IT & Intelligence Forum 2026: from AI that doesn't work to AI that surpasses human experts. And what's missing in between
The 95% of AI projects do not produce measurable results. 93% of AI budgets in companies go to technology, and only 7% to training and understanding of processes. 76% of employees use AI for work without policies, training, or knowing if it complies with GDPR or the AI Act. These three numbers, revealed during the Forum IT & Intelligence 2026 held on April 14th at Assolombarda in Milan, alone explain why artificial intelligence in Italian companies is still largely an unfulfilled promise. They also explain why the few cases where it really works deserve to be studied closely.
The 95% figure comes from the report The GenAI Divide: State of AI in Business 2025 by the MIT Project NANDA, which analyzed 300 public AI initiatives and conducted 52 interviews with executives. The fifth edition of the Forum, organized by For Human Relations with Datapizza as the main partner, gathered over 800 attendees and 120 speakers, and produced at least three stories that are worth putting next to each other.
Allianz: the AI assistant that has obtained legal validity
The first story is about Allianz Italia, narrated by Roberto Felici, Head of Future Lab, in a session led by Giacomo Ciarlini from Datapizza. Allianz’s health and asset insurance products are complex, with contracts spanning 500 pages that intermediaries only know partially. The solution was an intelligent assistant based on generative AI capable of answering any product-related questions, even in front of the customer.
What contradicts the prevailing narrative is that the project was entrusted to a 25-person startup after more renowned providers had failed, and the solution was implemented in three weeks. The traditional RAG (Retrieval Augmented Generation) architecture didn’t work for a modular product like Ultra Salute, where the same topic is covered in dozens of different documents with logical and syntactical variations because vector search was based on semantics and ended up getting confused. The startup built a multi-agent system that uses the LLMs themselves as search engines, with orchestrator agents, specialized and supporting agents dividing the logical space of tasks. The assistant surpassed human experts in accuracy tests, the responses gained legal validity, and the system now manages 15-16,000 prompts a month for a single product. The underwriters, freed from the help desk roles they had covered over the years, can now focus on risk assessment, which is their actual function.
Datapizza's keynote and Dual OS: the thesis of the AI-native company
The second story came from the keynote by Giacomo Ciarlini, co-founder and CIO of Datapizza, the Italian tech community with over 500,000 members that has transformed into the country's first AI Transformation Company. Ciarlini explained how in his company, a week of software development has now been compressed into a morning thanks to orchestrated AI agents, made possible by the “sustained correctness” of the latest models that can maintain accuracy throughout complex execution chains and recognize their own errors. The thesis of the keynote is that these models invalidate two fundamental assumptions of human organizations, namely intelligence as a scarce resource and the growing cost of coordination with the number of people involved, making it possible to build companies with a structure different from what we are used to, flatter with blurred role boundaries.
The thesis did not remain on the ideas level. In the Datapizza AI Lab, the team showcased live Dual OS, an AI-agnostic operating system with respect to vendors and models that self-configures onto the company’s tools. A meta-agent launches sub-agent explorers that recursively scan all data sources, map structures and relationships, identify inconsistencies, and produce a comprehensive map of the informational ecosystem. During the demo, a developer built a complete workflow in 15 minutes by chatting with the agent, and the system mapped a chaotic corporate dataset in a minute and a half, identifying anomalies that a consulting team would have taken months to uncover. Compared to products like Microsoft Copilot, Dual OS offers independence from vendors, agents with verifiable goals instead of pre-set steps, and complete transparency on costs and decisions. The product is still in the prototyping phase, with a stable version expected in a month and a half, and will be self-hostable on customers' cloud infrastructures.
The “messy middle” according to Jeremy Korst
The third story came in the afternoon, from the conversation between Ciarlini and Jeremy Korst, founder of Mindspan Labs and partner at GBK Collective. Korst is a co-author of the annual enterprise AI adoption study conducted with the Wharton School and Harvard, one of the longest-standing in the field, and has just published an article in Harvard Business Review addressing the issue of the “messy middle” in AI adoption.
