Fifty-eight percent of AI-using knowledge workers say they're producing work they couldn't have a year ago. The figure comes from Microsoft's 2026 Work Trend Index, surveying 20,000 people across ten countries. It is self-reported. It doesn't distinguish between work you lacked the skill for, work you lacked the time for, and work that simply required a tool you now have access to. These are three very different kinds of "couldn't."
Something real lives inside that number, though. People are producing competitive analyses, research syntheses, and strategic documents at a pace that would have required a specialist or a team eighteen months ago. The interesting question is what kind of capability this represents.
When you spend your days reviewing, refining, and redirecting AI-generated work, you develop something. You learn where the model hallucinates, which prompts produce brittle outputs, how to spot a confident paragraph that's subtly wrong. You build a detailed internal map of the system's failure modes. This feels like expertise. It involves pattern recognition, judgment, accumulated experience. By any experiential measure, you are getting better at your job.
A study in Scientific Reports complicates this. Researchers examining expertise in visual composition found that people who actively produce creative work develop different judgment than those who study and evaluate it. Both groups build what the researchers call "well-elaborated mental structures of categories." Both can assess quality. The difference is that practitioners maintain what the study describes as a professional interest in creating novelty. Curators and critics develop the categorical architecture without the generative impulse.
Applied to AI-augmented work, this suggests something worth sitting with. The mental model you're building of your agent's error patterns is real. It is also, potentially, a different thing than the domain knowledge it sits beside. You know where the AI gets pricing analysis wrong. Whether you could construct the pricing analysis from scratch is a separate question, and one you may never need to answer. Until the day you do.
The most advanced AI users seem to sense this. Buried in the same Microsoft survey: 43% of "Frontier Professionals" deliberately do some work without AI to make sure their foundational skills don't atrophy. Nearly half the power users are quietly hedging. They're maintaining a practice the way a pianist who conducts orchestras still plays scales alone, to preserve the ability to hear whether the orchestra is actually in tune.
The editing analogy goes part of the way. Great editors develop taste without writing every sentence. Film directors shape performances they couldn't give. But these roles emerged within ecosystems where evaluative and generative traditions co-evolved over decades, where editors had usually written extensively, where directors had studied their craft with obsessive granularity. The curator role in AI-augmented work is being invented in real time, without that co-evolutionary history, by people who may or may not have done the underlying work before the agent arrived.
What remains genuinely open: evaluative judgment might be sufficient for the work ahead. A study testing expert reviewers found they outperformed frontier AI models at evaluating research, but not by the margin their credentials would predict. Individual expert accuracy averaged 36.2%. Evaluative judgment is genuinely hard, even for people who spent careers building it through generative practice. Whether it can be built without that practice, through curation alone, is a question we're running the experiment on right now. And you can't tell the answer from the inside.

