Where did computer use AI come from?
Anthropic released computer use in October 2024, a way for its model to look at a screen as an image and send mouse and keyboard actions back. OpenAI followed with Operator in January 2025, an agent that drives a web browser to book, buy and fill in forms. Since then the idea has spread: you describe the task, the AI takes the screen, and you watch it work or walk away.
The underlying ability, reading a screen as an image, is the same one that lets an AI see your marketing. The split is what happens next. One kind of product hands the mouse to the AI. The other kind keeps the mouse with you and uses what it sees to tell you where to click.
Why is an ad account the worst place for an AI to hold the mouse?
Marketing accounts are the worst place for an AI to hold the mouse. A budget field, a campaign status, a published post: each is one click from a mistake that costs money or reputation, and an agent working on its own makes that click before you see it. The appeal is obvious, no more doing the work. The cost is that you no longer know what your own accounts are doing, and when something goes wrong you cannot say why.
There is a second cost. Google and Meta can see automated activity in an account, and both have terms about who may operate it. An owner handing their login to an agent is taking on a risk the agent's maker has not.
What is the one question to ask?
Before you use anything that offers to work your computer or your accounts: can it change anything, and what stops it. If the answer is "it only does what you ask", that is not a limit, because what you ask can be wrong. The limit you want is a hard one, written into the software: it can read, it can show, it cannot change. That is how CoraVision is built.
What happens when a daily budget is raised, two ways? (example figures)
An owner wants the daily budget on a Google Ads campaign raised from $40 to $50.
| Computer use AI | Screen-reading AI (CoraVision) | |
|---|---|---|
| Who opens the campaign | The agent | The owner, with Cora pointing at it |
| Who types 50 | The agent | The owner |
| Who clicks Save | The agent | The owner |
| If 500 is typed by mistake | The campaign spends $500 a day until someone notices | Cora reads 500 on the screen and says so before Save |
| What the owner learned | Nothing | Where the budget lives and what changing it does |
The second column takes thirty seconds longer. The owner can repeat it without help the next time.
Which side of the line is CoraVision on?
CoraVision is built on the second column of that table. Cora sees your screen during a session you start, as still images that are not saved, and in pill mode points at the next step while you do it. CoraVision reads Google Ads, Meta and Google Analytics 4 every night once they are connected, and only reads: it never creates, changes, pauses or deletes anything, and it cannot post to your Facebook or Instagram pages.
Questions
Does CoraVision control my computer?
No. Cora sees your screen during a session you start and points at the next step. You move the mouse. Nothing in CoraVision can click, type or change a setting on your computer or in your accounts.
Is computer use AI safe to use with my ad accounts?
Only with a hard limit on what it may do, and most tools do not offer one. An agent that can change a budget can change it the wrong way, and the mistake is live before you see it.
When would an owner want it?
For dull, reversible work with no money attached: renaming files, filling a spreadsheet, moving data between two tools. Not for anything that spends or publishes.