Salesforce AI gains 6,000 customers while quietly pulling back on large language models

While everyone talks about an AI bubble, Salesforce quietly added 6,000 enterprise customers in 3 months

While Silicon Valley debates whether artificial intelligence has become an overinflated bubble, Salesforce's enterprise AI platform quietly added 6,000 new customers in a single quarter -- a 48% increase that executives say demonstrates a widening gap between speculative AI hype and deployed enterprise solutions generating measurable returns. Agentforce, the company's autonomous AI agent platform, now serves 18,500 enterprise customers, up from 12,500 the prior quarter. Those customers collectively run more than three billion automated workflows monthly and have pushed Salesforce's agentic product revenue past $540 million in annual recurring revenue, according to figures the company shared with VentureBeat. The platform has processed over three trillion tokens -- the fundamental units that large language models use to understand and generate text -- positioning Salesforce as one of the largest consumers of AI compute in the enterprise software market. "This has been a year of momentum," Madhav Thattai, Salesforce's Chief Operating Officer for AI, said in an exclusive interview with VentureBeat. "We crossed over half a billion in ARR for our agentic products, which have been out for a couple of years. And so that's pretty remarkable for enterprise software." The numbers arrive amid intensifying scrutiny of AI spending across corporate America. Venture capitalists and analysts have questioned whether the billions pouring into AI infrastructure -- from data centers to graphics processing units to model development -- will ever generate proportionate returns. Meta, Microsoft, and Amazon have committed tens of billions to AI infrastructure, prompting some investors to ask whether the enthusiasm has outpaced the economics. Yet the Salesforce data suggests that at least one segment of the AI market -- enterprise workflow automation -- is translating investments into concrete business outcomes at a pace that defies the bubble narrative. Why enterprise AI trust has become the defining challenge for CIOs in 2025 The distinction between AI experimentation and AI deployment at scale comes down to one word that appeared repeatedly across interviews with Salesforce executives, customers, and independent analysts: trust. Dion Hinchcliffe, who leads the CIO practice at technology research firm The Futurum Group, said the urgency around enterprise AI has reached a fever pitch not seen in previous technology cycles. His firm recently completed a comprehensive analysis of agentic AI platforms that ranked Salesforce slightly ahead of Microsoft as the market leader. "I've been through revolution after revolution in this business," Hinchcliffe said. "I've never seen anything like this before. In my entire career, I've never seen this level of business focus -- boards of directors are directly involved, saying this is existential for the company." The pressure flows downward. CIOs who once managed technology as a cost center now field questions directly from board members demanding to know how their companies will avoid being disrupted by AI-native competitors. "They're pushing the CIO hard, asking, 'What are we doing? How do we make sure we're not put out of business by the next AI-first company that reimagines what we do?'" Hinchcliffe said. But that pressure creates a paradox. Companies want to move fast on AI, yet the very autonomy that makes AI agents valuable also makes them dangerous. An agent that can independently execute workflows, process customer data, and make decisions without human intervention can also make mistakes at machine speed -- or worse, be manipulated by bad actors. This is where enterprise AI platforms differentiate themselves from the consumer AI tools that dominate headlines. According to Hinchcliffe, building a production-grade agentic AI system requires hundreds of specialized engineers working on governance, security, testing, and orchestration -- infrastructure that most companies cannot afford to build themselves. "The average enterprise-grade agentic team is 200-plus people working on an agentic platform," Hinchcliffe said. "Salesforce has over 450 people working on agent AI." Early in the AI adoption cycle, many CIOs attempted to build their own agent platforms using open-source tools like LangChain. They quickly discovered the complexity exceeded their resources. "They very quickly realized this problem was much bigger than expected," Hinchcliffe explained. "To deploy agents at scale, you need infrastructure to manage them, develop them, test them, put guardrails on them, and govern them -- because you're going to have tens of thousands, hundreds of thousands, even millions of long-running processes out there doing work." How AI guardrails and security layers separate enterprise platforms from consumer chatbots The technical architecture that separates enterprise AI platforms from consumer tools centers on what the industry calls a "trust layer" -- a set of software systems that monitor, filter, and verify every action an AI agent attempts to take. Hinchcliffe's research found that only about half of the agentic AI platforms his firm evaluated included runtime trust verification -- the practice of checking every transaction for policy compliance, data toxicity, and security violations as it happens, rather than relying solely on design-time constraints that can be circumvented. "Salesforce puts every transaction, without exception, through that trust layer," Hinchcliffe said. "That's best practice, in our view. If you don't have a dedicated system checking policy compliance, toxicity, grounding, security, and privacy on every agentic activity, you can't roll it out at scale." Sameer Hasan, who serves as Chief Technology and Digital Officer at Williams-Sonoma Inc., said the trust layer proved decisive in his company's decision to adopt Agentforce across its portfolio of brands, which includes Pottery Barn, West Elm, and the flagship Williams-Sonoma stores that together serve approximately 20% of the U.S. home furnishings market. "The area that caused us to make sure -- let's be slow, let's not move too fast, and let this get out of control -- is really around security, privacy, and brand reputation," Hasan said. "The minute you start to put this tech in front of customers, there's the risk of what could happen if the AI says the wrong thing or does the wrong thing. There's plenty of folks out there that are intentionally trying to get the AI