, was less apocalyptic. He didn’t dismiss the risk, but he wasn’t ready to call it a jobs collapse either.

The disagreement is mostly about timing. Dimon is suggesting the shift is already inside the workforce. Solomon is more cautious, especially on the apocalypse framing. Both are reacting to the same pressure.

The evidence is more uneven than the headlines suggest.

A recent arXiv paper by Rudra Jadhav and Janhavi Danve gets at the issue from a more practical angle. Jobs are bundles of skills, and not all of those skills are equally exposed to AI. So they built a Skill Automation Feasibility Index, tested four frontier models across 263 tasks covering all 35 skills in the Department of Labor’s O*NET taxonomy, then compared the results with observed usage from the Anthropic Economic Index.

That’s the gap the paper is trying to quantify. A model can be strong on a task and still weak against the job, because jobs aren’t just collections of prompts; they come with context, judgment, handoffs, and accountability.

Math and programming scored highest on automation feasibility, at 73.2 and 71.8. Active listening and reading comprehension scored lowest, at 42.2 and 45.5. That part isn’t surprising.

Then the scores bend in the wrong direction.

The jobs most exposed to AI often depend on the skills where the models scored worst. Jadhav and Danve call this a capability-demand inversion.

This is the inversion. The better defined tasks are often the easiest ones for the model to take on. The rest of the role is less tidy. It depends on context, tradeoffs, handoffs, and accountability.

The usage data fits that read. In the Anthropic data, 78.7% of AI interactions were augmentation rather than automation. People are using the models to help with particular aspects of work more often than to hand work over entirely.

The comparison across models is also telling. All four frontier models produced nearly identical skill profiles, with only a 3.6-point spread, so it’s not really a provider problem. A better model may improve the experience, but it doesn’t appear to change the basic line between the tasks AI can absorb and the work that still needs a person.

There are caveats here. This is one preprint from a small group, and the authors are careful about its limits. Their index measures model performance on text-based representations of skills. It doesn’t measure the full execution of a job inside an organization, with incomplete information, social context, risk, judgment, accountability, and all the handoffs that make real work difficult to package.

So I wouldn’t overread the precise numbers. The pattern is the point.

It also lines up with sturdier labor-market analysis. Anthropic’s own economists, reading Current Population Survey data through late 2025, found no detectable rise in unemployment for workers in the most AI-exposed occupations since ChatGPT launched. Claude appears to cover roughly a third of the tasks it could theoretically handle in computer and math roles.

There may be one early exception. Hiring of workers aged 22 to 25 into exposed occupations has slowed relative to 2022. The effect is modest, barely significant, and doesn’t appear for older workers. If AI is touching employment right now, it may be thinning the bottom rung before it clears the building.

For leaders, the takeaway is this… The near-term return comes from extending skilled workers, changing the shape of their work, and shifting the cleanest tasks into the system. That may reshape staffing over time, especially at the entry level, but headcount plans built around imminent replacement are running ahead of the evidence.

That’s the question the inversion highlights. If AI is strongest where the work is easiest to isolate, what are you really automating? And who still owns the work the role was built around?

📄 https://arxiv.org/abs/2604.06906

Algorithm & Blues publishes Sundays.