Over the past few weeks, Salesforce has found itself at the center of an ongoing AI debate: can large language models (LLMs) be trusted to run critical business processes? Recent reporting suggested Salesforce was quietly losing trust in LLMs when operating Agentforce. Salesforce says that’s not quite right.
The Information shared a quote from Sanjana Parulekar, SVP of Product Marketing at Salesforce, which states that the company had “more trust in the LLM a year ago” compared to now. They reported that this has encouraged a more deterministic approach to Agentforce automation, where it functions on a set of predefined instructions. Yet Salesforce says this framing is an oversimplification of how it is actually approaching Agentforce.
Madhav Thattai, COO of Salesforce AI, told Salesforce Ben that the strategy is not about losing trust in LLMs, but about gaining greater control over how agentic AI operates in real enterprise environments. As more enterprises move from experimentation to production, Salesforce believes success depends on striking the right balance between LLM-driven reasoning and deterministic execution.
In our conversation, Madhav outlined how that balance is shaping Salesforce’s true vision for Agentforce, and why it matters for enterprises looking to deploy AI agents at scale.
Why the “Trust in LLMs Is Declining” Narrative Landed
It’s not that difficult to see why the idea that trust in LLMs is slipping struck a chord with some Salesforce or wider tech readers. Hallucinations are happening very often, and prompts are sometimes answered so confidently despite being plain wrong. In an enterprise setting, those mistakes aren’t always obvious until the damage is already done.
When we have AI agents involved in situations like financial transactions, healthcare decisions, or customer support at scale, those small errors still matter. A system that works “most of the time” can quickly become a liability if it fails in subtle, unpredictable ways. That’s especially true when mistakes can be introduced quickly and without users realizing anything has actually gone wrong.
Peter Chittum, Technical Content Director at Salesforce Ben, and I acknowledged this gap last year, when Salesforce CEO Marc Benioff said its AI agents were operating at around 93% accuracy. This might be an impressive figure on paper, but it falls short when measured against enterprise reliability benchmarks, such as Six Sigma, where near-perfect consistency is the goal (99.999%).
Against that backdrop, it’s easy to understand why comments suggesting Salesforce was rethinking its reliance on LLMs gained traction. But according to Madhav, this interpretation may be missing the bigger picture.
“The LLM Alone Is Not Sufficient”
When outlining Salesforce’s hybrid approach, Madhav explained that LLMs haven’t suddenly become untrustworthy. In fact, it might be quite the opposite.
“First and foremost, LLMs are a remarkable technology. They’ve unlocked our ability to communicate in natural language with software systems, which has changed the richness of the experience as we are interacting across channels – whether it’s voice or text,” Madhav explained.
“And then, really importantly, it’s also changed how we develop and deploy software in a really significant way. This is really meaningful. There should be no question about the role of LLMs and the importance of LLMs in this whole journey that we’ve been on over the past couple of years.”
The shift, he argues, comes from what Salesforce has learned as customers move beyond demos and pilots into real production environments.
“As customers are starting to move from experimentation into ‘is this technology going to drive ROI for me? And in what way am I going to get value from it?’ [It’s] really driven our product innovation, our investments, and where we think we want to take this going forward. From our vantage point – and we’ve said this from the start – the LLM alone is not sufficient.
“When we think about agent infrastructure, there are critical pieces that have to sit around the LLM to make an agent truly enterprise-grade. First is the data layer. A foundational LLM is never going to have all of a customer’s proprietary, structured, and unstructured data, so agents need secure, permissioned access to that context to be accurate.
“Second is process execution. We rely on LLMs for rich communication and flexible reasoning, but as workflows get more complex, their ability to reason consistently can waver.
“Third is governance. As customers scale, analytics, KPIs, and optimization become critical – and those are things you can’t rely on the LLM alone to provide.”
A 90%+ accuracy rate might seem good at first glance and even acceptable in early experiments, but in production – especially in regulated industries – it simply isn’t up to the required standard. As Madhav explains, Salesforce has seen firsthand that as workflows become more complex, LLMs struggle to consistently follow multi-step processes without drifting or dropping instructions.
“In industries like financial services or healthcare, 90% consistency is not okay,” Madhav said. “That’s why we pair LLM creativity with deterministic execution, so agents can be flexible when needed and precise every time.”
In essence, it feels as though Salesforce is simply looking to be much more deliberate about where LLMs are being used.
Why a Hybrid Approach Is Vital to Agentforce Success
Salesforce’s answer has been to lean into a hybrid reasoning feature called “Agent Script” – one that uses LLMs where they shine best, and deterministic systems where precision matters most.
LLMs still power the conversational layer, allowing agents to understand intent and adapt to different types of phrasing. But when it comes to executing core business functions, Salesforce will increasingly rely on deterministic logic to ensure consistency.
“We rely on LLMs for reasoning and intent, but one key observation from working with customers is that as processes become more complex, their ability to reason consistently can start to waver.
“In cases like Adecco – where agents are qualifying candidates for jobs – LLMs are great for creating a rich, natural interaction, but the underlying workflow has to be followed precisely. That’s why our approach to process execution is the best of both worlds; using LLMs for flexible reasoning and communication, while ensuring deterministic execution where accuracy really matters.”
In many ways, this reflects a broader and newer understanding of enterprise AI and how it’s really going to perform well. Early enthusiasm focused on what LLMs could do. Now, the conversation is increasingly about what they should do and what they shouldn’t really be trusted with yet.
When asked to summarize Salesforce’s stance on LLMs going forward, Madhav had this to say:
“We believe agentic technology will be transformative for businesses, and that hasn’t changed. Agentforce is central to how Salesforce will power these agentic enterprises as customer experiences evolve. LLMs play a critical role in enabling rich interactions and complex reasoning, but trust remains our number one value.
“These agents will handle some of the most important business processes and most sensitive data, so trust can’t be overstated. For us, enterprise-grade trust means trusting the data and context an agent has, trusting that workflows execute consistently, and trusting that customers have the visibility and control to manage, optimize, and improve agent performance over time.”
Final Thoughts
Whether this more structured approach will win over any skeptics remains to be seen, but early signs have shown customers are buying into the vision. Salesforce says more than 18,500 customers are adopting or deploying Agentforce, following a strong Q3 in which AI-driven revenue momentum continued to build late last year.
The framing that Salesforce has “lost trust” in LLMs, in reality, doesn’t tell the full story. What’s emerging instead is a more mature and realistic view of how generative AI should be used inside the enterprise, shaped by real-world deployments and risk tolerance.
Rather than dialing back any ambition, Salesforce appears to be refining its approach to optimize the Agentforce experience going into 2026, where LLMs are used effectively.