to do the wrong thing." Hasan noted that while the underlying large language models powering Agentforce -- including technology from OpenAI and Anthropic -- are broadly available, the enterprise governance infrastructure is not. "We all have access to that. You don't need Agentforce to go build a chatbot," Hasan said. "What Agentforce helped us do more quickly and with more confidence is build something that's more enterprise-ready. So there's toxicity detection, the way that we handle PII and PII tokenization, data security and creating specific firewalls and separations between the generative tech and the functional tech, so that the AI doesn't have the ability to just go comb through all of our customer and order data." The trust concerns appear well-founded. The Information reported that among Salesforce's own executives, trust in generative AI has actually declined -- an acknowledgment that even insiders recognize the technology requires careful deployment. Corporate travel startup Engine deployed an AI agent in 12 days and saved $2 million For Engine, a corporate travel platform valued at $2.1 billion following its Series C funding round, the business case for Agentforce crystallized around a specific customer pain point: cancellations. Demetri Salvaggio, Engine's Vice President of Customer Experience and Operations, said his team analyzed customer support data and discovered that cancellation requests through chat channels represented a significant volume of contacts -- work that required human agents but followed predictable patterns. Engine deployed its first AI agent, named Ava, in just 12 business days. The speed surprised even Salvaggio, though he acknowledged that Engine's existing integration with Salesforce's broader platform provided a foundation that accelerated implementation. "We saw success right away," Salvaggio said. "But we went through growing pains, too. Early on, there wasn't the observability you'd want at your fingertips, so we were doing a lot of manual work." Those early limitations have since been addressed through Salesforce's Agentforce Studio, which now provides real-time analytics showing exactly where AI agents struggle with customer questions -- data that allows companies to continuously refine agent behavior. The business results, according to Salvaggio, have been substantial. Engine reports approximately $2 million in annual cost savings attributable to Ava, alongside a customer satisfaction score improvement from 3.7 to 4.2 on a five-point scale -- an increase Salvaggio described as "really cool to see." "Our current numbers show $2 million in cost savings that she's able to address for us," Salvaggio said. "We've seen CSAT go up with Ava. We've been able to go from like a 3.7 out of five scale to 4.2. We've had some moments at 85%." Perhaps more telling than the cost savings is Engine's philosophy around AI deployment. Rather than viewing Agentforce as a headcount-reduction tool, Salvaggio said the company focuses on productivity and customer experience improvements. "When you hear some companies talk about AI, it's all about, 'How do I get rid of all my employees?'" Salvaggio said. "Our approach is different. If we can avoid adding headcount, that's a win. But we're really focused on how to create a better customer experience." Engine has since expanded beyond its initial cancellation use case. The company now operates multiple AI agents -- including IT, HR, product, and finance assistants deployed through Slack -- that Salvaggio collectively refers to as "multi-purpose admin" agents. Williams-Sonoma is using AI agents to recreate the in-store shopping experience online Williams-Sonoma's AI deployment illustrates a more ambitious vision: using AI agents not merely to reduce costs but to fundamentally reimagine how customers interact with brands digitally. Hasan described a frustration that anyone who has used e-commerce over the past two decades will recognize. Traditional chatbots feel robotic, impersonal, and limited -- good at answering simple questions but incapable of the nuanced guidance a knowledgeable store associate might provide. "We've all had experiences with chatbots, and more often than not, they're not positive," Hasan said. "Historically, chatbot capabilities have been pretty basic. But when customers come to us with a service question, it's rarely that simple -- 'Where's my order?' 'It's here.' 'Great, thanks.' It's far more nuanced and complex." Williams-Sonoma's AI agent, called Olive, goes beyond answering questions to actively engaging customers in conversations about entertaining, cooking, and lifestyle -- the same consultative approach the company's in-store associates have provided for decades. "What separates our brands from others in the industry -- and certainly from the marketplaces -- is that we're not just here to sell you a product," Hasan said. "We're here to help you, educate you, elevate your life. With Olive, we can connect the dots." The agent draws on Williams-Sonoma's proprietary recipe database, product expertise, and customer data to provide personalized recommendations. A customer planning a dinner party might receive not just product suggestions but complete menu ideas, cooking techniques, and entertaining tips. Thattai, the Salesforce AI executive, said Williams-Sonoma is in what he describes as the second stage of agentic AI maturity. The first stage involves simple question-and-answer interactions. The second involves agents that actually execute business processes. The third -- which he said is the largest untapped opportunity -- involves agents working proactively in the background. Critically, Hasan said Williams-Sonoma does not attempt to disguise its AI agents as human. Customers know they're interacting with AI. "We don't try to hide it," Hasan said. "We know customers may come in with preconceptions. I'm sure plenty of people are rolling their eyes thinking, 'I have to deal with this AI thing' -- because their experience with other companies has been that it's a cost-cutting maneuver that creates friction." The company surveys customers after AI interactions and benchmarks satisfaction against human-assisted interactions. According to Hasan, the AI now matches human benchmarks -- a constraint the company refuses to compromise. "We have a high bar for service -- a white-glove customer experience," Hasan said. "AI has to at least maintain that bar. If anything, our goal is to raise it." Williams-Sonoma moved from pilot to full production in 28 days, according to Salesforce -- a timeline that Thattai said demonstrates how quickly companies can deploy when