Maybe people skip AI tools not because they're bad, but because no one realizes the task could be automated at all.

This explores whether AI adoption stalls because people don't know a feature exists or what they'd use it for, rather than because the technology falls short. The corpus has no study that measures feature discoverability directly. What it does have points to something more interesting: the hardest part to discover isn't the feature. It's the problem the feature would solve.

The sharpest version of this comes from Evans Does easier tool-building actually solve enterprise adoption problems?. He argues that making AI tools easier to build or use doesn't fix the real bottleneck, because most workers don't see their own tasks as automatable. You can put a capable feature one click away, and someone who doesn't see their weekly report or approval chain as something a machine could handle will never click it. That's a recognition problem more than a menu-design problem. On top of it sits a second barrier: in companies, adoption is an organizational decision that cuts across departments and budget cycles, so even someone who spots a use often can't act on it alone.

A historical analysis that runs from GPS to modern AI agents makes a similar point at a larger scale Why do capable AI agents still fail in real deployments?. Capable systems repeatedly stalled because the conditions around them were missing: clear value, personalization, trust, social acceptability and shared standards. None of these is 'discoverability' in the narrow UI sense. Together they describe whether a tool fits the user's life well enough to be noticed and kept. On this reading, a feature that's hard to discover is often a feature whose value hasn't been made obvious in the user's own context.

Developers are a useful counterexample, because for them discovery is clearly solved. Stack Overflow's 2025 survey found 80% of developers using AI tools, while trust in the tools' accuracy fell from 40% to 29% Why do developers keep using AI tools they don't trust?. Once people do find and use a tool, the limit moves somewhere else. In this case it's the work of checking code that looks right but contains subtle errors. Discoverability may decide whether people start using AI, but it doesn't decide whether the use is deep or worthwhile.

The corpus suggests that discoverability limits adoption mainly as a matter of seeing your own work differently, not as a matter of finding buttons. If you want to go further, the Evans note is the best place to start, because it changes the question from 'can people find the feature?' to 'can people recognize their own work as something the feature applies to?'