The gap between what AI can do and what it's actually doing is closing. If your hiring process still optimizes for the implementation layer, you're selecting for the part that's being automated.
If you lead a software team, the way you evaluate and hire developers is shifting. Ignore it, and you'll miss strong people or hire for the wrong things.
This isn't theoretical. Anthropic just released labor market data, and it points to a real change in how we should think about technical talent. 94% of coding tasks could be handled by AI. Only about 30% actually are today. That gap is closing, and it's already changing what a "good developer" looks like.
The numbers
Peter McCrory, Anthropic's head of economics, shared more context in Fortune. Their March 2026 report, "Labor Market Impacts of AI: A New Measure and Early Evidence," introduced a framework called "observed exposure" — combining theoretical LLM capability with real-world usage data from Claude.
The top-line numbers stand out:
| Occupation | Share of tasks AI can perform |
|---|---|
| Computer Programmers | 74.5% |
| Customer Service Representatives | 70.1% |
| Data Entry Keyers | 67.1% |
| Medical Record Specialists | 66.7% |
| Market Research Analysts | 64.8% |
| Sales Representatives | 62.8% |
| Financial & Investment Analysts | 57.2% |
| Software QA Analysts & Testers | 51.9% |
| Information Security Analysts | 48.6% |
| Computer User Support Specialists | 46.8% |
More than 90% of the work done by tech and finance workers could — in theory — be replaced by AI. But the more important story is underneath.
The gap
There hasn't been a clear rise in unemployment for highly exposed roles since late 2022. Adoption in computer and math jobs sits around 33% compared to 94% capability. 30% of workers currently have zero meaningful AI task coverage in the data.
At the same time, job-finding rates for workers aged 22–25 in exposed roles are down 14%. Goldman Sachs estimates around 16,000 U.S. jobs being cut monthly due to AI, with Gen Z feeling it first.
The displacement isn't evenly distributed. It's hitting the youngest workers first — the ones with the least leverage, the smallest networks, and the most to prove.
Implementation vs. judgment
McCrory breaks knowledge work into three parts: asking the right questions, implementation, and expert evaluation. The implementation layer is getting saturated by AI. The other two aren't.
From what I see day to day as a CTO, that tracks.
The developers doing well right now aren't the ones who memorized the most syntax or can write a perfect binary search on a whiteboard. They're the ones who know what to build, can evaluate outputs, and can tell when AI is wrong. Implementation matters less than it used to. Judgment matters more.
That changes how I hire
I'm looking for people who can frame problems clearly, spot when something is off even if it compiles, and guide AI tools without blindly trusting them. People who can think in systems, not just code.
If your hiring process still rewards speed on basic coding exercises, you're optimizing for a layer that's getting automated. The people you actually need don't always stand out in those interviews.
McCrory compared this moment to electricity. The real impact didn't come from simply plugging machines in — it came from reorganizing work around it. We're still early in that shift.
The window
There's a bigger risk in the background. A downturn for white-collar work is possible. Anthropic's own economist has said as much. It hasn't happened yet, but decisions made now will shape whether it does.
That 94% vs. 30% gap isn't a comfort zone. It's a window.
For engineering leaders, using it well means rethinking who you hire, how you evaluate them, and what skills will actually matter soon.
Follow me on X — I post as @oldeucryptoboi.
Top comments (0)