they build on existing platform infrastructure rather than starting from scratch. The three stages of enterprise AI maturity that determine whether companies see ROI Beyond the headline customer statistics, Thattai outlined a three-stage maturity framework that he said describes how most enterprises approach agentic AI: Stage one involves building simple agents that answer questions -- essentially sophisticated chatbots that can access company data to provide accurate, contextual responses. The primary challenge at this stage is ensuring the agent has comprehensive access to relevant information. Stage two involves agents that execute workflows -- not just answering "what time does my flight leave?" but actually rebooking a flight when a customer asks. Thattai cited Adecco, the recruiting company, as an example of stage-two deployment. The company uses Agentforce to qualify job candidates and match them with roles -- a process that involves roughly 30 discrete steps, conditional decisions, and interactions with multiple systems. "A large language model by itself can't execute a process that complex, because some steps are deterministic and need to run with certainty," Thattai explained. "Our hybrid reasoning engine uses LLMs for decision-making and reasoning, while ensuring the deterministic steps execute with precision." Stage three -- and the one Thattai described as the largest future opportunity -- involves agents working proactively in the background without customer initiation. He described a scenario in which a company might have thousands of sales leads sitting in a database, far more than human sales representatives could ever contact individually. "Most companies don't have the bandwidth to reach out and qualify every one of those customers," Thattai said. "But if you use an agent to refine profiles and personalize outreach, you're creating incremental opportunities that humans simply don't have the capacity for." Salesforce edges out Microsoft in analyst rankings of enterprise AI platforms The Futurum Group's recent analysis of agentic AI platforms placed Salesforce at the top of its rankings, slightly ahead of Microsoft. The report evaluated ten major platforms -- including offerings from AWS, Google, IBM, Oracle, SAP, ServiceNow, and UiPath -- across five dimensions: business value, product innovation, strategic vision, go-to-market execution, and ecosystem alignment. Salesforce scored above 90 (out of 100) across all five categories, placing it in what the firm calls the "Elite" zone. Microsoft trailed closely behind, with both companies significantly outpacing competitors. Thattai acknowledged the competitive pressure but argued that Salesforce's existing position in customer relationship management provides structural advantages that pure-play AI companies cannot easily replicate. "The richest and most critical data a company has -- data about their customers -- lives within Salesforce," Thattai said. "Most of our large customers use us for multiple functions: sales, service, and marketing. That complete view of the customer is central to running any business." The platform advantage extends beyond data. Salesforce's existing workflow infrastructure means that AI agents can immediately access business processes that have already been defined and refined -- a head start that requires years for competitors to match. "Salesforce is not just a place where critical data is put, which it is, but it's also where work is performed," Thattai said. "The process by which a business runs happens in this application -- how a sales process is managed, how a marketing process is managed, how a customer service process is managed." Why analysts say 2026 will be the real year of AI agents in the enterprise Despite the momentum, both Salesforce executives and independent analysts cautioned that enterprise AI remains in early innings. Hinchcliffe pushed back against the notion that 2025 was "the year of agents," a phrase that circulated widely at the beginning of the year. "This was not the year of agents," Hinchcliffe said. "This was the year of finding out how ready they were, learning the platforms, and discovering where they weren't mature yet. The biggest complaint we heard was that there's no easy way to manage them. Once companies got all these agents running, they realized: I have to do lifecycle management. I have agents running on old versions, but their processes aren't finished. How do I migrate them?" He predicted 2026 has "a much more likely chance of being the year of agents," though added that the "biggest year of agents" is "probably going to be the year after that." The Futurum Group's analysis forecasts the AI platform market growing from $127 billion in 2024 to $440 billion by 2029 -- a compound annual growth rate that dwarfs most enterprise software categories. For companies still on the sidelines, Salvaggio offered pointed advice based on Engine's early-adopter experience. "Don't take the fast-follower strategy with this technology," he said. "It feels like it's changing every week. There's a differentiation period coming -- if it hasn't started already -- and companies that waited are going to fall behind those that moved early." He warned that institutional knowledge about AI deployment is becoming a competitive asset in itself -- expertise that cannot be quickly acquired through outside consultants. "Companies need to start building AI expertise into their employee base," Salvaggio said. "You can't outsource all of this -- you need that institutional knowledge within your organization." Thattai struck a similarly forward-looking note, drawing parallels to previous platform shifts. "Think about the wave of mobile technology -- apps that created entirely new ways of interacting with companies," he said. "You're going to see that happen with agentic technology. The difference is it will span every channel -- voice, chat, mobile, web, text -- all tied together by a personalized conversational experience." The question for enterprises is no longer whether AI agents will transform customer and employee experiences. The data from Salesforce's customer base suggests that transformation is already underway, generating measurable returns for early adopters willing to invest in platform infrastructure rather than waiting for a theoretical bubble to burst. "I feel incredibly confident that point solutions in each of those areas are not the path to getting to an agentic enterprise," Thattai said. "The platform approach that we've taken to unlock all of this data in this context is really the way that customers are going to get value."

