Two use cases in logistics
Ariston Group, an Italian multinational in thermal comfort with a presence in more than 70 countries and around 10,000 employees, presented two applications of agentic AI in its supply chain. The account came from Gianmarco Baleani, the group's AI and digital transformation manager, during a conference at the Politecnico di Milano, as reported by AI4Business. The first case concerns spot transport requests, so-called RFQs: non-standard shipments or shipments outside contractual quotas coming from production, sales and maintenance, around 30–40 a month, each of which previously required about two hours for data entry and information exchanges.
To reduce these inefficiencies, Ariston built an RFQ Agent based on Microsoft Copilot Studio. According to the account, the starting problem was data quality: whoever submitted the request, often without expertise in the applicable regulations, could enter partial or incomplete information. The solution is structured around three pillars, which the material does not detail further. Baleani cites as benefits increased data quality through guided compliance with company policies, the removal of software usability barriers and the elimination of internal emails, with a direct increase in productivity. The shift from about two hours to a few minutes is the summary provided by the source, which does not publish independent measurement metrics.
The second project concerns optimizing material loading for entering customer orders. Here, maximizing truck or container fill for flexible requests requires a high degree of customization. To speed up the process, Ariston used a coding agent, citing tools such as Claude Code, Codex or Cursor, to develop the Load builder application. The volume handled is around 30–40 order entries a day, with estimated savings of at least 20 minutes of work per order. The application also takes into account the slang and technical terms used by warehouse staff during load preparation.
The governance and security crux
The use of coding agents to rapidly develop micro-applications raises issues that the source places front and center. On the one hand, generating code in minutes offers flexibility; on the other, uncontrolled application creation risks fueling shadow AI. In an organization with more than 10,000 employees, Baleani observes, if every worker independently developed their own applications without central oversight, it would create IT chaos and operational disaster.
On the security front, coding agents operate by taking control of the user's terminal, with potential access to sensitive data, commands and screenshots. For this reason, Ariston emphasizes the need to have them work inside protected, isolated sandboxes. The source also reports a process question that executives should ask themselves: is the application meant to close a local procedural gap, or is it necessary to redesign the entire company process around AI? These are methodological indications, not measured results.
Training against the internal digital divide
Before the new training plan, Ariston conducted an internal analysis that identified three groups: a few curious people who used AI extensively, a majority who used it little or not at all, and a share who used it incorrectly. According to the source, this resulted in a productivity gap and the absence of a shared language. Among the risks cited: the leakage of confidential data to unauthorized public tools with servers outside European borders, the uncritical acceptance of model outputs, and a failure to understand operational preconditions, including both underestimation of potential and unrealistic expectations.
The program adopted does not focus on basic prompting but on problem formulation and output validation. A four-hour module for small groups devotes the first two hours to how AI works, its risks and Anthropic's 4D framework for structuring prompts and describing expected results; the second two hours explore concepts such as Agent Skills, MCP and RAG, with exercises on concrete cases. Baleani argues that small groups encourage participation and create a shared methodological foundation. The actual impact of these programs on productivity and error reduction remains to be verified: the material contains no follow-up data, nor details on the three pillars of the RFQ Agent or the costs of the initiatives.
