Big Pharma companies are still betting big on AI, with major players pouring resources into agent platforms. The goal: unify fragmented data, streamline operations from clinical trials to sales, and enable more personalized healthcare provider interactions.
Novartis has selected Salesforce’s Agentforce Life Sciences for a global rollout over the next five years. The deal follows deals from AstraZeneca, Haleon and Takeda.
Pharma companies have spent years accumulating data across separate tools and teams: clinical operations, medical affairs, commercial sales, patient support. “The way many pharma firms have had their environments to date has been very siloed or fragmented data across different personas,” said Tara Helm, VP and GM of Commercial for Agentforce Life Sciences. “It’s really about how they can have a seamless flow or orchestration from clinical through commercial, through medical affairs, through marketing.”
The proposed fix goes deeper than just a user interface. “Agentforce Life Sciences is one data model across everything that we’re supporting, from clinical through commercial through patient,” Helm said. “It enables us to really access the data across with all the right security.” That unified model is designed to move CRM from a “system of record,” where the rep has to go in and enter everything, to a “system of insight,” Helm said.
The near-term applications include post-call documentation, where a rep dictates notes and an agent drafts a summary; flagging potential medical inquiries for logging; and automating tedious tasks in clinical study management. “How do you find your trial sites? How can we help automate that? … How can you find your patients?” Helm asked, pointing to patient finding and site selection as key areas for agent automation.
Adverse event detection is coming
Adverse event detection, one of the most consequential potential applications for AI in pharma, requires special care. Safety reporting carries regulatory obligations. Missing an adverse event mentioned in a sales call can create serious compliance problems.
Helm described a vision where an agent functions like a “compliance officer in the background,” flagging language that suggests a safety issue and prompting the rep to log it properly. The functionality isn’t “live today, but it’s something we’re absolutely looking at.”
This gap between vision and reality highlights a core tension in generative AI: the conflict between its “probabilistic” nature, generating plausible-sounding outputs, and the “deterministic” requirements of legal compliance, which demand consistency and accuracy. Salesforce’s own experience underscores the problem. After reducing its customer support headcount by roughly 4,000 in 2025, a move the company characterizes as redeployment rather than layoffs, executives have increasingly acknowledged that it can be “hard to build reliable agents for enterprises.” In 2024, Salesforce touted how “quick and straightforward it is to build autonomous AI agents with Agentforce.” That year, CEO Marc Benioff wrote in Time of agents: “They can perform tasks independently, make decisions and even negotiate with other agents on our behalf. And unlike the traditional tech transformations of the past which required years of costly infrastructure buildout, these new AI agents are easy to build and deploy, unlocking massive capacity.”
While Benioff remains bullish on agents, other executives have acknowledged the challenges involved in agentic projects in enterprise settings. “All of us were more confident about large language models a year ago,” SVP of Product Marketing Sanjna Parulekar was quoted as saying in The Information in December. A Salesforce spokesperson clarified the company’s position: “We’ve learned from real customer deployments that the LLM alone is not sufficient. To be enterprise-ready, an agent needs critical infrastructure around the model that provides trusted data, governance and control.” “Agents when you first build them […] are interns at best,” Parulekar said in a YouTube video. In a subsequent blog post, she was more direct: “Language models are exceptional at understanding intent and context but they are, by design, probabilistic.” She continued: “They generate likely outcomes, not guaranteed ones. In low-stakes scenarios, that’s acceptable, even desirable. In core business operations, where policies, regulations, and customer trust are non-negotiable, it’s a liability.” She went on to add, “In these environments, hope is not a strategy. You can’t cross your fingers that your AI behaves correctly. You need to know it will, every single time.” She also highlighted that the nature of the technology requires “the right architecture to ensure that AI plays by the rules.”
Customizing agents
Agents can also require considerable customization. Salesforce markets Agentforce with pre-built capabilities for life sciences, but Helm acknowledged that customers typically don’t deploy it unchanged. “I’d be surprised if anyone took our off-the-shelf agent that comes with our product and just used it,” she said. “They’re probably going to add their company’s own flavor and their own roles to it as well.”
That customization extends to compliance. Helm described the platform’s guardrails: role-based data access, flagging for specific terms, Salesforce’s “Einstein Trust Layer” for data masking—but stressed that large customers add their own governance on top.
The same ambiguity surrounds AI agents more broadly. In June 2025, Gartner warned that over 40% of agentic AI projects could be canceled by 2027 owing to costs and unclear ROI, noting that most current efforts remained “early stage experiments or proof of concepts.” By August, the firm projected task-specific agents in 40% of enterprise apps by 2026—a bullish forecast that sits uneasily alongside its own cancellation warning. McKinsey in November found 62% of companies were at least experimenting with agents, but in any given business function, fewer than one in ten were deploying them at scale.
For now, the Novartis deal is a marker of where budgets are flowing: toward AI agent platforms, with long timelines and high expectations.
Filed Under: Uncategorized
