AI Adoption Is Social

New technologies do not all spread the same way. Some arrive with an obvious before and after. The smartphone did not need much persuasion once people saw a browser, a camera, maps, and email in one pocket. Cloud storage had a similar pull: lose one laptop, or try to work from two machines, and the argument was over. Other technologies spread socially. They need word of mouth, not because they are weak, but because the value is hard to understand from the outside. AI is mostly in the second camp. You can watch a demo and still miss it. You can read a benchmark and still not feel it. The adoption model is not "show me the score." It is "show me someone like me using it for my work."

That is why the recent Microsoft paper on command-line AI coding agents landed for me. Studying tens of thousands of engineers using Claude Code and Copilot CLI, the authors found that "first use spread primarily through social networks" and that adopters "merged roughly 24% more pull requests" than they otherwise would have. The important part is not just the number. It is the path. Awareness spread through peers and managers actually using the tool, not through a clean announcement from the top. AI adoption looks less like installing a mandated enterprise package and more like seeing the person next to you suddenly move faster, then asking what changed.

I felt that friction myself. For a long time I treated new AI tools as things to sample, not things to default to. I would open them for a special task, get a decent result, then go back to the old workflow because the old workflow had muscle memory on its side. The ice broke only when I started using them first instead of last: Gemini for presentations, Claude and Codex CLI for coding, AI Mode as the default shape of many Google searches. That shift is hard to explain without sounding like hype, because the gain is not one magic answer. It is the compound effect of making the first draft, the first search, the first outline, the first refactor, and the first critique cheaper. Going AI-first is tremendously powerful, but you have to use it to believe it.

The tools still have obvious gaps. The current CLIs are often too cautious in the wrong places and too bold in others. They ask permission for harmless moves, then miss the moment when a simple task should just be carried forward. They wait for user input when a reasonable next step is obvious, but can still wander if you let them go fully yolo without rails. They forget preferences after a /clear, lose project taste unless it is written down, and need too much babysitting around the small habits that make a workflow feel personal. The raw capability is already useful. The product shape is still immature.

Still, the direction is clear. The shell was always the best place for serious work because it is composable, inspectable, and close to the machine; AI agents make that even more true, which is why the best ones increasingly live in the terminal. But we should be honest about what the Microsoft result measures. A merged PR is a useful proxy for output, not proof of flawless production value. Today most of the benefit lives in mid-quality hobby projects, game experiments, internal prototypes, dashboards, migrations, tests, and glue work. Getting to bug-free production code that understands the broader system, the incident history, the architecture, and the hidden constraints will take a while. Use it anyway. The future usually arrives first as a habit.