The Augmentation Gap

The finding that mattered most was quieter.

Of all the ways people use AI at work, 52% involve augmentation: humans working alongside the tool, making decisions together, with the person retaining final judgment. Automation, where the AI handles a task entirely, accounts for 45%. The remaining 3% is somewhere in between.

In the fields of computer science and mathematics, AI could theoretically handle 94% of tasks. It's currently covering 33%. In business and finance, the theoretical number is also 94%. Actual adoption sits around 20%. Office and administrative roles: 90% theoretical, 25% real.

The gap is narrowing, but slowly. In January 2025, 36% of occupations had AI touching at least a quarter of their tasks. By November, that number reached 49%. That's real movement. But when you measure what AI is actually completing successfully, the spread is still wide.

Researchers call this the adoption gap. I think about it differently.

AI doesn't just speed up simple tasks. It speeds up complex ones faster.

Tasks requiring a college-level education were completed 12 times faster with AI assistance. Tasks requiring a high school education were completed 9 times faster. The more complex the work, the greater the time savings.

The popular version of the AI story assumes machines handle the simple stuff while humans handle the hard stuff. The data says something more interesting: AI's biggest productivity impact is on the work that requires the most training and expertise to do manually.

There's a deeper point here that goes beyond productivity metrics. For decades, organizations have treated expertise as scarce and rationed it accordingly. The senior analyst builds one financial model because that's all they have time for. The logistics coordinator maps one set of routing scenarios because the alternatives take too long to evaluate. AI doesn't replace the expertise. It removes the bottleneck that kept expertise from being fully deployed. The analyst builds five models and spends their remaining time interpreting them. The coordinator evaluates 30 routing options in the time it used to take to evaluate three.

That's a real concern. But the deskilling scenario assumes a passive response, one where AI reshapes job content by default rather than by design.

The alternative is deliberate. If AI compresses the preparation phase of expert work, redirect that time toward the work that benefits most from a human in the room: building relationships, reading clients, and understanding context that doesn't fit in a dataset.

The customs broker who no longer spends six hours on paperwork visits clients, maps their supply chains, and anticipates regulatory shifts before they arrive.

The U.S. has averaged about 1.4% annual productivity growth over the past two decades. Adding a full percentage point would be comparable to the surge that accompanied the personal computer and internet era.

But here's the qualifier the researchers are careful to include: those gains depend on adoption. The technology has to be used. Right now, AI use remains concentrated in a narrow band of occupations and tasks. The top 10 tasks account for 24% of all AI interactions, up from 21% at the start of 2025. Usage is getting more concentrated, not less.

The organizations capturing productivity gains are those that have moved beyond experimentation and built AI into their workflows. The rest are running pilots, attending webinars, and waiting.