Artificial agents, a government architecture is needed

"In Italy, by 2024, investments in artificial intelligence have risen to around EUR 1.2 billion, out of the total EUR 40 billion that the Information and Communications Technology market as a whole develops. We are at the beginning of a story, of a very fast revolution to be brought to companies and that will give results in the medium term. Patience is needed and investments are needed, but we have one certainty: the cost of implementing agent-based AI architectures is lower than the cost of implementing an IT architecture". The analysis of Stefano Sperimborgo, Data & AI Lead for Italy, Greece and central Europe at Accenture, is lucid, and brings clarity to one of the many issues that are accompanying the spread of new AI technologies in Italy. Which one? The failure to adopt these tools. The risk exists, but those who have already 'burnt out' in this regard (encountering premature project failures) are because they underestimated the management of AI consumption and the resources needed to maintain models and train them.

"The agent governance architecture is the key to winning the artificial intelligence bet," Sperimborgo explained to Il Sole 24 Ore, "because it is this architecture that makes it possible to keep costs under control, to organise and make quality data, and to be able to choose Gen AI models, from large-format to Small Language, in an optimised way and with respect to real needs and specific applications. Agents are an opportunity to rethink processes and make them more efficient, they are the re-imagination of work and the way of doing business'.

The agent revolution, in short, promises benefits, but it must be directed according to certain requirements, which are in fact those at the basis of the AI Refinery, the project that Accenture has initiated in Europe in close synergy with Nvidia. Since its launch last October, the company has filed 55 patent applications in 10 countries and has involved companies such as the German Kion AG group, the beauty start-up Noli (with the support of L'Oréal) and multinationals such as Nestlé and Unilever. Last March, it introduced a tool (Agent Builder) that allows agents to be easily customised in code-free environments, and the stated goal is to implement 100 agent solutions by the end of this year. The essence of this project? Helping large European companies to start customised adoption paths through AI agent systems (the first AI agent multi-system collaboration solutions are already available), maintaining full control over critical data and fostering the adoption of AI-based robotics in the most strategic sectors, benefiting warehouse, logistics and production processes.

"AI Refinery," Sperimborgo pointed out in this regard, "is a central node for creating agents and managing them, overcoming the limitations of a forced choice of one cloud architecture over another and becoming a key component of the hybrid architecture of tomorrow, which will combine IT components and AI tools. It is, at the same time, also an example of a technology agnostic approach that moves in the direction of digital sovereignty and infrastructure in line with the dictates of the Digital Act and the AI Act'. Nvidia's engineering expertise was instrumental in developing an architecture capable of governing costs, managing LLM model development frameworks such as FrugalGPT, and preparing the knowledge base (the data) to apply AI, vectorising documents and organising agents in a hierarchical logic. "Agents," added the Accenture manager, "are the new workforce to be entrusted with operational tasks, alongside humans, who are their supervisors. Refinery should therefore be understood as an accelerator for building agents according to the various vertical sectors, providing advanced tools for optimising models and empowering companies to use more than one of them and in an orchestrated way, so as not to lose control over costs and consumption".

Italy, on this journey, is at the beginning of the ladder, in line with the rest of Europe but far from the adoption levels of the US. No one, as Sperimborgo finally observed, has yet put in place an agent architecture, but it is certain that this is now a priority for many and that large companies are much more aware of its importance and necessity. The acceleration recorded in the last three to four months, with projects expected to be grounded in the next six to twelve, seems promising.