Most small shops have at least one piece of software that will not talk to anything else. A supplier portal with no export button. A scheduling screen on one old Windows computer. An accounting tool where every invoice means typing the same details again.

For years the only fix was a person with a spare hour. A new kind of AI tool is aimed at exactly that job, and the price is dropping fast enough that it is worth understanding.

A model that works the screen like a person

On October 26, OpenAI released GPT-6.1 Terra, which it describes as a model for “computer use”: it looks at a screen, fills in forms and clicks through apps the way a person would, instead of connecting to the software behind the scenes. According to its pricing page, it costs $1 per million input tokens and $5 per million output tokens, which is about a tenth of what its top model charges for input. Tokens are how AI tools measure the text they read and write. OpenAI says Terra is rolling out in ChatGPT Work and Codex, and through its developer platform.

OpenAI also says it is fast and accurate. Those are its own claims, and nobody outside the company has tested them yet. The price is a list price, and it is a fact.

Try the boring fixes first

Before you point any AI at a screen, check whether the screen is the only way in. A tool that clicks around a website is slower and more fragile than a direct connection. If any of these exist, use them instead:

  • An export. A “download CSV” or “export to Excel” button, even a hidden one in the reports menu.
  • An email option. Many portals can email a daily or weekly report on a schedule.
  • A connection. Many apps connect to others through Zapier or a built-in integration.
  • Asking. A short call to the software company often turns up an export nobody knew about.

Screen-clicking is for what is left over: the software with no other way in.

A made-up example

Here is a made-up example. An HVAC company’s office manager spends an hour every Friday copying completed job notes from a supplier’s ordering portal into the company’s accounting software. The portal has no export button, and the vendor says an integration is not planned.

She tries the boring fixes first, and none works. So she tests a screen-operating AI tool on the Friday job. She creates a separate login for it that can see job notes but cannot place orders. In the first two weeks she watches every run. The tool misses a field once, after the portal moved a button. By week three she reviews a short list at the end instead of watching.

The job drops from an hour of copying to about ten minutes of checking. If the portal changes again and the tool breaks, she goes back to copying by hand for a day while it gets fixed.

How to start safely

  1. Pick one repetitive job. Choose something you can describe like training a new hire: which screen, which fields, what “done” looks like.
  2. Give it its own login. Make a separate account with only the access that job needs. Never give it your own login or an owner account.
  3. Watch the first runs. Sit with it for two weeks. Count how often you correct it.
  4. Keep money with a person. Payments, refunds, tax filings and bank details stay with a human. The tool can fill in the form. A person presses the button.
  5. Expect it to break sometimes. When a website changes its layout, a screen-clicking tool can get lost. Know what you do on those days.
  6. Keep a record. Ask for a log of what it did on each run, so you can answer “what happened?” later.

Your one step this week: write down the one screen job in your office that a person does by copying from one place to another. Then spend ten minutes looking for an export button for it. If there is one, you have saved yourself the whole project.

If you want help working out which of your screen jobs are worth automating, here is what we build.

Sources

Kush AI Automation builds in accounts you own, and nothing that moves money happens without a person approving it first. Bring the screen job you most want off your plate and book a free call.