AI bubble bursting? Salesforce execs admit trust issues after laying off 4,000 techies, now scaling back use of AI models

Salesforce is reducing its reliance on generative AI. The company faced reliability problems with large language models. Executives admit confidence in AI has dropped. Salesforce is now prioritizing predictable automation. This shift follows reports of job cuts after AI agent deployment. The company aims for dependable business use of technology. Salesforce, one of the world's biggest enterprise software firms, is quietly pulling back from its aggressive use of large language models after facing repeated reliability issues. Senior executives now admit that confidence in generative AI has fallen sharply over the past year, even as the company laid off around 4,000 support staff following the rollout of AI agents, according to reports by The Information. The shift marks a sharp change in tone for a company that was among the loudest champions of AI-led transformation across corporate software. The change in strategy was publicly acknowledged by Sanjna Parulekar, Senior Vice President of Product Marketing at Salesforce. Speaking about the company's experience with large language models, she said, "All of us were more confident about large language models a year ago." Her remark highlights growing internal concerns about how reliable these models are when used for real business tasks. Salesforce is now moving away from open-ended generative AI and placing greater emphasis on predictable, rule-based automation in its flagship product, Agentforce. The company says this approach helps remove "the inherent randomness of large models," a clear retreat from the AI-first messaging that dominated the tech industry just months ago. Salesforce's AI push has already reshaped its workforce. CEO Marc Benioff revealed in a podcast appearance, cited by CNBC, that the company reduced its support staff from 9,000 to about 5,000 after deploying AI agents. Explaining the impact, Benioff said, "I've reduced it from 9,000 heads to about 5,000, because I need less heads." The comments underline how quickly AI tools have moved from experimental features to systems influencing staffing decisions, even as questions grow about their reliability. Behind the scenes, Salesforce engineers flagged several technical issues with large language models. Muralidhar Krishnaprasad, Chief Technology Officer of Agentforce, said the models begin failing when given too many instructions. According to him, once the number of instructions crosses eight, the models start dropping some directives. For businesses that rely on precision and consistency, this behaviour poses a serious risk. These issues surfaced during live customer deployments. Home security firm Vivint, which uses Agentforce to handle customer support for 2.5 million customers, experienced problems with follow-up surveys. Despite clear instructions to send satisfaction surveys after every interaction, Agentforce sometimes failed to do so for unexplained reasons, The Information reported. Vivint later worked with Salesforce to add "deterministic triggers" to ensure surveys were sent every time. Another problem emerged in what Salesforce executive Phil Mui described as AI "drift" in an October blog post. When users ask irrelevant questions, AI agents tend to lose focus on their main task. For example, a chatbot designed to guide users through filling out a form may get distracted when customers ask unrelated questions, reducing its effectiveness in structured business workflows. The pullback from large language models is a notable moment for CEO Marc Benioff, who has been a vocal advocate of AI-led change. In an interview with Business Insider, Benioff said he is now drafting Salesforce's annual strategy document with data foundations, not AI models, as the top priority. He cited concerns around "hallucinations" when AI systems operate without proper data context. Benioff also suggested Salesforce could one day rebrand itself as "Agentforce," telling Business Insider "that would not shock me," after learning from focus groups that customers are tired of hearing about cloud computing. Salesforce shares are down about 34% from their December 2024 peak. Still, the company expects Agentforce to generate more than $500 million in annual revenue. Salesforce's decision to scale back large language models reflects a broader industry moment, as companies try to bridge the gap between AI hype and dependable, day-to-day business use. (You can now subscribe to our Economic Times WhatsApp channel)