[Adobe Stock]
Pharma executives, observing peers in retail and banking tout GenAI payoffs “in weeks and months,” increasingly ask why their own firms can’t match that pace. This pressure for a ROI for an industry that has seen productivity gradually stall since the 1950s.
So, when pharma executives ask Sheetal Chawla, Capgemini’s head of life sciences: “When will I be able to show the ROI of these [genAI] investments?” Chawla recommended reframing the question. That is, embrace new metrics and new ways of thinking. Executives may need to conclude that such investments are difficult to gauge in “the same traditional way that we are measuring value in other parts of the organization.”
That “mindset change,” as Chawla put it, involves treating GenAI pilots less like discrete IT projects and more like building foundational capabilities. Pharma, in many ways, is more complex, with famously long and expensive R&D cycles. And the goal of organizational change requires a significant investment in change management.
Sheetal (Dingrani) Chawla
Beyond ROI, data security is also central. It’s “one of the top two or three things that comes up,” Chawla said. Companies in the highly-regulated and risk-averse sector often hesitant to adopt external LLMs owing to their chief information security officer’s concerns. She also noted that seemingly fundamental tasks, such as simply locating relevant data, are “a very common problem” given the complexity involved, spanning complex labs, clinical trial sites, and beyond. Her advice is to start small with a tightly controlled data environment, repeatedly test model outputs, and then build incrementally from there.
While discipline is involved, that doesn’t mean there isn’t room for optimism. Clinical trial design, patient engagement, and safety as areas with “a lot of investment and progress,” she said. She also noted that “diagnostic tools and also real-time imaging” are additional bright spots.
Start with the precise “why”
While genAI can still be viewed a bit like a bright, shiny object, Chawla recommended a steady focus on business problems when deploying it. Clients often “come to us and say, ‘We want to use AI or LLMs to help with this problem,’ but the problem itself isn’t very well defined.” The firm then helps define that problem, finding that the approach might be “agentic AI, LLMs, GenAI, sometimes automation, sometimes it’s a business process.” It might “not even [be] about AI… It could also be a human problem,” addressable by training or change management.
Tackle data realities before scaling
Once the problem definition step is in place, addressing data challenges comes next. Chawla mentioned Capgemini helps clients with “structured, unstructured data, using agents to build and kind of reverse-engineer knowledge that exists.” This careful data handling, often starting within a controlled scope, is also one strategy to help manage the risk of genAI model hallucinations. While limiting context is important, Chawla acknowledged the challenge: “if it’s too limited, how do you scale the use of what you’re building?… you would have to do that millions of times to achieve scaled value.” The approach, therefore, involves proving value in specific use cases before considering broader replication. Chawla notes that even for promising applications like using agentic AI in dossier submissions, where clients want to “eliminate errors… speed up submission timelines… [and] reduce a lot of manual processes,” achieving widespread impact can be elusive.
Capgemini’s pLLM for protein engineering: Cutting 99% of the data out of the loop
In addition to advising on GenAI strategy, Capgemini has been building its own tools. In a February news release, the company, earlier this year, announced a patent-pending methodology that can cut the datapoints needed for sequence design by more than 99%. On the Green Fluorescent Protein benchmark, Capgemini needed just 43 data points, down from thousands, to find a variant seven times brighter than the wild-type protein. The genAI-basd methodology supported a 60% increase in plastic degradation efficiency thanks to enhanced the cutinase enzyme’s ability to break down PET plastic.
While specific GenAI applications like Capgemini’s pLLM for protein engineering demonstrate progress, Chawla’s overarching message is that pharma’s AI future will be determined less by algorithmic breakthroughs and more by a steadfast focus on fundamental business and patient needs. This requires navigating the “business versus science conundrum,” as Chawla describes it, and continually asking, “what are we solving for? … Are we solving for it in a minutely unique way that isn’t differentiated nor adding value to a patient?”
Filed Under: Data science
