UX + AI Principles

Flat information architecture

The first idea I want to throw out there is flat information architecture. With artificial intelligence you don’t need a hierarchy of panels and menu entries to cross in order to reach what you need, whether it’s a document, a button, or the launch of a routine. You can pull things up by asking directly for what you want. There’s no need for a hierarchy.

That concept of the shelf with the documents on it, and inside the document you find the sub-document, the subfolder, this is stuff to abandon in favor of a flat information architecture with direct access and no navigation. Navigation only comes into play when the user can’t find what they’re looking for. And often they don’t even have to go searching for what they need, because we’re in a phase where the user’s experience with AI isn’t searching for a thing you need, it’s getting an answer, with a system that goes and finds the sources needed to answer that question. So I’ll put it provocatively: not only do we have to abandon the hierarchy of information, the searching and the navigating within information, we have to abandon the very concept of search and move to the concept of the answer. Easier said than done, of course.

Mixing conversation and graphics

Although the graphical interfaces proposed by the likes of ChatGPT and Claude are “you chat, you do everything through chat”, you’re starting to see very interesting experiments where the conversation and the classic graphical UX are used at the same time. You can give input and ask the system questions both by writing and by clicking or moving sliders.

And it’s interesting, in my opinion, to understand how you orchestrate an interface that’s both conversational and graphic, old style. For example, you have a bar at the bottom where you write what you need, and instead of the WhatsApp-style, classic-ChatGPT-style list of messages, you have an interface that stays put, fixed, visual, above the box where you type your messages, and that upper part changes based on what you write and offers you buttons to react directly with it. Or it’s a matter of still having a conversation where, message after message, the bot sometimes replies with graphic panels instead of text messages, and in those graphic panels there are buttons.

These graphical interfaces are typical, I’m seeing them a ton in enterprise agent products, like Microsoft’s Copilot 365 or Google’s Agentspace. So the big players are experimenting hard on this too, and it’s all still to be figured out. In my view there’s a lot of interesting stuff here to see and to do. In fact, when I hear “AI takes work away from programmers, it takes work away from artists”, no, it doesn’t take work away from anything, folks, if anything it finally spares us from doing the things we don’t feel like doing.

Human in the loop

The third point concerns the so-called human in the loop, a concept that’s pretty widespread by now, which expresses the need to have a human decision-maker inside the decision cycle of the AI, of our automation. And this human in the loop is all the more necessary the more sensitive the decisions are. There has to be a human upstream, deciding for instance which automation gets launched and which doesn’t, or downstream, that is, after the automation has run, when I want to validate the system’s answer. For example, I’m a doctor and I want to validate the diagnosis an automatic system made, because the responsibility is mine as the doctor.

Or at the intermediate points: the automation starts and every so often comes back to me saying “I got stuck at this point, I need a human choice here.” And this is even more interesting, having automations that from time to time come back to you and ask what to do because they need your opinion. This is another under-explored aspect, extremely interesting, and it’s so important that even the protocols now emerging in the deeper technology of artificial intelligence provide for defined communication modes for the bot, so it can come to you and ask for information when it needs it.

There are, in fact, plenty of middle grounds between full automation and doing things entirely by hand. There’s a gradient. This gradient has to be explored based on the use cases, based on the vertical, based on the situation of whoever’s using the system. And thinking about full automation is just as childish as thinking about using no automation at all. This slider was also mentioned by Karpathy, a great AI researcher, in one of his latest talks: we should imagine a kind of slider where you say, okay, I want this thing fully automatic, or less automatic, and everything in between.

The structured versus unstructured continuum

Beyond this automation gradient, I’d like to raise the point of another continuum, the one between structured information and unstructured information, free conversation. Sometimes we want structured information that has to be exactly that. When I make a booking I definitely want to converse, but there’s a point where I have to see exactly the data I’m talking about, I have to be aware of the various stages of the booking, of my trip, of what I’m doing. Imagining user experiences tied to AI that are at once fast, that ask for as little input as possible from the user, but that are also as solid as possible in making you give the OK at the points where it’s needed, only at those points, and that guide you through the different phases. In some cases you can speak freely, in some cases you have to say exactly what you need, and there maybe it’s not so much writing and talking, it’s precisely selecting, or getting it and pressing the specific button.

This, in my view, is extremely interesting. In the Cheshire Cat, for those who’ve never heard it named, in English it’s called Cheshire Cat AI, an open source project born in Italy to build AI agents, and it’s made specifically to be invisible, in the sense that you can use the Cat inside third-party applications, say inside the invoicing management software, inside the fridge’s control panel. With the Cat we experimented a lot with the guys, with the contributors, we tried to imagine interactions that are conversational but goal-oriented, where the bot keeps a very precise structure in its memory, for example the pizza order. That information has to be reduced, at some point, to precise data, and it has to be validated. For instance, conversationally I say I live at Via Peppino 82, but then the system has to be able to take that address, convert it into coordinates, and the bot has to know the latitude and longitude, whether they fall exactly within the area it can serve, because otherwise I’m in Torbellamonaca and they order the pizza from Madagascar, and that can’t be done.

These conversational formats are made, first of all, to know in advance which data the AI wants to collect. And it’s the AI that asks you, and then it also gives you the final confirmation. In asking you for that final confirmation, it can show you a panel with the status of your booking. You say OK, or press an OK button, and the pizza order actually goes out.

The long tail begins

To recap, folks: the fanboy moment, the cult of yet another model, in my view, is finally over. Now we can dedicate ourselves to the long tail, a period in which we design, we create, and we make available our own artificial intelligences, our own agents, and our own mixed interfaces, conversational and graphic, which are truly fascinating.

My appeal: share this video with your friend who works in design and is passionate about AI, your programmer friend who’s starting to tinker with agents, your friend who wants to launch a startup. And what do you all think about this stuff? Where is it discussed? We could create a space to go deeper on this kind of conversation and move into a phase where we actually have fun. Like, share, and leave me your number.