If customers will increasingly rely on their own AI assistants, should companies still invest in apps and websites? Spoiler: Yes. But those investments need to build more than the channel your customers happen to use today.

Customers arrive with a situation

For years, companies have worked to make digital journeys easier. We studied behaviour, removed steps, turned complex processes into usable flows. That work matters. Digital experiences are considerably better because of it.

It also taught us to decide everything in advance. We pick which task your customer is trying to complete, which question comes next, which route leads to an answer.

Now imagine your flight has been cancelled. You are travelling with two children, your connection is gone and you need to be home tomorrow morning. You do not arrive with a neat request to rebook a flight or claim a refund. You arrive with a situation.

The intent was always there. What we did was attach it to an action our customer is expected to know. If this happens, open that section. If you want this, take that route. If you have that problem, complete this form. Your customer does part of the organisation’s work before the real service can start.

From fixed journeys to an orchestrated experience

AI makes another way of working possible. A person explains what is happening, and the service works out which actions make sense. A fundamental change.

Someone who listens to your situation and sorts it out on the spot, at eleven at night, without a queue. That used to be a room full of trained people, and only a few companies could afford that. Your customers get it for free now, from an assistant that has nothing to do with you. That is the standard they arrive with.

When I say this, most people will immediately picture a chat window. That makes sense: conversational AI is what made intent tangible for everyone. But capturing intent does not require every experience to become a chat. A tap, a search, past behaviour, an abandoned flow or data the customer has chosen to share can reveal just as much.

This is personalisation, though not the kind most companies have been buying. Until now it mostly meant changing what sits inside a screen that stays the same for everyone. A name in the header, a product recommendation, an offer aimed at a segment. The experience itself never really changed. That kind of personalisation is built from what someone did before. What an interface can use now is what is happening while it happens. And when that is not enough, it can ask, and adapt around the answer.

A generative interface changes the screen itself. Which parts appear, in what order, how much explanation comes with them, whether the next step is a form, a question or a person: all of it can differ from one customer to the next. Segments were as close as we could get before. Now you can design for one person, in one situation.

Take the cancelled flight again. The experience can begin in the airline’s own app and use what it already knows: the booking, the missed connection, the two children, the seats still open tonight. It might offer one clear route home, let you compare two alternatives when the choice genuinely matters, or help you book a hotel when nothing gets you back in time. It adapts and generates around your situation.

Though, you will not remember the screens.
You will remember that someone got you what you needed in that moment.

We will still design the ingredients, the rules and the boundaries. We will be less certain about the order in which any given customer meets them.

Start inside the channels you own

None of this means every company needs its own version of ChatGPT. People still need overview, hierarchy, comparison and a sense of control. A set of buttons with clear CTA’s often beats a conversation.

The work available today is to bring intent and context into the experiences you already have. What does your product already know about a customer that it never acts on? The telco that watches someone run out of data every month. The bank that makes a returning investor start from an empty form. The insurer that asks for documents and inputs before it says anything reassuring.

Four questions tell you how far along you are. Can your app recognise why someone arrived? Can it use context without becoming intrusive? Can it change the next step instead of sending everyone down the same flow? Can it explain why an option is on the screen, ask an additional question or hand over to a person when its confidence drops?

Those are questions about trust more than technology. Technology can find the patterns. It cannot decide what a good relationship should feel like, or what a customer should still be able to describe six months later. That stays an organisational choice.

The interface is already moving

I still remember asking my father for his credit card to order a T-shirt from a band’s website in 1998. He was sure the account would be emptied. The technology worked, yet his confidence had not caught up, and it took years. It will not take years this time. People already talk to an assistant every day, and they will not need convincing to let one rebook a flight.

Their assistant becomes the place where it starts. They will ask it to compare offers, change a plan, make a booking or open a claim. At times you will deal with the customer’s agent first and with the customer afterwards.

The foundations for that world are already being built. Google’s AppFunctions, still an experimental preview, lets an Android app expose selected functions to authorised assistants and agents, so your capability becomes something another system can call. Runway’s Solaris research looks further ahead: an interface generated frame by frame as someone interacts with it. Early work, with real limitations, but the direction is hard to miss.

A screen generated as you use it never existed before you arrived and will not exist again after you leave. No two customers see the same screen. There is no artefact to ship, to audit, to hand to a developer. You design the rules it gets made by, and never the thing itself.

That is the sharpest version of the question this article started with. Paying for something that will not exist tomorrow goes against everything we were taught about investment. That discomfort is the shift. What you fund now is whatever assembles the experience, and it outlives every screen it makes.

Build what can travel

Nobody knows which assistant, device or interaction pattern wins. Predicting the winning interface is a poor foundation for a strategy meant to hold five years. The reusable parts are on every architecture slide already: identity, consent, payments, the route to a human being, …

That is the easy half. The capability that decides whether any of it works is harder to buy: being able to run a service whose shape you cannot fully predict.

Today you approve screens. From here you approve rules, boundaries and the range of what the service may decide on its own, because nobody at your end sees the exact answer a customer gets before they get it. It is the same capability you need when that customer only reaches you through their assistant. You will not control the wording, the order or the moment. You control what your service is willing to do, and what it does when it is unsure.

What travels is what they remember

That is why the channels you own still matter. They are where customers learn what you are like, where trust is earned, and where you can practise this way of working before it becomes the only way.

One thing outlives the interface and you cannot build it. An assistant can handle a request perfectly and leave nothing behind. A customer who remembers nothing about you has no reason to ask for you by name, and no reason to give your name to anyone else.

So keep investing. Improve the channels customers use today, and make sure every investment also strengthens your ability to understand intent, use context, and assemble a response you did not draw in advance. First inside your own channels. Later through the channels your customers choose.

The journey will become harder to predict. The quality of the relationship does not have to be.