Salesforce

Salesforce wants to take AI out of the testing phase and into business processes. This was the message coming from the Agentforce World Tour Milan, an event that brought together about 6,000 people, with over 90 sessions and more than 30 customer testimonials.

For Salesforce, the point is no longer to use artificial intelligence to write emails, summarize documents, or increase personal productivity. That phase has already started. The issue is different: many companies use AI as a personal assistant for various tasks but fail to scale it in critical processes. Customer service, sales, marketing, banking, transport, manufacturing. This is where AI needs to come in to generate a real impact.

Vanessa Fortarezza, SVP & Country General Manager of Salesforce Italy, spoke about the “agentic gap”: on one side are companies still stuck on pilots, often hampered by incomplete data and weak governance; on the other are companies already bringing AI agents into operational processes.

Headless 360: Salesforce Becomes a Backend for Agents

The main novelty announced in Milan is Headless 360, a solution that allows AI agents to operate regardless of the developer.

Headless 360 exposes the platform's functionalities as an MCP service, the Model Context Protocol. In practice, data, workflows, configurations, and Salesforce rules can be called from external environments like Slack, Claude, Gemini, ChatGPT, Cursor, Windsurf, or Claude Code.

"For years we have built applications for people. Today, we have a new family of users: AI agents. Headless 360 gives them access to everything companies already have in Salesforce: data, processes, rules, regardless of the platform they operate on," explains Nicola Lalla, VP Solution Engineering at Salesforce.

During the day, the topic of vibe coding was also discussed, which the company itself is using. However, the value of the platform does not lie in the code generated more quickly. It resides in the business knowledge accumulated over time: customer data, workflows, permissions, rules, interaction history.

UniCredit and Trenitalia: Agents Within Customer Relationship

Among the cases presented, UniCredit demonstrated how AI can work on personalizing offers. Through Data 360, customer interests, behaviors, and spending propensity are analyzed daily. The system generates dynamic suggestions and insights that can be used by the sales network.

The human agent no longer starts from a blank screen. They find prepared offers, signals, and prompts. AI does not replace the relationship but reduces the time needed to turn data into a sensible commercial proposal. In a regulated sector like banking, the theme is also about control: personalization yes, but within a governed perimeter.

Trenitalia introduced a case that is closer to the end user's daily experience. In three months, it released three agents on the Salesforce platform to support end users and sellers. The agents can assist in ticket sales, customer support, and managing recurring requests.

One of the scenarios involves Agentforce Voice: purchasing tickets via voice call, speaking with an AI agent. Another concerns delays. If a train experiences a service disruption, the agent can contact the traveler via WhatsApp and propose an alternative solution. When needed, the request is passed to a human operator, who receives a summary of the case and suggested responses.

Francesco Cacciapuoti, Chief Sales Officer of Trenitalia, explained that Salesforce is helping the company experiment with new contact and service modes, particularly on WhatsApp, with simpler and more integrated conversational experiences.

Gefran: AI for Choosing the Right Product Among 50,000 Codes

The case of Gefran brings the topic of agentic AI into manufacturing. The Brescia-based company produces sensors, components, and power controllers for industrial automation. It has over 250 product families and about 50,000 finished product codes. In such a context, customer service does not only handle standard requests but must understand technical applications, usage constraints, different markets, and product configurations.

Davide Bettera, CIO of Gefran, explains that the company has been using generative AI tools as personal support for offices and individual users for over a year. By 2026, however, work is moving towards more vertical applications integrated into business systems.

"Today we are focusing on more vertical implementations, on applications like Salesforce or SAP, where established processes and larger volumes of data exist," states the manager.

The chosen area on Salesforce is customer service, both in pre-sales and post-sales. In pre-sales, AI must help validate the technical solution and identify the product most suitable for the customer's request. Until now, this was a task assigned to experts in the individual product families. An effective model, but difficult to scale when the catalog is extensive, and it is necessary to integrate new people.

"We have people with twenty years of knowledge of Gefran. We want to introduce agents to support product selection," says Bettera.

The AI agent is trained on datasets and technical documentation. It can analyze the customer's request, check for any competing products, consult documents, and suggest the most suitable product or configuration. But it does not decide alone.

"It is a decision support tool; it does nothing autonomously. There is always a concept of 'man in the middle' who makes decisions," clarifies the CIO.

For now, the solution is designed for internal use, supporting Gefran's technical staff. In a subsequent phase, it could also be deployed on the web, on the company's site, to manage a first level of orientation or customer assistance.

Not Just Chatbots: AI Must Enter Enterprise Applications

Personal tools like ChatGPT or Copilot are useful but primarily generate local efficiencies. They help people work better but do not alone change business processes.

"Personal tools are very useful and create efficiency in micro-business solutions. This, however, is a different approach, more linked to processes and IT structures," explains Bettera.

The point is to bring AI into systems where the company already has data, procedures, and rules. Gefran has been using Salesforce since 2014 and SAP since 2001. This prior investment now accelerates adoption: according to Bettera, it takes three to four months to go from decision-making to actual use of a solution, precisely because the processes are already embedded in established applications.

There is also the issue of lock-in. Integrating AI on central platforms can strengthen ties to suppliers. However, Bettera downplays the issue:

"AI is not the driver in this regard. Anyone doing my job knows that replacing an ERP like SAP for a company like ours, with ten connected production sites around the world, is already a complex endeavor."

Gefran collaborates with Atlantic Technologies, part of Engineering, already a partner for Salesforce sales, marketing, and service components. Its role also includes helping the company choose the most suitable agents and models for different use cases. Because not all models work the same way on structured data, technical documentation, or web searches.

"When I need to do research on a datasheet, on the web, or on structured data in the database, it is essential to understand which language model is best suited to achieve precise results," says Bettera.

For manufacturing, accuracy matters. A wrong answer is not just a formal error; it can lead to recommending an incorrect product, an unsuitable configuration, or a solution that is not useful to the customer.

The Game is Played on Data Already Present in the Company

The cases of UniCredit and Trenitalia show what can happen in relationships with millions of customers. Gefran describes a more industrial application: placing AI beside those who must manage complex products, technical documentation, and specialized assistance requests.

This is where agentic AI makes the difference: when it helps an operator find the right product, qualifies a lead, prepares a coherent offer, or manages a service disruption, leaving a trace of what it does. Within governed processes. With reliable data. And with humans still in the right place in the decision-making chain.