Much of today’s “AI computer use” involves simulating a human user: moving a mouse, reading pixels, typing into UI elements. It’s a crude stop-gap — necessary for tasks that rely on software without APIs or structured interfaces yet.
But longer term, we need to distinguish between two kinds of tasks.
For the long tail of computer tasks, this operator-style approach is probably unavoidable. But for the common, repeatable tasks we perform every day, it’s both inefficient and wasteful.
Tasks like spreadsheet calculations, drafting slide presentations, or other interactions with productivity tools don’t need to be routed through an AI pretending to be a person. These are better served by headless engines: systems that operate directly on files, data, formulas, and logic — without a UI in the way.
Consider the comparison between ChatGPT’s Agent Mode and an LLM using a dedicated spreadsheet engine for a typical calculation task. Agent Mode spins up a virtual machine, uploads the file, interprets the layout, and tries to imitate a human solving the task — taking minutes. A purpose-built engine calls precise functions, completes the task in seconds, and provides a clear, visual explanation of its work.
Yes, AI will need operator-style tools to interact with obscure software and niche workflows. But for everyday work, we don’t need a Rube Goldberg machine. We need fast, headless systems designed for AI-first use.
The future isn’t about letting AI “use software.”
It’s about building software that’s useable by AI.
Originally published on LinkedIn.