I recently read a paper from Microsoft Research called Working with AI. It analyzed 200,000 conversations with Bing Copilot and introduced a new metric: the AI Applicability Score. The authors frame this score as a way of measuring how AI can augment jobs—not automate them away.

Two things stood out to me.

First, the data. The study used Bing Copilot, which isn’t as widely used as models like ChatGPT, Gemini, or Claude. If the same study were run with a more popular and capable model, would the case for augmentation hold up as strongly? And what about specialized, domain-specific AI models that could reshape certain jobs even more dramatically?

Second, the language. Calling it “AI applicability” sounds much friendlier than “task encroachment,” a term that hints more directly at automation and job loss. The framing felt a little like a soft-sell for Bing Copilot. Not surprisingly, the jobs with the highest “applicability” happen to overlap with the fields where Microsoft is most eager to scale its AI products.

One part I did find helpful was the breakdown of user goals and AI actions—it’s a useful lens for thinking about human–AI interaction. But from my experience in both academia and industry, what’s often missing from these models are the business objectives that ultimately determine whether AI enhances or replaces work. Without factoring that in, we risk ending up with an overly optimistic view of applicability.