Google Cloud AI Live Milan 2026: the era of AI agents is here, and Italy is preparing
There is a precise moment when a technology stops being a promise and becomes infrastructure. For AI in the enterprise sector, that moment is now. This is the message Google brought to Milan today with Cloud AI Live 2026, an event that has changed scale, tone, and content compared to previous editions. There is no longer talk of experimentation or pilot projects: we are discussing autonomous agents already in production, thousands of billions of tokens processed each month, and Italian companies that have redesigned their processes using Gemini models and Google Cloud as a cornerstone.
The guiding thread of the entire day is a precise concept: the Agentic Enterprise. Artificial intelligence is no longer an assistant that answers questions; it is an ecosystem of agents capable of reasoning, planning, calling tools, and acting independently on complex processes. A paradigm shift that, according to the data presented on stage, is already producing measurable effects. The slide shown in Milan did not indicate cumulative consumption, but the real-time processing rate of Google's models: tokens per minute increased from 7 billion to 16 billion between October 2025 and April 2026, with 10 billion already by January 2026.
The Italian Market: €1.8 Billion and a Window Not to Be Missed
The numbers for the Italian AI market are clear. According to the most recent data from the Politecnico di Milano, the sector is currently worth €1.8 billion, with a 50% year-on-year growth and a compound rate of 54% over the past three years. Seven out of ten large companies have already initiated AI projects, and six out of ten report finding real value in the solutions adopted. A study commissioned by Google to Frontier Consulting estimates that the full-blown impact of AI on the Italian economy could reach 8% of GDP, translating into a value between €150 and €170 billion over the next ten years. This is an enormous potential, but with a critical variable: each year of delay in adoption generates a loss of competitive advantage that is hard to recover.
Google is addressing this challenge with concrete investments in the territory. In addition to the continuity of the two regional data centers in Turin and Milan, the company has announced co-investment in the Blue-Raman submarine cable, which will connect Italy to the Middle East. On the human capital front, $2 million has been made available for Italian universities to train and certify 13,000 students in the field of AI, along with 100,000 free licenses for courses for professionals. In October 2025, a partnership with CDP Venture Capital initiated a program dedicated to Italian startups in the AI sector, offering expertise, operational support, and Cloud credits provided directly by Google.
The Platform: Build, Scale, Govern, and Optimize Agents
The technological basis of the day was the Gemini Enterprise Agent Platform, presented as the fundamental infrastructure for managing the entire lifecycle of an agent: from building to scalable production deployment, from governance to continuous performance evaluation. The platform adopts a low-code approach that, through tools like Antigravity, allows agents to be created even in a declarative mode, simply by describing in natural language the agent's goal, the type of model to use (which can be Google, third-party, or open-source like Gemma), the skills available, and the tools with which it can interact with business systems.
A fundamental architectural element presented during the keynote is the Model Context Protocol Server, now integrated by default in all Google Cloud Platform services. This protocol allows agents to inherit or read the context of the business systems they interact with, thus making every response contextualized to the organization's real data. This is complemented by the novelty of the Conversational Analytics API, a managed service that automatically translates natural language requests into SQL queries, eliminating the need to reinvent the layer for accessing structured data each time.
Scalability is managed by the Agent Runtime, a dedicated service that handles production deployment with guarantees of reliability, network security, and compliance. Governance is entrusted to the Agent Gateway, which monitors and filters in real-time every execution of the agents and proactively blocks any prompt or response that violates company policies. Each instance of each agent is tracked with a unique cryptographic ID to ensure full traceability. Finally, the optimization phase leverages continuous evaluation tools (evals) that allow performance measurement of agents against benchmarks defined by the company, both before go-live and in production.
Why Agents Change Everything: The Value of the Multi-Agent Approach
One of the most discussed topics during the day, also taken up in the session dedicated to the Multi-Agent Quest, is the deep reason why agents represent a qualitative leap compared to using AI in a monolithic way. A single general-purpose model must manage potentially huge contexts, with broad prompts and composite requests that increase the complexity of processing and reduce the accuracy of responses. Agents, on the other hand, work on narrow and well-defined contexts: each agent has a specific task, a specific area of knowledge, and a delimited set of tools. This means that it can be much more accurate, faster, and controllable. In a multi-agent architecture, multiple specialized agents collaborate under the supervision of an orchestrator, each contributing without having to “know everything,” just like in a well-structured human organization (or even within the same human brain, taking a moment to dive into the biological dimension). The result is an overall system that is more robust, scalable, and verifiable than any holistic AI approach.
