What can ChatGPT computer use actually do for your business?
A real run: one prompt, five Austin venues researched and five tailored emails drafted, with the send button still ours.
ChatGPT computer use lets the desktop app drive your browser and apps the way you would: opening tabs, filling forms, and drafting emails while you work elsewhere. It is slower than doing the task yourself, so the payoff is unattended time, not speed. Start with work that is annoying, repetitive, and easy to check.
What computer use actually is
Computer use is the model driving your machine instead of talking to you about it. It opens a browser window, reads what is on screen, clicks, types, and moves between apps the way a person would. You give it a goal in plain language and then you go do something else. That last part is the point.
This is different from a chat that hands you a research summary and leaves the clicking to you. It is also different from an integration built for one specific tool. Computer use does not need an API for the site it is working on. If you can do the task with a mouse and a keyboard, it can attempt it.
You need the desktop app
This does not run in the web version. Download the desktop app for Mac or Windows and start a chat there. The first time you invoke it, it asks for permission to control the screen. Grant what you are comfortable with.
There are two ways to trigger it. The official way is to tag computer mode in your message. The informal way is to say "use computer mode" and then describe the task. Both work the same.
Pick the stronger reasoning model for anything involving computer use. It has to look at a page and decide what to click, and that takes more thinking than answering a question. Save the fast, cheap models for straight lookups. Asking a top-tier model where to get lunch gets you a thirty minute answer to a ten second question.
A real run: five venues, one prompt
Here is the task we gave it. We are hosting an AI workshop in Austin for twenty to thirty people, and we needed venues downtown or south of it, inexpensive, with a screen we could present from. We started in an ordinary chat and asked for five options.
It came back with five, each with capacity and an hourly rate, plus a note on what to confirm before paying: whether the screen is big enough and how the Wi-Fi actually performs. Then we handed that list to computer mode. Go to all five, find their contact forms, ask about the screens, the Wi-Fi, and availability for a specific date.
It opened its own browser and went. You can watch it work, and it is unsettling the first time. Tabs open that you did not open. The cursor moves on its own. A small indicator in the corner tells you the screen is in use.
It drafts. You send.
Partway through, it opened Gmail and started composing. That is the moment people tense up, so it is worth being precise about what is happening. The email connection in this setup is read and draft only. It can read the inbox and it can write a draft. It cannot press send.
Ten minutes later there were five drafts in the drafts folder, each addressed to the right contact, each asking the questions we specified. We opened the folder, read them, and sent them ourselves. That boundary is what makes this usable day to day. The agent does the tedious ninety percent and stops at the step where a mistake would be public.
It is slower than you. That is fine.
Ten minutes for five emails is slower than doing it by hand. That is not the pitch. The pitch is that you are not in the chair for those ten minutes. You can work on the other monitor, take a call, or leave the house entirely. The return is time you were not going to spend anyway, not raw speed.
Read the output, because it slips
The drafts were not perfect. On the first two it garbled the opening sentence and dropped a stray period into the middle of a phrase. By the fourth it had cleaned that up without being told. Nothing was wrong enough to matter once we read them, but any of it would have been embarrassing to send unread. Read what it writes before it goes out. That is the whole quality process.
Pick your first task with ACE
Do not start with wiring money or sending a message you would regret getting wrong. Use ACE instead.
Pick one task that is annoying, common, and easy to verify.
- Annoying. Something you actively resent doing.
- Common. Something that comes up daily or weekly, so the setup effort pays back.
- Easy to verify. You can tell in ten seconds whether it did the job right.
Venue outreach hits all three. So does most of the clicking you do to move information from one place to another. You already spent the judgment deciding to run the event. Finding five contact forms is not where your brain earns its keep.
Other jobs worth handing over
- Pipeline hygiene. Walk the CRM and flag records that are stale, half filled in, or duplicated.
- Vendor checks. Verify a supplier's details against their own site before you sign anything.
- Cross-referencing. Compare a list of people or companies against work you have done before.
- Moving data. Anything that is currently you copying out of a CRM into a spreadsheet, or back the other way.
The test is simple. If the task is mostly you clicking, it is a candidate.
And from your phone
The mobile app can kick off a run on the computer sitting at your desk. You are at the gym, you ask it to pull six contacts and tell you who to reach out to first, and it goes and does that at home while you finish your set. That setup deserves its own walkthrough, so we will cover it separately.
The short version
- Computer use runs in the ChatGPT desktop app only, not the web version.
- Use a stronger reasoning model for computer-use runs and a cheap fast one for plain lookups.
- An email connection set up read-only can draft but not send, so the last click stays yours.
- Expect it to be slower than you at the task. The win is that you are not at the keyboard.
- Read every draft before it goes out. It still garbles a sentence now and then.
- First task should be annoying, common, and easy to verify. Never money, never anything emotional.
Questions this raises
Can ChatGPT send emails on its own?
Not in this setup. The email connection was read and draft only, so it could read the inbox and write drafts but could not press send. The five venue emails sat in the drafts folder until we opened them and sent them ourselves. Check what your own connection is permitted to do before you rely on that.
How long does a computer use run take?
The five-venue run took about ten minutes end to end. Expect it to be slower than doing the task yourself. The value is that you are not at the keyboard while it runs, so you can work on something else or leave the house.
What should I hand it first?
Something annoying, common, and easy to verify. Research and outreach, CRM cleanup, or moving data between a spreadsheet and another system all qualify. Do not start with anything involving money or a message you would not want sent incorrectly.