The most interesting data point concerns a perceptual fracture that explains many things. Over two-thirds of senior executives believe they are getting a positive return from AI, but when it comes to middle managers, the picture changes sharply: they are skeptical, overwhelmed, and fear for their own job security. Executives fund, intermediate managers resist, and in between accumulate “proof of concept” projects that no one can bring into production. The study adopted a deliberately broad definition of ROI, measuring confidence in continuing to invest rather than pure financial impact on P&L, which makes the contradiction with the 95% figure from MIT merely apparent: one can believe they are getting something good while simultaneously lacking any measurable results on the balance sheet.
In the workplace, data significantly nuances the doomsday narrative: about 90% of respondents think AI can enhance skills, 70-72% believe it will replace specific tasks, and 40% fear it will render important skills obsolete. Three truths simultaneously that leadership should navigate differently rather than applying uniform approaches to all teams and functions.
Regarding the European delay, Korst was straightforward: it is visible, measurable, and for those competing in global markets, it's a concrete risk. But he added something usually missing from these analyses, namely that for companies in local markets, there is a “strategic procrastination” that can make sense because technology changes so quickly that investing too early can be equally costly. Provided they remain ready to move when the moment arrives, which is the part that many underestimate.
Two observations from the dialogue deserve separate mention. First: if an investment in AI is a strategic bet on the company’s survival, asking for a traditional business case means not understanding what one is doing, especially since the cost of experimentation has collapsed to the point that what used to take months and tens of thousands of dollars can today be prototyped in an afternoon. The second: AI as a “layer of intelligence” traversing all functions will converge previously separate processes and lead to the collapse of some organizational silos. LinkedIn is already creating roles like “Full Stack Builder”, multidisciplinary individuals who build functional products by working cross-functionally on sales, support, and marketing. And large organizations, paradoxically, are often lagging behind smaller ones in adoption, which opens a window of opportunity for those agile enough to slip in.
What Ciarlini thinks: exclusive interview
On the sidelines of the Forum, we spoke with Giacomo Ciarlini. Regarding the replicability of what Datapizza does in traditional companies, Ciarlini was honest: it isn’t replicable, at least not immediately, because their company was designed that way from the start. What is replicable is the method, which involves starting from the process and redesigning it from scratch, imagining that current models and those of the next 12-18 months are available. The problem, he said, is that companies do the opposite: they attach a piece of AI to the existing process and then are surprised that nothing changes, a bit like someone in 1998 putting a paper catalog in PDF and calling it e-commerce.
On the 95% failure rate of projects, Ciarlini diagnosed that the problem is almost never the technology. It’s the lack of a clear organizational decision on what they want to change and what they are truly willing to change. And when the conversation touches on issues of power, hierarchical structure, and professional identity, many organizations prefer simply not to have it, which condemns them to remain stuck in the pilot phase, where they can continue to experiment without questioning anything.
According to Ciarlini, companies that manage to scale share a leadership that uses AI tools firsthand, because otherwise, they don’t understand what is possible, what is difficult, and especially how fast everything is changing. Those who make AI strategy once a year and then archive it are already losing ground. And they have the courage to dismantle and rebuild, which for a CIO means accepting that part of what they built over twenty years needs to be rethought, something that doesn’t come naturally to those who have spent their careers optimizing the existing.
On the Allianz case, Ciarlini commented that the story should worry traditional consulting companies: the advantage of the startup wasn’t the technology, accessible to anyone, but the speed, adaptability, and ability to help the client understand what they were building together, while the large system integrators tend to stuff the problem into their packaged solution. Regarding “Dual Intelligence”, the concept that gives its name to Datapizza’s vision, Ciarlini stated that his teams are already producing the same output with half the people, not because they work harder but because coordination, handovers, delays, and rework are managed by agents. All this leads towards flatter organizational structures with fewer managerial layers, and the implications for career paths and how people think about their roles are considerable. The ambition for the coming months is to complete the journey towards fully automated software factories and help Italian companies do the same, starting with culture rather than technology.