On this point, Paolo Spreafico, Country Director of Customer Engineering Italy at Google Cloud, who we met on the sidelines of the event, was very direct: “An agent employs an extremely intuitive interface, which is typically natural language.” Spreafico also emphasized how the simplicity of creating an agent is rapidly lowering the accessibility threshold: “Creating an agent has become conceptually quite simple. You have an instruction, you need to indicate what type of model you want to use, define some skills, and then a series of tools that allow actions to be taken. In fact, an agent can potentially also become an autonomous entity that allows for the automation of many business processes.”
Security in the Agentic Era: Wiz and the Three-Agent Model
The accelerated adoption of agents also brings new security challenges. The keynote dedicated ample space to this theme, highlighting how the same AI models that companies use to defend themselves are now also available to attackers, resulting in an exponential increase in the speed, scalability, and sophistication of attacks. There are already self-modifying malware capable of responding to corporate countermeasures. In this context, Google introduced the integration of Wiz (a recently acquired reality) into the Gemini Enterprise Agent Platform, through an agentic approach to cybersecurity based on three specialized agents that collaborate autonomously: the Red agent proactively and continuously tests corporate defenses for vulnerabilities; the Blue agent analyzes identified vulnerabilities and evaluates their impact; the Green agent proposes and applies code fixes. As it was clearly stated on stage: “Right now, perimeter defense cannot be done solely with a human approach. The help of agents is crucial to having an effective response time to attacks.” Spreafico also provided a clear perspective on this front: “These tools, which make everything done faster and more scalable, can also be used by attackers. It is important to have tools that can respond to the renewed speeds and complexities that attackers may have at their disposal.” An observation that perfectly summarizes the stakes in security in the agentic era.
The New Models from Google I/O: Gemini 3.5, Omni, Spark, and Workspace Tools
The Milan event was also an occasion to delve deeper, in front of the Italian audience, into the announcements made at Google I/O just a few days ago. The first protagonist is Gemini 3.5 Flash, described as the most powerful agentic and coding model ever released by Google. It surpasses Gemini 3.1 Pro on the main programming benchmarks and is four times faster in terms of tokens generated per second compared to comparable competitor models. Its economic impact is remarkable: if the leading companies that currently process around a thousand billion tokens a day shifted 80% of their workloads to Gemini 3.5 Flash, they would save over a billion dollars a year to reinvest in business. During the keynote, it was also shown directly how the new generation of TPUs (Tensor Processing Unit) from Google, with the separation of the training phase from the inference phase, has nearly tripled throughput and response speed, ultimately bringing the infrastructure to around 120 thousand trillion floating-point operations per second.
Gemini Omni is instead the new multimodal model that unifies the management of text, audio, images, and video into a single platform. The goal is to simplify and centralize the production of visual and multimedia content to make generation and editing available via natural language. On stage, it was shown how Stellantis has already used this technology, specifically the generative video model Veo, to create a television advertising campaign (the ad “Italieno” from Padina Hybrid), starting from a concept developed by the company’s creative team and then expanded into a whole visual universe thanks to AI.
Gemini Spark completes the triad of new models: it is the AI personal agent active 24/7 within Gemini Enterprise and Google Workspace, capable of independently executing multi-step workflows in the background on behalf of the user.
On the Workspace front, two innovations complete the picture. Google Pics brings advanced control over the generation and editing of images to productivity suites: it is now possible to move, resize, and transform individual objects, modify or translate texts within images independently, all to simplify complex operations such as updating global marketing campaigns or adapting creative elements for different layouts. The new hands-free voice features for Gmail, Docs, and Keep allow brainstorming, organizing information, and completing tasks without touching the keyboard, directly within the most used Workspace apps. Additionally, the new AI Content Detection API provides companies with a tool to identify AI-generated content in support of responsible media governance.
