When AI Enters Processes: Two AWS Partners Compare Cloud in Italy

At the AWS Summit in Milan, AWS lined up the numbers on cloud and artificial intelligence adoption in Italy, and the picture that emerges is that of a country in motion but still distant from maturity. Today, 40% of Italian businesses claim to use AI, with a 33% increase in one year, while cloud is now present in 70% of companies. The critical point is not adoption, but the depth with which technology enters processes. Even in the North-West, the most advanced area, only 18% of companies using AI have reached a transformative stage, a share that drops to 8% in the South and 6% in the Islands. Among small and medium-sized enterprises, adoption jumped from 29% to 38% in one year, yet only 11% have gone beyond experimentation.

On stage at the Summit, presenting the study, Giulia Gasparini, Country Leader of AWS Italy, focused on this very issue: “The cloud is the prerequisite for scalable AI,” she noted, because where digital foundations are weaker, the leap towards transformative use becomes more difficult. This is exactly the space that separates AI that summarizes documents from that which changes the way of working. This gap measures the partners AWS has built around itself in 15 years of Italian presence, and two of the longest-standing ones, beSharp and Reply, have navigated it from the start with almost opposite models. The two interviews conducted with Edge9 in the days following the Summit convey the same diagnosis and two different recipes.

Within that 40%, there are four levels of maturity. Simone Merlini, CEO and co-founder of beSharp, invites us to unpack that percentage, as it contains very different phenomena. At the first level, AI is a tool for personal productivity, the co-pilot that summarizes documents and prepares drafts of emails, now widely accepted but still poorly governed. Above it lies an optimization of business flows focused on two use cases, automated customer assistance and enhanced search on internal knowledge bases (techniques known as RAG), the simplest to realize. Then come the higher plans, where the share thins: integrated agentic AI in processes, capable of making decisions autonomously, and finally the same agentic approach brought into the products and services the company offers its customers. The higher one goes, the more the technical complexity and the knots of security, governance, and regulatory compliance increase, and that initial 40% gets pressed down to the lowest rung.

Right at the top of the ladder, Merlini sees a film replaying. Generative AI has entered companies as the cloud did in the past, through the back door of personal tools, and today the risk is the same as in the era when employees circumvented IT by uploading company files to consumer services. This is shadow AI, the confidential documents ending up on a free platform without anyone knowing, and Merlini frankly admits that for a moment, even beSharp was about to find its collaborators signing up for external services with personal accounts before rushing to take precautions.

The knot for beSharp is the governance of agents. Many, by their very design, inherit the permissions of the user who initiates them, and the potential for harm grows with their autonomy. Just think of a powerful agent like OpenClaw, which appeared only a few months ago: such tools open concrete possibilities but also a very wide margin of risk if something goes wrong. The right approach, Merlini argues, is almost Socratic, the full awareness of not knowing in a field that continuously changes. “We are moving in a minefield, where the mines continue to change position,” is the image he uses, wary of those who promise platforms capable of bringing security and AI governance to order once and for all. In a market inflated by marketing, he ironically reminds us, every database has suddenly become the best for artificial intelligence.

The story of beSharp with AWS begins long before all this. When the company was founded in 2011, there was yet no local AWS structure in Italy, and Merlini was in dialogue with a contact in Luxembourg, Nicola Previati, who later headed AWS Italy for years, who noticed him for a curious reason: on his personal credit card, there was active what was probably the largest AWS cloud subscription in the country at the time. From there, beSharp has seen AWS establish its local structure in Italy, which has grown over the years to include hundreds of people. The company itself now counts just under 70 collaborators and is fully expanding, with a strong root in Pavia and a close tie with the university, a legacy of a community and dissemination model.

That model, Merlini recalls, marks the difference between the two waves. In the beginning, the cloud had to be pushed: the few pioneering partners spent more time evangelizing the market than realizing projects. With AI, the opposite happened; it is the market that pulls in a hundred different directions, and the partner struggles to keep pace with a demand that is as confused as it is pressing.

Among the cases that best tell the story, beSharp cites Satispay, for which it managed the very first migration to AWS in 2015, and which, ten years later, is still a client, recently engaged in rewriting part of its older code. However, the clearest example of the leap toward agentic AI is Planeat, the meal planning service beSharp uses for its cafeteria. Weekly planning, with 21 meals to decide in advance, was the main friction point for users. BeSharp built an automatic planner, first in generative form and then in agentic terms, the real turning point. The result, Merlini sums up, is as simple as it is eloquent: “A 10% increase in revenue. That's how the story is told.”

