e15.cz: AI Week Milan 2026: From Fragmentation to Orchestration
The biggest European AI event of the year, AI Week in Milan (19–20 May 2026), brought together the founders of key platforms and global business leaders in the middle of transformation.
A cross-section of the main presentations offered a picture very different from the superficial enthusiasm around chatbots. The questions are no longer about whether and how to deploy AI. They are about how to rebuild companies so they can function in a world shaped by artificial intelligence.
What is at stake is a fundamental rebuild of how organisations operate, how software is created, and where value actually lies.
From Tools to Architecture: The Problem of Organisational Silos
The strongest theme of the conference had nothing to do with specific AI models. Alex Ball, Vice President of Genesys (a global customer experience platform), named the core paradox: “Are you scaling intelligence, or are you scaling fragmentation?” Most large enterprises today have dozens of AI tools, each isolated in a different department. This is an architectural problem. Systems were designed so that each one solved a single workflow. However, AI is by nature dynamic and fluid. It works best when rigid boundaries between systems do not exist.
The solution is an orchestration system with memory across the entire interaction and visibility of all agents simultaneously. Ball added a landmark data point: 2026 is the first year in which more than half of customer intent comes from channels companies do not own: YouTube, Reddit, ChatGPT. He describes the shift from workflows to outcomes as “a tenfold increase in business value.”
Four Mistakes Companies Make When Deploying AI
Claudio Ricci from TIM Enterprise (Telecom Italia) named four systematic mistakes. First: “pay and pray” — buying licences without an implementation strategy. Second: excessive ROI focus. “ROI measures the value of today’s technology, not the value of the knowledge you gain along the way.” A creative agency that calculated ROI on AI video in 2023 would have refused to invest. The one that tried anyway had a two-year head start by 2025.
Third: deploying new tools into old processes — “like putting a Ferrari engine into a horse-drawn carriage.” Fourth, and most critical: not changing the organisation. “If you quadruple someone’s AI co-workers overnight without changing processes, that person becomes the bottleneck.” According to Ricci, 20% of companies currently capture 75% of the value from AI. The gap is widening.
A strategic framework was offered by Scott Likens, Global Chief AI Engineer at PwC: “Don’t have an AI strategy. Have a convergence strategy. Breakthroughs come at the intersection of technologies.” Conway’s Law is the key: technological architecture mirrors organisational silos. AI is the first technology capable of breaking through that pattern. The measure of success must also change: “Return on Experiment matters more than Return on Investment. Short cycles, one to four days. Not fail fast, but learn fast.”
Agentic AI: From Snapshots to Real Time
The difference between traditional ML and agentic AI was illustrated by Francesca Fortunato and Angela Sebastianelli from VAR Group through a retail case study. Traditional ML works with data snapshots and reports results with a delay. Agentic AI “lives in the stream of customer events. It perceives in real time. It responds immediately.” The key premise: “A customer as a stable, coherent and predictable entity does not exist. What exists is their behaviour.” One person can be a loyal customer in November, a deal-hunter in July, and completely inactive the following year.
Sebastianelli defines four capabilities of an agentic system: perception (real events, not snapshots), decision-making (weighing alternatives), action (personalised to a specific customer at a specific moment), and coordination — the work of a conductor. Inaction always figures among the options as a conscious decision to do nothing. On governance she adds: “Compliance does not slow development. It delivers competitive advantage.”
The gap between agent text output and how people actually consume information was raised by Rong Yan, CTO of HeyGen (used by 80% of the Fortune 500): “Text works for agents, but people prefer video.” HeyGen Avatar 5 (April 2026) generates unlimited video in unlimited languages from just 15 seconds of recording. Project Hyperframes (open source, 20,000 GitHub stars in three weeks) integrates video generation into any agentic workflow. The next billion videos will not be limited by computing power. They will be limited by the message.