Lastly, on the developer side, the Managed Agents API allows technical teams to create and activate custom agents with a single API call, without having to manage the underlying infrastructure: agents are deployed in secure remote environments hosted by Google to free teams from the infrastructure stack and allow them to focus exclusively on agent behavior. Connected to this, CodeMender is the new security agent integrated into the Gemini Enterprise Agent Platform: it autonomously identifies vulnerabilities, recommends fixes, tests them securely, and applies patches to dependent systems, always under the supervision and approval of the development team.
On Stage: Technogym, Stellantis, and the Evolution Bank Demo
The agency vision was brought to life by clients who took the stage. Technogym presented its AI Coach based on BigQuery and Gemini models: a system capable of conversing in natural language with users, adjusting training programs in real-time based on progress, training location, and personal goals, and providing fitness center operators with automated insights to optimize operational activities. The starting point is a comprehensive assessment called Technogym Check-Up based on six parameters that include body composition, cardiovascular functionality, strength, mobility, flexibility, and mental reflexes, from which the system derives the so-called Wellness Age, the functional age of the user. Technogym is the official supplier for the last ten editions of the Olympics and currently has 40 million users connected to its ecosystem, 100,000 wellness centers, and 500,000 homes worldwide.
The demo of Evolution Bank (a fictional bank, used only for demonstration purposes) showcased in real-time an agent capable of querying internal business data and external sources to process the analysis of a new financial product, a credit card designed for digital nomads, and automatically generate a complete proposal with presentation slides, market data, competitor analysis, and a value package. The demo also featured a fictional moment of tension, but useful to understand how to respond promptly: a temporary error in the model was immediately managed by the monitoring system of the platform that detected in real-time a drop in the quality of the response and automatically activated a new version of the agent. The demonstration revealed one of the most interesting aspects of the Gemini Enterprise Agent Platform: the ability to evaluate and replace agents in production without service interruption.
The Google Cloud AI Groundbreaker Awards 2026: Italian Excelences
The day also hosted the second edition of the Google Cloud AI Groundbreaker Awards, honors awarded to Italian organizations that have distinguished themselves for the use of Google Cloud technologies in high-impact projects. Four categories were awarded. For AI Agentic Transformation, the winner was Acea, for redesigning the customer experience with conversational AI solutions based on Gemini: the service is now active 24/7, waiting times have been reduced, abandoned calls have decreased by 60%, and about 35% of interactions are handled end-to-end by AI without human intervention. For AI for Creativity, the award went to OTB Group, represented on stage by Renzo Rosso, the famous founder of Diesel, for a virtual try-on system based on the Gemini Enterprise Agent Platform that generates personalized visual content directly on the customer's device, increasing engagement and conversion rates in retail. The recognition for AI for Culture was awarded to D4Science from CNR, for a scalable platform that democratizes access to AI in scientific research through Google Kubernetes Engine, Compute Engine, and Gemini Enterprise Agent Platform. Finally, the AI Accelerator award went to Generative Bionics, a startup active in humanoid robotics with applications in logistics and healthcare based on Google Cloud infrastructure.
Interview with Paolo Spreafico: AI as a Competitive Obligation
On the sidelines of the event, we spoke with Paolo Spreafico about the role Google intends to play in the Italian market and the concrete challenges companies must face in adopting an agentic approach. Spreafico emphasized how holding this event in Italian has a precise cultural impact: “The fact that it is in Italian brings the concepts closer and makes it clear that these technologies are really close to us. There is this type of barrier that gets removed.” Regarding the priorities for Italian companies, he indicated two transversal lines: improving customer experience and increasing productivity. “The agentic approach definitely makes the workforce more effective in carrying out a whole series of tasks that would otherwise be done less productively.” On the future of work, Spreafico took a pragmatic and responsible stance: “It is important that the use of AI is always aimed at the interest of humans. Workers cannot have no interactions with artificial intelligence if they want to be productive, fast, and effective in their work. AI is complementary and should be seen as a friend in respecting our work.” And regarding Google’s position in the global AI landscape, he emphasized how the Transformer architecture underlying all generative AI was invented by Google engineers, and concluded with: “Google is in a position of extreme leadership in this matter. Our approach aims to always be responsible, because we need to frame usage within norms that respect humans and the use of everyday life.