Our relationship with AI is shifting from a Q&A chat into a coworker hand-off. The keys to this shift are folder access and context management. Today we walk through the principle from the ground up.
Most people who use AI today have fallen into the same habit. You open a chat window, type a question, copy the answer, paste it into your own document. Today I want to walk you through why that picture is about to feel old — and what the next one looks like. We'll go slowly.
Let's start with the principle. Good working relationships move from 'exchanging information' to 'delegating work.' When a new colleague joins, they ask about everything for the first month. By month three, you hand them tasks. By year one, you just receive finished results. The reason this drift is natural is simple — swapping questions and answers every time is more tiring and less accurate than sharing context and receiving outcomes.
Among humans this is so obvious it doesn't need saying. But the moment AI enters the room, the common sense disappears. I regularly see people spend 30 minutes in a chat window doing a job that should take 10. Why? Because our first impression of AI — the 'answer machine' — was too strong. ChatGPT in 2022 really was just that: ask, receive. Three years of that shape has stayed in our bodies.
To make this concrete, let me use a feature Claude launched in research preview in January 2026. It's called Cowork. Other companies are building similar things, so don't read this as 'a Claude thing.' Read it as a single frame from a much bigger movement.
Open the Claude desktop app and the window splits in two. Chat on the left. Cowork on the right. The difference is simple.
Say I type into Cowork: 'Prepare materials for tomorrow's meeting.' It asks a few questions — 'What kind of meeting?' 'What output do you want (PPT, script)?' 'Is the audience familiar with AI?' — then plans on its own, searches the web, builds the slides, reviews and revises itself once, and drops the final file into the folder I designated. I just open the file when it pings me.
Here's the easiest way to feel the shift. Imagine an office. When you have a work question, you might text a friend. Your friend answers, you take their answer, and you build the document yourself. With a secretary it's different. 'Draft tomorrow's materials,' you say, and the secretary researches alone, leaves a draft on your desk, and you just check and edit.
Chat is the friend. Cowork is the secretary. The friend makes you move. The secretary moves for you. You need both, but they have different jobs.
For AI to become a secretary, though, one thing is non-negotiable: folder access. A friend can't open your desk drawer. A secretary can. Without the drawer, there's nowhere to put the result. That's exactly why Claude stayed an 'answer machine' for over two years — browser policies wouldn't let it touch your computer. So every result lived only as words.
Let me show you the actual difference with a small experiment. Same task: a 10-slide PowerPoint plus a demo script for a meeting.
| Method | Time | User involvement | Final output |
|---|---|---|---|
| Chat | ~40 min | ~20 back-and-forth copy-pastes | Text, manually moved to PPT |
| Cowork | ~12 min | 2 short answers | .pptx + .docx files directly |
40 minutes becomes 12. Roughly 3.3× faster — but speed isn't the real point. The feel is. After chat I feel like I worked. After Cowork I feel like I came back from a break. Same output, different body.
Here's the aha.
AI is moving from 'quick answers' to 'handed-off results.' The less you move, the better you're using it.
To delegate well, just two things.
First, scope the folder. Don't open your entire company drive. Make a small, dedicated folder like Documents/cowork-lab, and let AI read and write only inside it. You don't want AI accidentally touching something important. Take a backup first.
Second, make it plan before it executes. Tell AI: 'Don't do it yet — explain the method first.' If the method is good, say go. If not, fix it. This is called 'planning mode.' As of 2026, Claude Code has it as an official mode, and similar features are appearing in other tools. Without a planning step, it's hard to intervene midway and token usage spikes. One minute of planning saves ten minutes of execution.
The smallest possible experiment you can run today:
1. Install the Claude desktop app (mac or windows)
2. Create a 'cowork-test' folder in your home directory
3. Open Cowork, grant access to that folder only
4. Type: 'Organize my week into a planning sheet'
5. Review the plan, approve execution
6. Open the generated file
As of January 2026, Cowork is research preview — only Claude Max subscribers get it, and only on Mac. Windows isn't supported yet. General release is coming soon. If you can't access it yet, you can feel the same principle through Claude Code's planning mode, ChatGPT's Canvas, or Gemini's Deep Research. The tool name doesn't matter. The hand-off feeling does.
One more thing — safety.
Delegating means handing over permission. Permission means accidents are possible. Three rules keep you safe. First, never grant access beyond the dedicated folder. Don't open your whole desktop or documents folder — AI might alter something important by accident. Second, keep backups separately on iCloud or Google Drive. Assume the AI folder could evaporate any day. Third, watch for prompt injection. If you feed AI a document from the internet, hidden instructions inside that document can hijack it. Such attacks are being reported in 2026. Don't hand sensitive folders to AI.
Summary.
The relationship with AI moves from 'ask and answer' to 'delegate and receive.' Two technical keys make the move possible. Folder access — giving AI a place to put results. And context management — letting AI remember yesterday and think about tomorrow. In 2026 this shows up under the name Cowork, but that name will change. The principle won't.
Each time you reach for AI from now on, ask yourself: am I chatting with a friend, or handing off to a secretary? Just picking the right mode for the moment changes the rhythm of your work. The era of 'results that live inside a chat' is almost over. Delegated results become the default. When that happens, 'experience delegating' becomes the most expensive skill.
Technology changes. Relationships evolve. The principle stays.
Not asking but handing. Not answers but artifacts. Not talking but trusting.