Our own software Solutions Websites InfinyAI School How we work Blog News Contact View 3D website
Analysis

How to automate legacy software with no API using AI agents that use the computer

Almost every small business has software that does not connect to anything: a 15-year-old invoicing program, a supplier portal or a desktop application that only works with a mouse and keyboard. As long as it has no API, the only way to get data into it was for someone to copy it by hand. AI agents that use the computer the way a person does change that, and GitHub has just taken a clear step in that direction.

What GitHub announced

On October 1, GitHub introduced computer use for Copilot, its AI coding assistant, in public preview in Copilot CLI and in the GitHub Copilot app, on macOS and Windows. Copilot reads app content and what is on screen, clicks, types, presses keys, scrolls and moves information from one application to another. According to GitHub, it is meant precisely for legacy or GUI-only software, with no API, command line or MCP integration.

Three details matter for a business: Copilot asks for approval before controlling each app, the list of always-allowed apps can be reviewed or reset at any time, and organizations can disable the feature through their managed settings.

Which tasks fit

An agent like this does well on repetitive tasks with clear steps that are easy to check. For example:

  • Entering orders that arrive by email into the management software.
  • Copying data from a supplier portal into the spreadsheet where you track costs.
  • Updating prices or product records in an application that cannot import a file.
  • Putting together the same weekly report from several screens.

Tasks where a mistake is expensive and not caught in time do not fit, such as making payments or sending documents to customers without reviewing them.

How to start safely

  1. Pick a single task that takes hours of copying and pasting today, and measure how long it takes.
  2. Test it with sample data or a copy of the software before touching real data.
  3. Do not give permanent permission to sensitive apps, such as online banking or company email. It is better to approve each time.
  4. Review the result and also the path: which windows the agent opens and what it does to finish.
  5. Decide who in the company can turn it on. The organization can switch the feature off if it does not want it.

When it is not worth it

Today it is a preview inside a tool for developers, so in a small business it will usually be tested by whoever handles the technical side. And if your software does have an API or can import files, connecting to it directly is usually faster and more reliable than an agent moving the mouse. We believe an agent that uses the screen is the answer for what has no other way in, not a shortcut for everything.

What matters is the signal: automation no longer depends on the software being ready for it. If you have a task like this, now is a good time to write it down.

More analysis