Everything runs on AWS, on a native platform, with the AI segment entrusted to Amazon Bedrock and the AgentCore family of services. Behind these results, insists beSharp, there's often overlooked point. “Many forget that AI is only possible thanks to the cloud,” Merlini observes. It is the cloud that enables training large models, access to frontier models, and the ability to quickly turn on and off many environments to try different scenarios, a necessary condition for innovation to remain sustainable even in costs. When attention focuses solely on AI, the underlying level disappears from view, and it is there that much of the outcome is determined.

The network model and measurable value Reply reads the market in the same way but approaches it from a different perspective, that of a large group organized in a network of dozens of specialized companies, each dedicated to its own area. Luca Gardini, Global Head of Strategic Partnership and Alliances and Partner at Reply, who closely follows the relationship with AWS, recalls a partnership lasting nearly 20 years, one of the first in Italy for a medium-large system integrator. He frames it in three phases. The first, cloud-native, was that of infrastructural migration, the shift of workloads from on-premise to cloud, and application modernization, with projects like that of Enel leading the way. The second was about managed services, when increasing complexity made it necessary to ensure operational continuity, security, and, almost immediately, cost control, known as FinOps, the discipline of cloud cost optimization. The third phase, currently under full development, is that of AI.

To bridge the gap between experimentation and transformation, Reply focuses on three levers. The first is vertical expertise, the sector specialization that allows generative AI to be applied in the specific processes of a factory, a bank, or a retail company, rather than applying it indiscriminately. The second is the ability to take complex projects from prototype to large-scale production, with modular cloud-native architectures that provide the customer a quick return but are ready to grow. The third responds to a clearly emerging data point from the same AWS study: the lack of internal expertise complained of by half of companies, resulting in training programs and change management, as well as handling regulatory complexities. “We don’t propose three-year AI adoption programs,” Gardini summarizes, “but critical processes engineered in modules that show their utility right away.”

This approach recently received formal recognition. In June, Reply obtained the AWS Business Value Realization Competency, a specialization that distinguishes partners capable of linking cloud and AI projects to measurable business outcomes, and which involves the group companies specialized in AWS technologies, from Storm Reply to Data Reply. This is the point Gardini returns to repeatedly in the narrative: “Being technologically good is not enough; you need to know how to link everything to measurable business results.” The case he refers to is that of BMW, for which Reply modernized the data analysis systems with a generative AI-based approach, reducing manual activities by over 50% and achieving an accuracy of up to 90% in conversion.

The concrete examples, in Gardini's account, span various sectors. For Aeroporti di Roma, along with the group company Storm Reply, Reply created a voice virtual assistant accessible via WhatsApp, capable of responding in Italian and English, in real time, on flights, services, and shops at Fiumicino airport, based on Amazon Bedrock and a multi-agent architecture. In manufacturing, connected sensor solutions (IoT) and machine learning analyze factory data to predict failures and prevent machine downtime, which are worth millions of euros. In the financial and insurance sector, systems built on Bedrock assist consultants by analyzing thousands of pages of documentation in seconds and providing verified responses. In retail, predictive models for customizing offers have increased online sales conversion rates by 20-30%.

On a point, Gardini insists as much as Merlini, albeit from a more economic angle. Just as in the transition to the cloud, cost control quickly became essential, with AI the theme resurfaces amplified. Reports of companies that replaced personnel with models only to discover a token bill much higher than expected tell a real problem, he notes. Reply builds safeguards into projects and aims for as predictable costs as possible because every competitive advantage must close the books with a return on investment.

Read together, the two narratives converge on a conclusion. AI produces measurable results only when it rests on solid data foundations and ties to a precise process, and the season of experimentation for its own sake is behind us, as evidenced by AWS's very decision to certify business value before technical skill. On this common ground move two opposing figures: on one side, beSharp, grown from the bottom up, attentive to technical culture and the risks of agent autonomy still difficult to govern; on the other, Reply, which industrializes transformation through its network model and measures it with business metrics. What unites them is the idea that the cloud is not the backdrop for AI but its condition of existence, the part of the infrastructure that, precisely while everyone looks elsewhere, continues to make a difference.