Open Google Maps on your phone and, next to the search bar, there’s a button marked “Ask Maps”. Tap it and you get a text box and a cursor. Google shipped it in March, first in the US and India, alongside what it called the app’s largest update in more than ten years, and Miriam Daniel, who runs Maps, put the pitch neatly: instead of research and sifting through reviews, “you can just tap the ‘Ask Maps’ button and get your questions answered conversationally”.

She’s right that it’s better. It’s also the moment I’ve been waiting on for a decade, and my first reaction in front of that empty box was: ask it what? Which is when it occurred to me that menus were doing two jobs all along, and we only ever complained about one of them.

The first job is navigation — getting you from where you are to the thing you want. That’s the job we resented, and it’s the one AI does better. Nobody misses drilling four levels into a settings panel to find the thing they use every Tuesday. Back in 2016 I showed clients a slide whose benefits bullet read “no menus or UI to navigate — just ask”, and I meant it as an unambiguous good. I got most of that prediction wrong, but never that bit.

The second job is less obvious and much harder to replace. A menu is an inventory. It’s the closest thing most software has to an honest account of what it can do, sitting there in public, whether you use it or not.

That’s how most of us learned the tools we’re good at — opening menus we didn’t need, spotting a command we hadn’t known existed, filing it away for later. Not the only way we learned, but the reliable one, the one that worked without anybody teaching us. The menu showed you the shape of the thing.

Take it away and the software can do far more while the person using it knows far less about what it can do.

The evidence is already in

Nielsen Norman Group ran usability sessions on Amazon’s AI shopping assistant, Rufus, in early 2025. Participants liked it once they’d been shown it. It knows what you’ve been browsing, it answers real questions, and one participant, shown it mid-session, said “this is dangerous” because it would make her buy more — presumably what Amazon was hoping for.

None of them found it on their own. Not one. The researchers had to point it out, and that included people who reported using Amazon daily. Nor did anyone touch the AI prompt suggestions Amazon had tucked under the search bar: not a single participant clicked one voluntarily across the whole study. One of them explained why he’d never type a question into the search box. “Because that’s where you find products.”

That sentence is the whole problem. The capability was there, on the screen, free, and nobody’s mental model had a slot for it.

Those participants deserve some sympathy, because we’re all going to be them shortly. Nobody arrives at a blank box carrying a list of everything the software could do for them. Keeping that list was the software’s job.

What this does to product design

Discoverability is now a teaching problem. You used to solve it by putting the feature where people look, which stops working when there’s nothing to look at. The product has to volunteer what it can do, at the moment it’s useful — the accounting tool noticing you’ve exported the same report to a spreadsheet three months running and offering, right there, to make it a recurring one. That’s a much harder brief than “add it to the File menu”, and it’s the brief now.

Habits outlive the routes that made them. The Amazon participant wasn’t ignorant, he was practised: he had a way of getting things done and no reason to abandon it. People keep taking the old route for a long time after a better one appears, and when the old route is removed rather than replaced, they mostly just do less. So the teaching has to happen where the old habit lives, not in a welcome tour nobody reads.

The gap between your best and worst users stops being capped. Two people on the same licence, paying the same money: one asks the AI to tidy up an email, the other has it read six months of client notes and draft the renewal proposal. That difference used to be bounded by the menu, because the menu showed everyone the same list. Unbounded, it shows up in your usage data as a handful of people getting enormous value while everyone else uses the tool for the first thing they thought of.

Which leaves the suggestion doing the work the menu used to do. I tried to build something along these lines years ago — a dashboard that told you what to do next rather than showing you numbers and leaving you to work it out. The technology wasn’t ready. It is now, and it’s where the design effort has to go, because a blank box with a friendly placeholder isn’t a substitute for a menu. It’s an exam.

Whether that actually works, I’m not sure. Amazon’s prompt suggestions were in-context suggestions, and nobody clicked them either. It’s possible we’re about to spend a decade building recommendation layers that people ignore just as thoroughly as they ignored the sparkles icon, and I don’t know how you’d tell the difference in advance.

Some of this will sort itself out. Nielsen Norman make the reasonable point that mental models shift — their youngest participant went looking next to the search bar for the AI, because that’s where Instagram put it. Give it five years and “ask the thing” may feel as ordinary as double-clicking. I’d design for the five years rather than the destination, though. I made the same bet in 2016, was about six years early, and the products that shipped into that gap mostly didn’t survive it.

What to do in the meantime

Watch what your team actually asks your tools for when nobody’s looking over their shoulder. A vendor demo won’t tell you that; your own logs will, and it’s the cheapest piece of research you’ll ever run. My guess is you’ll find three or four prompts doing all the work and a lot of expensive capability sitting untouched.

Then, when you’re evaluating anything with an “ask me” box in it, put a blunt question to the vendor: how does a new user find out what this can do? If the answer is “they’ll pick it up” or “there’s documentation”, you’re buying a capability your team will use at a fraction of its value.

For forty years the interface was the product’s honest account of itself. We’re taking it away because it was slow, and it was. But something still has to tell people what the software can do, and until we work out what, the most capable tools we’ve ever built will keep getting used for the three things everybody already thought of.