We are at a point where we do not need to redesign every piece of software for AI. An agent can operate the software as it is.
The buttons, menus, and text fields are already there. Give the agent the tools to see the screen, click, and type, and it can work through that interface. Both OpenAI and Anthropic now document computer use for apps that do not have a dedicated integration.
One small example: in its 2025 Computer-Using Agent report, OpenAI reported an agent creating a grocery-shopping project in Todoist and adding the items through the interface. The agent used the same controls a person would use. A shopping list is ordinary work. That is why it is a useful example.
The agent still needs access, and it can get things wrong. Longer workflows can need another attempt, and the result still needs checking. Direct integrations can be faster and more reliable, so there are still good reasons to build them. But a dedicated integration is no longer the only way in.
That is the change I find interesting: AI is learning to use the software we already have. We do not have to rebuild every application before it can take part.