Tired of AI? Put it under the hood

You know when you design some feature and think: it would be so good if we knew something about the user right here. We could skip a step, say something that actually applies to them, give them a custom UI, nudge them towards their use case.

There's so much potential in personalisation, but it's also so hard to pull off. We need some data about the user (usually we try to collect it during onboarding), and mostly we either don't have it, or the data we have is too messy to use for the solution we want.

So usually we shrug: would be nice, can't do it, let's live with the generic view.

But hey, AI and everything, we can actually do all of this now pretty easily.

We're having the wrong AI debate

Please hold for this quick interim rant.

When designers talk about AI right now, we mostly talk about process: Figma MCP, Claude Code, prompt-to-design, which workflow to learn before it's too late. I care about those tools a lot, but this entire debate is so extremely self-centered: how we work, how fast we produce mockups, and whether we're still needed. This has nothing to do with the actual job to be done:

Solving real problems for real people.

So I wonder: What can we design now that we couldn't design before.

LLMs as design material

An LLM inside your product doesn't have to be a chatbot. Nobody needs to type a prompt or see a sparkle icon. It can sit under the hood and quietly turn messy data into something you can design with.

Take the classic tooltip tour. We know users skip them, and still, sometimes a feature really is worth nudging. Now the nudge can know the user: "You create about ten items every week, one by one. There's a bulk add, here's how it works for your case." I'd actually read that tooltip.

Or the empty profile. Apps want filled profiles and users don't fill them. But the user already told you who they are, in their freeform input and in the way they use the app. An LLM can turn all of that into structured profile data while the user just uses the product.

Or: "Untitled (37)". Naming, tagging and sorting is housekeeping that users skip and then suffer from, and all of it can now happen from the content itself, silently.

Even the signals you already have become usable. A user who creates one item per month is telling you something about their activity level. That information has been sitting in your timestamps all along, and now you can read it and act on it.

None of these problems stayed unsolved because we didn't know the fix. They stayed unsolved because the fix needed data we didn't have, so we asked the user, badly. Or it needed content written by hand for every single case, so everyone got the generic version. LLMs remove both constraints.

And to the user, none of this looks like AI. The product just seems to get them.

Read on, how this can look in practise:

Reference

15 classic UX problems you can now just solve

A growing list of UX problems that stayed unsolved for years, and how an LLM under the hood solves them.