.webp)
There is a difference between adding an AI feature and rethinking how a product works. It is easy to do the first and assume it counts as the second. That mismatch has real consequences: for architecture, for cost, and eventually for the product's competitive position.
Why mobile app interfaces are changing
I can see that something more fundamental than chatbots is happening around me. Users are changing how they relate to digital interfaces altogether.
For the past decade, mobile apps have taught users to think structurally: open a menu, tap a category, filter results, complete a form. The assumption built into every navigation flow was that users would invest time in learning the app's own logic. Many did. But in the AI era, expectations are resetting.
People who interact daily with voice assistants, AI-enhanced search, and large language model tools are developing a different instinct. State your goal and let the system figure out the path. This is a behavioural shift already under way outside your app, independent of whether your product has caught up.
Recently, I noticed it with Revolut. I have been using the app for nearly a decade, and one thing has always stood out: the product team keeps rethinking how interaction should work. Right now they are testing what happens when the interface itself stops being necessary. If it works, it will probably influence how other banking products evolve too.
In complex, context-heavy products, interfaces that force users through a predefined structure are starting to feel like friction rather than familiarity. Not everywhere, and not all at once. But in specific contexts such as financial planning, product discovery, healthcare guidance, and utility configuration, users increasingly want to just say what they need, while the interface keeps routing them down a predefined path.
Conversational AI or structured UI
Here is where realism needs to come in, because the temptation to bolt a conversational layer onto every product is strong and mostly unjustified.
A well-designed structured UI is still the better choice for any linear, predictable task flow. If a user needs to top up a prepaid card, check an invoice, or update a delivery address, a clean native UI with two taps and a confirmation screen will beat a chat interface every time.
Conversational layers introduce latency, ambiguity, and failure states that structured flows simply do not have.
The case for a conversational layer is strongest when several conditions hold at once:
- Product maturity. A conversational interface needs something to draw from: a data model, a content layer, a history of what the user has done. Without structured, queryable data, the AI has nothing to personalise and no context to reason from. Revolut already shows what this looks like. Land in another country and its Pay-per-Day travel insurance can switch on by itself, triggered by location, with no screen to open and nothing to confirm. The app stops waiting for input and starts acting on context. For a product with years of transaction and travel data behind it, that is a small step. For one without it, it is not possible at all.
- Genuinely complex, non-linear choices. Where users face decisions with many variables, natural language removes friction that navigation menus create. Choosing a financial product, a health plan, or a configuration across dozens of parameters is the obvious case. The user should not need to learn the categories. They need to describe their situation.
- An established user base with real behavioural data. The quality of a personalised AI interaction is proportional to the quality of the available signal. Without it, you have a generic assistant that could have been built for any product, and users will treat it that way.
If those three conditions are not met, traditional UI wins. The goal is never to add AI. It is to reduce friction and improve outcomes. Sometimes that means a better information architecture, not a language model.
How we built Rossi for ROSSMANN
The ROSSMANN CLUB app is a project we have been developing since 2019. What started as a loyalty app and web has grown into a personalised retail ecosystem used by over 1.2 million customers.
The challenge that led to Rossi is familiar to any retailer with a large catalogue: reproducing, at scale, the guidance a knowledgeable shop assistant gives in person. In a physical store, a staff member asks the right questions and points you to the right product. In a digital product, filters and recommendation carousels approximate that, but they are reactive and impersonal by design.
Rossi, the virtual advisor we built into the app, was designed to do exactly that.
It runs on a RAG architecture over the full product catalogue and handles product discovery, personalised recommendations, stock availability, and loyalty queries. It is scoped strictly to ROSSMANN context, so it will not answer questions outside that domain or recommend competitor products.
The decision that made it work was treating Rossi as a core product feature, not a bolt-on. It has access to the catalogue, purchase history, location, and contextual signals. Within the first two weeks after launch, queries increased by 80%, with around 50% leading to a same-day purchase.
Where the AI actually runs
Not every AI feature needs the cloud. For bounded tasks over local content, on-device models are now capable enough to do the job with no network call, no latency, and no user data leaving the phone. That covers a lot of real cases:
- summarising content that is already on the device
- answering questions from data the user already has
- classifying something against a known user profile
That last property, keeping data on the device, matters a lot in regulated markets. My colleague Antonín “Tony” Šimek has written separately about what building on Apple's Foundation Models taught him, including where the local approach holds up and where it still hits a wall.
The short version: for complex reasoning, long documents, and open-ended conversation you still reach for the cloud, but a growing share of useful features no longer needs it.
The practical pattern in production is usually a hybrid. A capable cloud model for the complex, high-value interactions, and smaller or local models handling the volume. Treating every query as if it needs the largest available model is neither necessary nor affordable at scale.
The post-navigation direction
This is closely linked to the rise of Super Apps. There is often a step before that. In practice bigger companies rarely have one app to maintain. There is a loyalty app from one team, a payments app from another, something shipped externally two years ago that nobody really owns. Pulling those into one coherent product comes first.
Once you do consolidate, and then keep adding payments, mobility, and lifestyle services on top, the navigation gets impossibly deep. I have seen this firsthand: at some point no menu structure can hold it together. An intent-driven agent becomes the layer that keeps the whole thing usable.
We see the same pull closer to home. With ČSOB, we have been steadily integrating its AI assistant, Kate, deeper into the core experience, letting users manage payments, cards, loans, or insurance through natural voice and chat instead of navigating traditional banking flows.
The infrastructure is moving the same way. Google recently introduced AppFunctions for Android, which effectively turns app logic into an API that agents like Gemini can call directly.
You can already see early versions of this in map apps. The agent skips the interface and runs the routing logic directly, without ever opening a screen. Navigation is leading the way, and plenty of more complex applications will follow.
The human side of this is stronger than people expect. My mother-in-law is 69 and already prefers controlling her phone by voice over digging through menus. She is not an exception. When the experience is reliable, this kind of interaction removes friction for a wide range of users.
I don't think traditional interfaces will vanish overnight. As an alternative to them, though? Absolutely.
There is no need to remove buttons tomorrow. What matters is understanding where the market is heading, and starting to move in that direction before it forces the question.
The hard part is building a product where AI makes the experience meaningfully better than what was there before. That takes businesses longer and costs more upfront than it looks. A good intent-driven layer is something you then operate for years and your users will appreciate it.
We tend to think about these considerations as the team that stays with the product afterwards, not the one that hands it over at launch.
.webp)
.webp)
.webp)
.webp)
.webp)
.webp)
.jpg)
.jpg)
.png)
.avif)
.avif)
