The future of UX design: designing for AI agents and beyond GUIs

6 min. read

Podcast

Constantine Gavrykov, Senior Director of Product Design at Decathlon.

Constantine has spent more than 20 years in product and UX design. After moving from Ukraine to Amsterdam, he built the team behind TommyHilfiger.com and worked on Calvin Klein's mobile apps, then moved through Bestseller, Intersport and Adidas before joining Decathlon.

Today Constantine is Senior Director of Product Design at Decathlon, the world's largest sports retailer, where he connects emerging tech trends to a design vision for the business. He has also written a three-part series on the shifts he sees reshaping digital: autonomous agents, hyper-personalisation and generative UI.

We got into designing for humans and AI agents at the same time, why personalisation-first is replacing mobile-first, what AI overviews are already doing to organic traffic, and where brands show up once agents skip the interface.

Watch Delivered with Constantine Gavrykov, Sr Director of Product Design at Decathlon.

Key takeaways#

How do you design for AI agents as users?#

Design for dual interactions. People still click, tap and scroll, so they still need good UX. Agents read the page differently: they rely on the DOM structure. Constantine's advice is to go through your touchpoints, work out which ones humans hit and which ones agents hit, and design for both. He calls the emerging practice agent-responsive design, a direct parallel to responsive web design. It's already relevant. A study of 3,000 websites found that 63% already receive traffic from AI.

Try this: pick your 5 highest-traffic pages and check how they read with styling switched off. That's roughly what an agent gets.

Why is personalisation-first replacing mobile-first?#

Because talking to a system in natural language feels more natural than working through a graphical interface. Constantine sees most of today's AI work as making horses faster: the same mobile-first interfaces with AI added on. A real personalisation-first world needs 2 things. People have to adopt AI assistants that know them, from what they ate yesterday to how they train. And brands have to make their data systematically available to those assistants, beyond an API or a page to click through.

Try this: ask your team how a customer's own AI assistant would get your product data today. If the answer is scraping your site, start there.

What is hyper-personalisation in e-commerce?#

Constantine describes 3 tiers. The first works from clicks and behaviour patterns. The second needs a login and purchase history: you bought running shoes, so we know whether you run marathons or jog casually. The third is hyper-personalisation, when an agent has access to your personal data and preferences. At that level a retailer can sell a package of services: the shoes, recurring nutrition supplements, a training plan, a running community. Content can adapt too. If he likes running on the beach, the lookbook shows the shoes on a beach. Fit is the other big one. He calls size and fit a notorious problem the internet never solved, and sees computer vision combined with generative AI as the way through.

Try this: map one customer journey across the 3 tiers. Note what data you'd need to reach the third, and whether you'd be comfortable asking for it.

Which UX patterns stop working when agents browse for you?#

The ones that slow down getting to information. Infinite scroll, modals and pop-ups get deprioritised by agents because they get in the way. Constantine suggests plain pagination as a near-term fix, since it's easier for an agent to scan. The end state is agents reading raw data directly. His view on graphical interfaces is direct: we built them so people could operate computers, and they're going to be a thing of the past.

Try this: audit your key flows for anything an agent has to dismiss or scroll through to get to the product. Each one is friction for both audiences.

How are AI overviews affecting organic traffic?#

They're cutting it already. Constantine points to publications such as HuffPost and Business Insider seeing 40% to 60% less organic traffic, driven by Google's AI Overviews and ChatGPT's web browsing. Georgios shared his own number: around 80% to 90% fewer click-throughs since AI overviews rolled out. He also told the story of a dryer purchase at home. ChatGPT compared 10 options and gave buy links. No one in the house saw a landing page, a brand story or a product detail page before checkout. Constantine's read: this is happening right now.

Try this: split your organic traffic reporting by query type and track the informational queries separately. That's where AI answers bite first.

How do you keep humans in control when agents act on their behalf?#

Design the human back in at the sensitive moments. Constantine calls it human-in-the-loop design: the agent does the work, and the person gets control back for things like entering payment details or confirming a purchase. Privacy is the open problem. An agent filling in forms handles phone numbers, addresses, card and bank details, and that data travels to be processed elsewhere. He sees direct integrations, such as with Shopify, and payment companies like Visa and Mastercard building for agents as ways through. He also expects convenience to win, and people to hand over more control than they'd admit today.

Try this: list every step in your checkout where a customer's agent would need personal data. Decide which ones should always hand control back to the person.

Where do brands show up when agents skip the interface?#

Outside the screen, mostly. If an agent picks the product, the customer may never see your landing page. Constantine expects first impressions and brand stories to move to social media, TV, billboards and physical brand activations. He also sees a new idea coming: branded agents, an evolution of the brand ambassador. Nespresso has George Clooney. A future version might license a voice, a likeness and mannerisms into an agent that speaks for the brand.

Try this: ask what your brand would sound like as an agent answering a customer's question. If the team can't describe it, your tone of voice needs work.

What should leaders do first about AI?#

Build AI literacy, then set a vision. Constantine wants people upskilled enough to ask the right questions: how does this change customer shopping habits, logistics and the value chain? The common mistake is grabbing AI and pointing it at any problem. He starts from customer pains and gains instead, and treats AI as a secondary measure. His cautionary example: a company that spent 4 months with 25 people automating a task that took 2 hours a quarter. He also cites BCG's advice to make the AI agenda business-led, let business units experiment rather than centralising everything, and have IT and AI teams scale the foundations and governance. And give people time and space to use AI in their daily work.

Try this: before approving the next AI project, write down the hours it will save per quarter next to the hours it will cost to build.

How does AI improve digital accessibility?#

Constantine calls it a massive leap already. A graphical interface asks you to learn how to use it. Chat and speech don't, because pretty much everyone knows how to speak. Computer vision is his favourite part: point your camera at something and ask what it is, how it works or what's wrong with it. He notes it isn't fully available in Europe yet. Georgios added an example of AI changing format for the listener: he turned a 50-page McKinsey report on AI into a podcast with NotebookLM.

Try this: test one key task on your site using voice only. Note every point where you need to see the screen.

What comes after AI agents in UX?#

Brain-computer interfaces, in Constantine's view the biggest leap. He also lists robotics, which he thinks is arriving faster than people expect, along with biotech, personal medicine and quantum computing. He's sceptical of the invasive approach Neuralink takes, with implants in the brain, and asks how anyone can be sure a device installed today will serve them for life. The real step, for him, is non-invasive: a headset or cap that reads brain activity. He'd rather wear a cap than have holes drilled.

Try this: pick one emerging interface, voice, vision or wearables, and prototype one interaction with it this quarter.

Who is responsible for the ethics of AI-driven design?#

Designers, leaders, engineers and entrepreneurs. Constantine puts the responsibility on the people building these systems. That includes demanding that governments regulate AI content, so people can tell what is AI-generated, AI-tweaked or original, and avoid falling for deepfakes. He also wants people to keep a choice. If someone feels nostalgic, they can ask ChatGPT, Gemini or Grok to show information as a classic website.

Try this: add one line to your design principles about how you label AI-generated content. Decide it before a campaign forces the question.