ChatGPT Can Now Use Your Apps

Introduction

Every business has a list of jobs nobody wants and nobody can automate. Someone logs into a supplier portal every Monday to download invoices, then types them into the accounting software one line at a time. Someone else fills in the same government form every quarter, copying numbers from a spreadsheet into boxes on a website. Neither system offers a way for other software to connect to it (an API), the vendor isn’t going to build an integration for a customer your size, and so the work stays manual.

At the end of October, OpenAI released GPT-6.1 Terra opens a new window , a model built for what OpenAI calls computer use: it looks at a screen, reads what’s on it, and clicks and types its way through apps and websites. OpenAI’s own examples are the kind of work on that list: reading incoming invoices and entering their details into accounting software, filling out forms, and interpreting technical drawings. It runs in ChatGPT Work, the agent side of ChatGPT, on plans a lot of businesses already pay for. Like Claude in Google Workspace opens a new window , it’s a capability that shows up inside a tool you already have rather than a new purchase.

That makes the question practical rather than theoretical. If you already pay for Plus, Business, or Enterprise seats, you can try this next week without buying anything new. What follows is how we’d decide which jobs to hand over first, how to watch it while it works, and the places it will let you down.

What actually changed

Computer use isn’t brand new in ChatGPT. It arrived with the desktop app relaunch in July, and OpenAI’s larger models could already do it, at a cost in speed and usage that made it hard to justify for routine work. What’s new is that Terra is fast enough, and light enough on your plan’s usage allowance, to make everyday back-office jobs worth trying.

In practice, it works like this. You open the ChatGPT desktop app on a Mac or a Windows machine, switch to Work, and turn on the Computer Use opens a new window plugin. Then you describe the task, naming the app and the flow: “Open the supplier portal in Chrome, download this week’s invoices, and enter each one as a bill in our accounting software.” ChatGPT takes screenshots, reads them, and moves the mouse and keyboard the way a person would. It asks your permission before it uses each app.

According to OpenAI’s release notes opens a new window , Terra is rolling out in Work and in Codex, OpenAI’s coding tool, on the Plus, Pro, Business, Enterprise, and Edu plans, with availability depending on plan, region, and workspace settings. It isn’t offered in regular Chat, and on Enterprise and Edu workspaces an admin may need to grant access before people see it. Your plan’s usage limits don’t change, but OpenAI says the same allowance covers more tasks with Terra than with its larger models.

The jobs worth handing over

The best candidates share three traits, close to the ones we described in Automating Repetitive Work with Skills and MCP opens a new window . The work lives in an app or portal with no integration, it follows the same steps every time, and a person can check the result quickly.

Invoice entry is the obvious one, and the one OpenAI leads with. The model reads a PDF or a portal page, finds the vendor, date, amounts, and line items, and keys them into the accounting software. A bookkeeper still reviews the bills before they’re approved, but reviewing ten bills usually takes a fraction of the time it takes to type them.

Supplier and customer portals are the next tier. That means downloading statements, checking order status across five vendor sites, updating your company details everywhere after a move, or pulling remittance advice. Each of these is a few minutes of clicking that nobody budgets for and everybody resents.

Government and compliance forms fit too, with a caveat we’ll come back to. Moving numbers from a spreadsheet you already prepared into a web form is exactly the kind of tedious, visual task this is good at. Submitting it is a different question.

Then there’s anything that involves reading a picture. OpenAI put a lot of weight on visual reasoning in this release, and its examples include turning technical drawings into models. For a business, that might mean pulling dimensions and part numbers off a drawing into a quote, or reading scanned delivery notes that never came in as data.

Notice what’s not on the list. If your accounting software and your CRM already talk to each other, or there’s a proper connector, use that. OpenAI’s own documentation says the same thing: when a dedicated plugin exists, it’s more structured and repeatable than having the model click around a screen. Computer use is for the gaps integrations don’t cover.

How to supervise it

Think of this as a new hire who is quick, tireless, and has never seen your systems before. You wouldn’t give that person the company credit card on day one, and the same instincts apply here.

Start by watching. Run the first few tasks with someone sitting in front of the screen, and treat them as training sessions for you rather than time saved. You’ll learn where your instructions were vague, which screens confuse it, and how long the job really takes. OpenAI’s documentation says you can stop the task or take over the computer at any time, and that you should cancel if it starts working in the wrong window.

Keep a human on the last click. That is the caveat on government forms, and it applies everywhere else too. Have ChatGPT prepare the bills, fill the form, or stage the order, then stop. A person reviews and presses submit or approve. In our experience, this one rule covers most of the risk, because the costly mistakes are the ones that leave your building: a payment, a filing, an email to a customer. ChatGPT will sometimes ask for permission before sensitive actions on its own, but you shouldn’t rely on its judgment of what counts as sensitive in your business.

