Paolo Zaccardi: "Stopping hiring young people is madness"
The slide that made a splash in the audience was titled "When AI builds itself": 80% of Claude Code is now written by Claude Code, and the productive capacity of Anthropic's systems has grown eightfold from March to March. In the subsequent interview with Paolo Zaccardi, CEO and co-founder of Fabrick, he adds a detail that unsettles me more than the numbers: reading the documents that Anthropic publishes, the speed of self-improvement quarter over quarter has surprised even those who are building it. "The surprising thing is that it is a surprise for those who are doing it," he tells me, admitting that he was struck by the transparency with which the company shares its data, including its candor.
Zaccardi leads Fabrick, the Sella group's open finance platform, which has Mastercard among its investors. His intervention was titled "The Invisible Advantage" and brought together three converging forces: embedded finance, artificial intelligence, and autonomous agents. According to Mordor Intelligence data cited on stage, the embedded finance market is currently worth 156 billion euros and will surpass 450 billion by 2031.
Automation is a done deal
The starting point of the talk is that discussing automation no longer makes much sense because it is already here. Waymo, projected numbers in the room: 500,000 paid rides per week, 14 million trips in 2025 alone, a tenfold growth in less than two years, and a system that makes thousands of micro-decisions per second. Jensen Huang's quote from Computex in June also came up on stage: "every edge device will become autonomous." For Zaccardi, AI should be treated as the new operating system for people and businesses, and now the interesting question, he says, is what to automate: current processes as they are, or new processes, redesigned with the human at the center.
The Turing trap
When I ask him about layoffs related to AI, Zaccardi shifts the discussion to how companies measure adoption: almost all measure it in terms of reduced personnel. "The first metric you set is: I succeed if next year, instead of 3,000 people doing the same thing, I have 2,500." And the most painless way to reduce the workforce is well known: "there’s no better way to replace people than not hiring them, because you have no friction." The reference he cites is an essay by Stanford economist Erik Brynjolfsson: the Turing trap, the obsession with systems that replicate humans instead of extending them, with distorted incentives that reward replacement even when increasing capabilities would create more value.
The right question, for Zaccardi, would be a different one: "can I, with my differently educated people, achieve eight times what Anthropic does?" He admits that a price is paid anyway, the classic J curve where results worsen before improving. And regarding Italy's micro-enterprises, which will move more slowly due to their size and regulations, he sees a paradox: if large American companies choose the brutal path of layoffs and we don't choose any path, we will never embark on that curve.
Entry-level, critical thinking, and the EPOCH framework
On stage, the question from Peter Norvig (Stanford HAI) came up: if AI does the work that trained juniors, where will the next generation of seniors come from? Alongside this, a statistic attributed to the World Economic Forum: in the United States, entry-level jobs have decreased by 35% in eighteen months, largely due to AI. When I ask him if he agrees with American executives who call it madness to halt hiring young people, the response comes without hesitation: "it's madness." His reasoning: if a junior comes in to program and the system already programs better than him, it makes no sense for him to enter; if he comes in with critical capacity, with an understanding of complexities, he can leverage an automated system from day one.
The word that comes up again is training, both inside and outside companies. He cites the EPOCH framework from MIT (authored by Isabella Loaiza and Roberto Rigobón), which maps the skills where humans remain ahead: empathy, presence, judgment, creativity, and hope, understood as vision. He tells of a professor from New Jersey he met at a retreat with Stanford HAI: his university is redesigning its courses so that there is no program without critical thinking, Socratic dialogue, and dynamic communication built in. For Zaccardi, critical thinking has a precise role in the age of delegating to AI: understanding contexts, improving judgment, and reducing bias, "the true great risk of delegation," with the old case of Amazon recruiting serving as a warning. HR will also change, he says, as it must cultivate what he calls contextual intelligence.
If the agent pays the wrong thing, who puts up the money?
The most fintech part of the talk concerns agent-based payments, software that pays itself based on given objectives. The example on the slide: the flight is three hours delayed, the agent activates the insurance coverage, files the claim, and credits the refund, all before you’ve opened the app. Projected estimates indicate that 90% of B2B commerce will be mediated by agents by 2028 (Gartner) and 30% of B2C by 2030 (Deloitte), and already on the American Black Friday of 2025, about half of the purchase intentions arose from searches on LLMs.
At that point, I asked him the question any Italian customer would ask: if the agent pays the wrong thing, who bears the cost and who takes the blame? The answer comes from someone who handles the issue: the profiles of responsibility are not yet regulated, but the technical mechanisms to define them already exist. It’s called Verifiable Intent, an open-source protocol launched by Mastercard with Google: the user defines an intent, meaning objectives, boundaries, budget, and conditions, the intent is tokenized on the existing rails of card circuits, the agent executes the payment within those boundaries, and the merchant verifies the intent before selling, with everything being tracked. Within the boundaries, the responsibility lies with the purchaser; it changes if the agent steps outside the perimeter or if the merchant skips verification. Hence the slide with the lighthouse that summarizes his thesis on trust: "the more AI becomes invisible, the more trust must become visible." And that, even more concisely, on competitive advantage: "the transaction is the endpoint, the advantage is built beforehand."
At the end of the interview, since he was the only Italian of the day, I gave him a copy of my book, which discusses this very topic: what remains for humans as AI takes over tasks. We parted with the question that closed the stage, which is also why it was worth being there: "not what can we do with AI, but what do we really want to become thanks to AI." And for now, there’s no agent to delegate to.