Unite Paris 2026 Recap: 5 Lessons on the Future of Experimentation

On October 1, 2026, we brought the international experimentation community together in Paris for another edition of Unite, at La Felicità, in the heart of Station F.

Zalando, Samsung, Vinted, AS Watson and many others shared their vision for the future of A/B testing and digital optimization.

One thing is clear: AI is no longer just accelerating experimentation. It is changing the rules. Roles are evolving, teams need to keep working together, ideas are multiplying, and the capacity to test them is becoming a precious resource.

In this new landscape, measuring real impact and maintaining control over decisions are becoming two major priorities.

Here are my 5 key takeaways from Unite Paris 2026.

‍

1. AI Does Not Replace Teams. It Redefines Their Role.

At Zalando, experimentation involves around 1,200 employees each year, spanning 14 functions and more than 30 business domains. AI agents are already significantly accelerating their work.

By July 2026, Zalando had already launched as many experiments as it did throughout all of 2025, with no increase in headcount. But for Marcel Toben, Head of Engineering for Economics & Experimentation at Zalando, the transformation goes beyond this productivity boost.

He sees CRO teams increasingly orchestrating experimentation. Product Managers set the objectives, developers build the environments in which agents operate, and Data Specialists ensure results are reliable.

Collaboration becomes easier, execution can be delegated, and human judgment becomes more valuable. Strategy, creativity, ethics, and complex decisions remain central to teams’ responsibilities.

2. The New Challenge: Working with AI Agents.

While a lack of technical resources has long held experimentation back, AI is now removing that barrier.

At Samsung, Kameleoon’s PBX agents are now part of CRO teams’ daily work. PBX Build generates 95% of the tests that previously would have required coding, with an average of just three to four instructions needed to finalize a variation. The Ideate, Configure, and Analyze agents also support teams, from hypothesis development to analysis.

But while tests can be created faster, technical teams want to retain control over the changes being made. The bottleneck has shifted from building tests to putting them into production. The next challenge is therefore no longer about speeding up each step, but about working together.

That is the vision shared by our CPO, Frédéric de Todaro: moving from specialized agents to a system capable of orchestrating the entire experimentation cycle, from the initial objective to execution, with human approval and safeguards.

3. As Ideas Multiply, Traffic Becomes Precious.

AI can generate dozens of hypotheses and create variations in minutes, but the number of visitors available to test them remains limited.

Jean-René Boidron, CEO of Kameleoon, summed it up in his opening remarks: “AI multiplies ideas. Not the traffic available to test them.” Traffic, like the time spent on analysis and decision-making, is becoming a resource to use carefully.

The question is no longer “What can we test?” but “What is truly worth testing?”

This new constraint calls for more careful hypothesis selection, before a test even begins. It is one of the reasons we have been exploring synthetic personas: simulating visitor reactions to help identify the most promising ideas before testing them with real traffic.

‍

4. More Testing Does Not Mean More Value.

With around 300 experiments per month, Vinted is among the organizations experimenting at scale. Yet Jev Gamper, Decision Scientist at Vinted, reminded us of a truth that is too often overlooked: “More tests do not necessarily mean more impact.” Running more tests or releasing more features does not prove their contribution to growth.

It is more meaningful to measure the cumulative impact of changes actually rolled out than the number of experiments conducted. Each validated decision contributes to a shared business objective, such as increasing the number of transactions.

The true measure of success is no longer the number of tests launched, but the value they collectively create.

‍

5. Greater Autonomy Requires Greater Trust.

As AI agents become more autonomous, one question becomes central: how much decision-making should we entrust to them?

At Zalando, Marcel Toben presented four levels of autonomy, ranging from assistance to systems capable of proposing, validating and rolling back their own changes.

This progression rests on three foundations: reliable metrics, predefined rules and full decision traceability.

We share this conviction at Kameleoon: autonomy should not mean a loss of control. We design our agents to act within clearly defined boundaries, with human approval, safeguards and, rollback options.

‍The challenge is not to choose between autonomy and control, but to build the trust needed to bring the two together.

And Because Unite Is Also About Community...

Unite Paris 2026 was more than a series of talks.

This year, we made a deliberate choice: fewer talks, more conversations.

‍

The two roundtable sessions gave participants the opportunity to share their experiences, challenges and ideas with professionals from a wide range of backgrounds. These spontaneous exchanges are what make Unite so rewarding.

Then there is everything that gives Unite its atmosphere: meeting over coffee, conversations that carry on over lunch, thoughtful touches for our guests... and even an unexpected musical interlude, with our CEO, Jean-René Boidron, on the accordion.

‍

It is hard to imagine a better setting than La Felicità, in the heart of Station F. A huge thank you to its teams and the Big Mamma group, which we are also delighted to count among our clients, for their hospitality and energy, and to the Barth & Pearson agency for helping us organize the event.

And of course, thank you to all our speakers, partners and participants for making this edition both inspiring and welcoming.

What Comes Next?

AI enables us to test more. It is up to us to make sure it also helps us test better.