AI Agents Without Software: Why Salesforce Does Not Believe in the 'SaaS Apocalypse'
The Title
Salesforce is among the most affected by the market narrative known as the "SaaS apocalypse," the idea that next-generation language models could render the software platforms that drive much of business processes redundant.
Marco Hernansanz, Executive Vice President & CEO for Southern Europe, Middle East and Africa, used the online roundtable with a select group of European publications to present the opposite thesis: in the era of the agentic enterprise, software does not disappear; it becomes the trusted operating system for running AI agents.
The business context that Hernansanz brought to the table aligns with the narrative. Salesforce closed the FY26 fiscal year at $41.5 billion and has set a target for the new fiscal year between $45.8 and $46.2 billion, with a goal of $63 billion for FY30, implying a compound growth rate of over 11% annually. According to the manager’s interpretation, these numbers should be kept distinct from stock fluctuations and notes from analysts and financial institutions that suggest a transition phase lasting up to a year and a half. Hernansanz avoided delving into the financial climate and shifted the discussion to commercial traction, claiming to see growth in the Agentforce customer portfolio, now around 30,000 globally and in the hundreds in markets like Italy and Spain.
Agentic AI from Pilot Projects to Production
The central point of the narrative was dedicated to why, according to Hernansanz, most AI projects have not yet reached production. This refers to the now-famous statistic of 95% of failed pilots, derived from the MIT NANDA study "The GenAI Divide, State of AI in Business 2025." A figure that has become somewhat of a cliché in enterprise discussions, but should be viewed as a snapshot of a phase that is partly behind us: as emerged in the Q&A, in recent months, that percentage has progressively decreased, and a growing number of agentic use cases are indeed reaching production. However, Hernansanz's diagnosis remains useful to argue his thesis: language models alone do not know the business, do not have access to business data, are non-deterministic, and can provide wrong answers with assertive tone—especially critical limits in the regulated industries upon which Salesforce's customer base relies.
The architectural response that the group proposes is articulated in four layers. The first is the context system, built around Data 360 and the zero-copy model that allows the agent to read information directly from external data lakes like Snowflake or Databricks without moving the data. The second is the work system, i.e., the workflows, integrations, and applications that Salesforce has layered over 27 years of activity, providing agents with concrete action points on business systems. The third is the agency system, entrusted to Agentforce, which manages governance, orchestration, and telemetry necessary to scale beyond the individual agent to ecosystems where different agents also communicate with third-party agents through protocols like MCP. The fourth is the engagement system, handled by Slack as a conversational interface spanning both people and agents.
Hernansanz's vision presented is of a granular agentic enterprise, where each process corresponds to a specialized vertical agent, not a single super-agent. The Agentforce library includes around 300 pre-packaged agents for specific industries, from generating an insurance quote to handling a credit card refund, from onboarding a banking client to scheduling a technical home visit. Salesforce itself presents as the first use case: the group’s help desk, now managed by Agentforce, reports 1.6 million requests resolved before a ticket was generated, $100 million in annualized savings on support, $130 million in sales pipeline intercepted by agents qualifying website traffic, 650 new customers from previously unreachable demand, and 78,000 employees assisted.
Market Proof: Telepass and Boggi Milano
In Italy, Salesforce has already fielded two significant references. The first is Telepass, which has brought into production a multilingual self-service agent integrated into the customer portal and a second internal agent to support contact center operators. As reported by Marco Gaeta, Chief Information Technology Officer of the company, in the customer case published by the group, the agent handles 40,000 conversations per week, generates about 1.4 million monthly calls to the underlying LLMs, and autonomously resolves 87% of requests, compared to a previous chatbot that forwarded more than half to an operator. The release of the first application took six weeks, and the return on investment was achieved in ten, with the Testing Center automating about 70% of the tests.
The scale context in which these numbers are read is significant: Telepass manages over 75% of Italian highway tolls, 1.4 billion cashless transactions a year, and over 10 million customers, with a contact center that the company reports varies between 300 and 400 operators depending on seasonality. The internal agent has helped halve the average call handling time, while the onboarding of new partners on the Salesforce Community has dropped from about three months to one or two weeks.
The second Italian reference is Boggi Milano, a men’s clothing chain with over 290 retail locations in more than 60 countries, a Salesforce customer since 2015. In March 2026, the brand announced the launch of MYAgent, an AI agent based on Agentforce integrated into e-commerce and WhatsApp, with Jakala as its implementation partner. The declared functional scope includes real-time access to orders and catalogs, style recommendations, management of returns and size exchanges, and multilingual service. The launch presents itself as a first step towards personalizing interaction with customers across all direct channels of the brand.
The trajectory described by Hernansanz for company adoption mirrors that seen in previous technological cycles. It starts with internal use cases supporting employees, averaging in production in three months and fully operational within six, then moves to customer-facing use cases, first informational and then transactional, and finally arrives at the composition of internal and external agents via standard protocols like MCP—a stage currently reserved for more mature companies. Regarding the infrastructural pressure raised in the Q&A about the consumption of agents and the large infrastructures dedicated to AI built by model providers, the manager acknowledged the problem but indicated that the prevailing direction is to use smaller and specialized models, orchestrated together rather than a single massive model. For Salesforce, in other words, the real challenge is not to defend enterprise software from the rhetoric of the 'SaaS apocalypse' but to translate into functioning agents, and in rapid timelines, the advantage accumulated over 27 years of workflows, integrations, and data on the installed base of its customers.