Democratisation of Creation: The End of the 0.6% Monopoly
The visually strongest argument came from Aino Bergius of Special Projects at Lovable (valuation 6.6 billion dollars): a grid of 2,500 squares, each representing 3.2 million people. Just 14 orange squares — 0.6% of the world’s population — had been deciding since the nineties what digital software looks like. The barrier was structural: years of learning, capital, a team. “Most people never even tried to start.” Today Lovable counts 40 million projects and 600 million visits per month. Sabrina from São Paulo built an app for checking criminal records without any technical background. A sales leader at Uber Eats created a presentation generator without waiting for IT.
Michele Catasta, President and Head of AI at Replit (valuation 9 billion dollars), described the same philosophy from a product design perspective. In conversation with journalist Karen Hao (author of Empire of AI) he said: “We go through every pixel asking whether it is truly necessary. I don’t want users thinking about which model to use.” He described a mother who built a spelling app for her child with dyslexia as an expression of where the field is heading: “Models are far ahead of products. Most visible products target elites. I would like to see technology in everyone’s hands.”
From Executor to Supervisor: Three Phases of Team Transformation
Luca Mastella from Learnn added the organisational dimension. Three phases of team transformation: from executor to orchestrator (AI takes over routine tasks), from orchestrator to supervisor (AI produces output without prompting, the human checks quality). His warning: “If we save time through AI but do not point it in the right direction, we simply fill our calendars with pointless tasks.” AI-native companies as data context: Italian Lexroom grew from zero to 12 million ARR in 18 months. Gamma reached 100 million ARR. Base44 sold to Wix for 80 million in six months with a single founder.
Giada Franceschini from Boosha AI named the practical intersection of democratisation and automation: “Programming today means composing actions, not writing code.” Three levels of working with AI agents — description, skill (a formalised competence triggered by a single command), and routine (automation running without the user’s presence) — are accessible to anyone in an organisation. “If I do something more than twice, I build a system that does it for me.”
The Natural Limits of AI: Empathy, Journalism and Post-Transformer Architecture
Where AI naturally runs into its limits was illustrated by digital mental health clinic Serenis. Its CEO Silvia Wang worked with Ipsos survey data: 50% of young people use AI to talk about personal problems, citing the absence of judgement as their main reason. After deploying AI to eliminate administrative tasks, the result was captured in a therapist’s words after six months: “It is easier to talk with AI because it does not judge you. But AI also agrees with you. And that is where the crucial difference lies.” True understanding begins with challenge, not agreement. “The risk is not that machines will become like us. The risk is that we will slowly become like machines.”
The impact of AI on the media ecosystem was mapped by Nic Newman, Senior Research Associate at the Reuters Institute at Oxford. Social networks and video platforms have overtaken television as the primary source of news for the first time in history. AI chatbots for news are used by 7% of the population (14% of young people), doubling year on year. Publishers expect to lose half their search traffic within three years. Google referrals have already fallen by a third. The response is a focus on distinctly human formats: field reporting, narrative journalism, community and live events. Podcasts are growing because audio and video are harder for AI to repackage. BBC uses AI regional accents for sports briefings as an example of personalisation that would have been economically impossible without AI.
The Man Who Invented the Transformer Says Its Era Will End
The least expected perspective came from Llion Jones, CTO and co-founder of Sakana AI in Tokyo and co-author of “Attention is All You Need” (over 250,000 citations). That paper underpins ChatGPT, Claude and Gemini. The paper emerged from a hardware problem. Recurrent neural networks were slow on new TPU chips. Removing the recurrence produced a thousandfold speedup. Jones frames the lesson as a challenge: “The real breakthrough was in the hardware lottery: we designed the architecture to match the hardware.” His conclusion surprised the entire room: “Stop rearranging components. You will not find the next breakthrough by reshuffling existing parts. The next big thing will not look like a transformer at all.” Sakana AI is explicitly working on post-transformer architecture.
A Cross-Sectional View: Three Axes of Transformation
The ideas from AI Week Milan 2026 can be read as three axes of transformation that reinforce one another.
First: the shift in value from tools to architecture. AI ceases to be an application and becomes the layer through which an organisation operates.
Second: the shift in human work from execution to design. Agentic AI does not mean fewer people but different people in different places.
Third: the shift in approach from ROI to Return on Experiment. Organisations that bet on fast learning rather than certain returns build a lead that cannot be bought back retrospectively.
Source: e15.cz
