By the time I open my laptop, a scout has already checked job boards for me, a few coding agents are working in separate branches, and anything I write goes through a style checker before anyone reads it. I’ve replaced about half of my subscriptions with software I built this way. It’s easy to forget how unusual that is.
The numbers
Pew Research found in 2025 that 34% of US adults had ever used ChatGPT. Among adults under 30 it was 58%. Two thirds of Americans had never tried it, and 20% said they had heard nothing about it.
The people who do use it mostly use it as a better search box. In an NBER working paper, a team of OpenAI and academic economists analyzed 1.5 million conversations and found that practical guidance, seeking information and writing make up nearly 80% of them. Coding is a small share. By June 2025, 73% of messages had nothing to do with work, up from 53% a year earlier.
For most people, using AI means asking a question, reading the answer and closing the tab. That’s a real product with real value. It just isn’t the product most of us in tech are building for.
Chat makes you the orchestrator
What I get from AI comes from things a chat window doesn’t do by default. My agents carry context across days, load skills I wrote once, run while I sleep, and work in parallel. In a plain chat, you are the memory, the planner, the reviewer and the scheduler, one message at a time.
Most people don’t want that job, and they shouldn’t need it. Asking someone to become an orchestrator before AI is useful to them is like asking them to learn SQL before they can use a spreadsheet.
Building for everyone else
I’ve built on both sides of this. At Microsoft I shipped AI features to Windows users who never asked for an agent, and now I build for scientists who want the result, not the workflow. Three things matter more than model quality:
- Meet the question where it happens. Most usage is practical guidance, about the thing in front of you. The best place for AI is inside that thing: the settings page, the search bar, the lab notebook. When Copilot could look at a user’s PC instead of guessing, its troubleshooting could start from the real problem.
- Do the orchestration for them. Memory, skills and background work shouldn’t be power-user features. A product should remember what you told it last week and get better at your tasks without you writing a prompt library.
- Ask before acting. The step from answering to doing is where trust breaks. People let software act for them once they can see what it will do and say no. Consent isn’t friction. It’s the feature that makes delegation possible.
The gap is the opportunity
The distance between how I use AI and how most people I know use it isn’t intelligence or curiosity. It’s product. Every capability that makes my setup useful, from memory to delegation to skills that improve with use, exists today, wrapped in tools that assume you’ll configure them yourself.
The next billion users won’t configure anything. Whoever packages the power-user stack so it works for them without setup will matter more than whoever tops the next leaderboard.