Saved approvals deserve some care. When ChatGPT asks to use an app, you can approve it once or tell it to stop asking. Apps you’ve approved permanently show up under “Always-allowed apps” in the settings, where you can remove them. Grant that standing approval only for apps where a mistake is cheap to undo. If your company is on a Business, Enterprise, or Edu plan, your admin can also restrict which apps Computer Use can touch, and whether people can save approvals at all.

It also helps to give it its own space, because computer use acts as whoever is logged in. If it uses your browser, websites see its actions as yours, with your permissions, on every account you’re signed into. A separate browser profile, or a separate machine or virtual desktop, signed into only the accounts the task needs, limits the damage a mistake can do. On Windows this matters for another reason too: ChatGPT takes over the mouse and keyboard while it works, so you can’t use the same desktop at the same time. A narrower setup also keeps the model away from data it has no reason to see, which is the question we worked through in Safely Leveraging AI: Privacy and Security opens a new window .

Finally, check the output rather than the activity. Watching the cursor move is reassuring and tells you very little. What matters is whether the bills in the accounting software match the invoices. Build a quick check into the process, like a total that should match or a count of records, so the review is a comparison, not a re-do.

Where it breaks

The marketing focuses on what Terra gets right. Here is what we’d plan for.

It makes mistakes, and they look like correct work. OpenAI made real progress on factual errors in this release. On its own internal set of prompts designed to trip models up, the share of answers with an error dropped from 53.6% on the previous Terra model to 16.7% at low reasoning effort, and sits at 4.3% at maximum effort. That’s a big improvement, and it’s also a reminder that a transposed digit or a misread line item will slip through sometimes. In data entry, those errors don’t announce themselves. This is why the human on the last click isn’t optional.

It also hasn’t seen your screens. OpenAI measures Terra on benchmarks like OSWorld, a test of tasks in desktop apps, and Agents’ Last Exam, a set of long multistep jobs. The scores are impressive relative to earlier models. None of those tests include your vendor’s portal from 2011, with its pop-ups, session timeouts, and a date picker that only works if you click it twice. Expect the first runs on any new system to be rougher than the demos, and plan the pilot accordingly.

Logins and security prompts stop it. Computer use can’t authenticate as an administrator or approve the operating system’s security and privacy prompts, and it can’t automate terminal apps or ChatGPT itself. We’d expect two-factor codes, CAPTCHAs, and “confirm it’s you” emails to be the points where a person has to step in. OpenAI also advises against handing it tasks that need passwords or other secrets unless you’re present to approve each step. For most back-office work, that means someone logs in first and lets the model take it from there.

Then there are the websites themselves. A page the model reads can contain instructions written to hijack it, known as prompt injection. OpenAI says Terra showed strong resistance in its tests, and that’s good news, but strong resistance isn’t immunity. OpenAI’s own safety testing opens a new window also found that Terra was more willing than GPT-6 Astra and GPT-6.1 Sol to work around a warning or access restriction: in 29.6% of test runs at maximum reasoning effort, it saw the barrier and finished the task another way. The reason OpenAI gives is that it sometimes reads a legitimate restriction as an attack and ignores it. Those tests ran without OpenAI’s production safeguards, so the real-world rate should be lower, but the tendency is worth knowing about. Both are reasons to keep the model away from accounts it doesn’t need and away from the payment button.

And some tasks still need the bigger model. OpenAI itself recommends GPT-6 Astra for the hardest tasks that depend on screenshots or visual judgment. If Terra keeps failing on a particular screen, switching models for that task is worth trying before you give up on it.

Conclusion

In this post, we looked at what GPT-6.1 Terra adds to ChatGPT Work, which back-office jobs are worth handing to it, how to supervise it, and where it still breaks. The practical next step is small. Pick one job from your list. It should be repetitive, live in a system with no integration, and have a result someone can check in a few minutes. Invoice entry from a single supplier is a good first candidate. Write the steps down the way you’d explain them to a temp, run it supervised a handful of times, and time both the run and the review. If the review takes less time than doing the job by hand, and the error rate is one you can live with, you’ve found work your ChatGPT seat can take off someone’s plate.

Deciding whether a task belongs in computer use or needs a proper integration is a close cousin of the questions we covered in When Does Your Business Need an AI Agent? opens a new window and Finding the Right Problems to Solve with AI opens a new window . Not sure which jobs on your list are good candidates? Send us a message opens a new window and let’s look at it together.