Personalization

Earlier, I shared about Discover, our proactive AI assistant within Copilot that delivers highly personalized content based on deep user understanding. It’s a simple concept: Get to know your users so well that your system knows what they want to see before their finger even hits the keyboard. But how do you do this effectively? Pull back the curtain, and things get a lot more complicated.

Personalization is one of those problems that sounds solved until you try to do it well.

Most systems today are good at the easy version: Your user clicked on running shoes; here are more running shoes. Collaborative filtering, embedding similarity, engagement-optimized ranking? These are well-understood techniques. They work. They’re also shallow: they model behavior, but not the person behind it.

The hard version is: Your user is training for a marathon, cares about gear but is price-conscious, prefers video instruction over articles, and is more engaged in the evenings. Six months from now their interest will have shifted - either they ran the marathon or they didn’t, and the system needs to know that without being told.

That’s not a ranking problem - it’s a user modeling problem. It requires maintaining a persistent, structured representation of a person that updates continuously from heterogeneous signals across devices and contexts. Confidence estimation over inferred interests. Temporal decay that distinguishes a dormant passion from an active one. Goal inference from behavioral sequences, not just topic classification from click streams. Real-time signal fusion, not batch recomputation.

And then the product constraint that makes it genuinely hard: The user needs to be able to see what the system believes, understand why, and correct it; and those corrections need to propagate immediately and visibly. Explainability isn’t a nice-to-have. It’s a forcing function on the architecture.

Personalization is more than the shoes - it's the whole story.

There is a lot of work ahead to realize this in a comprehensive and privacy-safe way. The ambition here is genuinely large. If you want to work on personalization that actually means something, talk to me!