Image courtesy of Adobe Stock
Pharma and biotech are fielding AI pilots across discovery, clinical and commercial, but the gains still hinge on having the right plumbing, culture and talent. That means integrated platforms, governed data, and trained users. The harder truth: AI tends to magnify dysfunction rather than fix it. This foundational work is especially vital when AI-driven recommendations are put in the hands of field teams, where inconsistency can undermine the entire strategy. “We don’t want representatives that are in the field to be analysts,” says Conexus CEO Akshay Kapadia, in an interview. “We need to be able to provide them some very specific, prescriptive information of what they should be doing so actions are consistent.”
Demand is rising, even as readiness varies
While “AI bubble” talk has lingered for years, Kapadia says executive pressure is real and growing in biopharma. “The mandate is there for most of our customers to get into this space,” he says. “The CIOs are telling their teams, ‘What can you show us?'” The result: “I don’t see it slowing down. If anything, I feel the opposite. There is a lot of traction that I’ve seen in the last couple of years.”
Akshay Kapadia
That traction, though, can sometimes collide with uneven maturity. Many teams still juggle siloed systems and patchy data governance. The result can involve promising proofs of concept that stall at scale, or pilots that never graduate to production because the data model or workflows aren’t ready.
Kapadia points to an emerging “start small” pattern, especially in platforms that live where the data is. At Veeva’s recent R&D and Quality Summit in Boston, he observed that “AI was at every session that was presented at the summit.” The company outlined a roadmap to embed scores of agents directly into the platform. “It’s a bunch of agents that Veeva is trying to put and build into the system,” Kapadia explains. “A percentage of those will be done by Veeva as out of the box, and the others would be configured, customized by partners like us for customers specific to their needs.”
The target is on measurable task automation within existing workflows, not moonshot reinvention. “One of the goals stated by Veeva CEO Peter Gassner was improving productivity by about 20% by 2030,” Kapadia notes. Ambitious, but plausible when agents sit next to governed data and standard processes instead of calling across brittle integrations.
There’s an old saying about data being the new oil, but the pendulum is swing around towards having the right structure for data. The volume is almost a given. “The amount of information that is now available just continues to grow at a rapid pace,” Kapadia observes. “As more and more data is available in a structured format, as opposed to all across the board, spread out in different systems and all of that, which used to be a story of the past. With all of the data being available, there’s more that you can do with that data.”
The platform split: vertical depth vs. horizontal AI
Salesforce is also leaning into agents via its Life Sciences Cloud. Kapadia expects the next couple of years to clarify differences between the approaches: Veeva pushing deeply vertical with agents that run inside its clinical and commercial data context; Salesforce bringing horizontal AI infrastructure and then layering life-sciences-specific capabilities on top. For buyers, the practical question is less about logos and more about data gravity: which stack already houses your core records and will minimize context drift.
On the commercial side, Conexus’s home turf, Kapadia sees steady progress in field force effectiveness. “Next best action, based on the data and the trends,” is how he describes it—systems that “recommend to the sales rep what the next best action is based on the historical data that is available.” The trick is less about algorithmic cleverness and more about fit and trust.
“What you don’t want is different interpretations and therefore different actions being taken out of the results,” he says. “We need to be able to provide them very specific, prescriptive information… so that there is consistency.”
That implies clear governance around data sources, feature definitions and override rules. In other words, biopharmas need training plans that respect how reps actually work.
Clinical and quality: dashboards before “do-everything” agents
Even outside Conexus’s primary focus, Kapadia sees the value of cleaner systems of record and live dashboards that reduce spreadsheet herding. “The systems today are way ahead of” the Excel-driven past, he notes. “Forget AI, just having a system of record where all the data is available with easy-to-access dashboards to look at status and all that just makes a very, very big difference.”
In multi-stakeholder environments like clinical trials, simply centralizing status and handoffs yields tangible wins. Those priorities come before you ask an agent to automate cross-team tasks. Mature organizations sequence the work: stabilize the source of truth, standardize the workflow, then insert agents where they won’t trip on missing context.
Life sciences constraints shape AI design. “In the pharma world, things are very regulated,” Kapadia explains. “You cannot have data that goes out into the ether, so to speak. There’s confidential information, there’s firewall information, there’s certain information even within a pharma company that cannot be shared between teams, you know, commercial teams and their medical science liaison (MSL) teams and all of that. So those controls have to be in place.” That pushes teams toward agents that operate close to governed data, with audit trails and explicit boundaries.
Avoiding pilot purgatory: vision first, then projects
Kapadia’s playbook starts with a clear vision, not a shopping cart of isolated experiments. “Otherwise, customers will start one-off projects and get limited success,” he says. But there’s more: “Having a vision, putting together a plan, and especially in pharma and life sciences, doing things in a compliant manner, those are very, very critical before launching any projects. Because otherwise, you may get some results, but whether you can scale those solutions or not becomes a challenge.”
The sequence he favors begins with defining the outcomes and compliance requirements. Second, map the data you’ll trust and the workflows you’ll change. Third, pilot narrowly where the data is strongest. Then, train users on how to interpret outputs and when to escalate. “You need to make sure that people are trained to utilize and interpret the outcomes from these programs,” he emphasizes. Finally, scale only when the results are consistent across teams. “AI can come up with very fast results, but then how do you leverage that becomes critical.” It sounds unglamorous. It’s also how pilots become products.
Agentic AI is still nascent, but it is gaining real momentum in pharma, especially where platforms already anchor the data and the work. But the leaders won’t be the ones with the flashiest demos. They’ll be the ones that made the comparably “boring” investments: integrated platforms, governed data and trained users, so that the exciting stuff actually sticks.
