What to know: what an AI can do depends on the tools and access you give it, more than on the model. Chat apps can already search the web, read what you upload, and with connectors reach your email, calendar or files. How much varies a lot by app: ChatGPT does far more than Gemini, for example. If you want it to do more, it needs more tools and more access. Each rung on this ladder gives it more of both.
More reach means more it can do for you, and more that's exposed if it gets something wrong or reads something it shouldn't.
- In the chat. It answers from what you upload or paste, searches the web, and reaches whatever you've connected to it, like Gmail, Google Drive or your calendar. Most people are here.
- In your browser. For example, Claude in Chrome. It clicks and types in your own browser, and it's already signed in as you, so it can reach every site you're logged in to.
- On your files. For example, Claude Code, or Codex in ChatGPT's work version. You give it access to your computer's files, usually starting from one folder. It can read, make and change files there, run programs and write its own tools. This is where the Basics layer picks up.
- Running in the cloud, always on. For example, ChatGPT dots, Claude Duet and Grok Bot. The provider runs it for you, and it keeps working when your laptop is off. Some give each one a cloud computer of its own. On others, every agent on one account shares one computer, with the same logins and files. It reaches whatever apps and logins you connect to it.
- On your computer. For example, Claude Cowork in the Claude desktop app. When there's no connector or browser route for a job, it uses what's called computer use: it sees your screen and moves your mouse and keyboard. Everything open or signed in on that computer is in reach: email, bank, work systems. It asks before it uses each app, but nothing walls it off from an app once you've said yes, and it can see anything left open on screen.
How to choose: it's a trade. Each rung up saves you more work, and puts more at risk. So start by naming the problem you're trying to solve. Then weigh what solving it is worth against what the AI could reach.
For the rung you're looking at, ask:
- What could it reach? Only the apps you've connected, or your files, or every site you're signed in to, or everything on your screen?
- Can a mistake be undone? A changed file can be restored from a copy. A sent email, a payment or a deleted account can't.
- Whose information is in there? Yours, or your customers', your family's, your employer's? Whatever it reads goes to the AI company.
- What will it read along the way? A web page, an email or a document can carry hidden instructions written for the AI. The more it can reach, the more that matters.
- Who's watching? On rungs 1 to 3 you usually are. On rungs 4 and 5 it can work while you're away or not looking, so it runs on whatever rules you set beforehand.
Then take the lowest rung that solves the problem. Your files are often enough. Your whole computer rarely is. Some tools sit on more than one rung: Claude Code works on a folder, and can also be given computer use.
These products change. Check what yours can do today.
Reveal on the spot: say which rung the person is on now, from what you can see of your own tool. Ask what job they'd most like off their plate, and say the lowest rung that would do it. Then say what that rung could reach that the current one can't.
Unlocks: knowing what each kind of AI tool can reach, and picking the smallest step that gets the job done.