If AI is going to be a game changer for drug discovery and development, a hard question must be answered: How? Simply adding technology to a decades-old R&D process rarely produces lasting change. Since AI can profoundly improve patients’ lives, getting it right is critical. On McKinsey’s Eureka! podcast, six leaders spanning big pharma and academia, and one of McKinsey’s own China experts, describe what it takes to get AI past the demo stage and into the daily work of discovering and delivering medicines.
Investment in AI drug discovery has grown sharply, but much of it is flowing toward problems that are technically easier for AI to tackle. As our recent research argues, funding has concentrated on molecule design, where structured data and rapid experimental feedback make AI tools more mature. The industry’s central scientific bottleneck lies elsewhere: identifying and validating the disease mechanisms those molecules must address. Designing a molecule faster offers little benefit if it targets the wrong biology.
Turning that speed into clinical success will require greater investment in target validation, better biological data, and closer integration across the drug discovery process. What distinguishes the organizations that are pulling ahead is how they’ve rebuilt data foundations, governance, talent, and culture around AI.
On the Eureka! podcast, these guests shared their perspectives on what’s working, what’s hype, and what’s still unsolved:
- Christian Diehl, chief data and digital officer, biomedical research, Novartis
- Sai Jasti, senior vice president and head of data science and AI, Bayer
- Bob Li, senior vice president, global head of medical affairs, oncology, AstraZeneca
- Greg Meyers, executive vice president and chief digital and technology officer, Bristol Myers Squibb
- Fangning Zhang, partner, McKinsey, Greater China Life Sciences Practice
- James Zou, associate professor of biomedical data science, Stanford University
Speakers highlighted several common themes as they discussed data platforms, cancer care, AI research, and operating in China.
What’s changed since last year
Last year’s series made the case that bolting AI onto business as usual wouldn’t deliver, and that success would rest on a set of enablers spanning strategy, technology, data, talent, and change management. This year’s series built on that case, with podcast guests describing practical challenges they faced as they made AI part of everyday research and development.
From whether to adopt AI to how to do it. With the case for redesign largely made, experts focused more on what happens after the commitment and why so many efforts still stall. Bayer’s Sai Jasti urged companies to move from “being opportunistic” to “committing” to AI. Bristol Myers Squibb’s Greg Meyers explained why that transition can stall: Companies experiment but struggle to put those tools into everyday use, leaving them in “prototype purgatory.”
Agents are moving into day-to-day work. Last year, guests raised questions about which decisions AI agents could make independently and when humans should intervene. This year, guests shared examples of agents at work. They talked about overseeing them “almost like associates” and using other agents to check their output for errors. In one virtual lab, agents designed nanobodies—small antibody fragments—that researchers then tested in a physical lab.
A sharper eye for hype. Podcast guests were confident in AI’s long-term potential but called out where today’s reality falls short. This included concepts like “cool does not scale,” “agent-washing,” and AI’s limits in the parts of biology that haven’t been sampled.
Beyond individual companies. Guests also pointed to China’s rise as an innovation engine and to patient access as the real measure of success.
Data remains the bottleneck. Our experts spoke more candidly about the years of work needed to clean and organize research data before AI tools can use it.
Five lessons that separate scaling from stalling
1. Make AI “boring,” and be honest about what’s real
A successful AI rollout makes AI routine and baked into everyday work. Novartis’s Christian Diehl said, “My experience, to be provocative, is that cool does not scale.” He described watching how demos wow a steering committee, then checking back six months later to find nothing had been implemented. The fix is to make AI “part of the daily process rather than something you marvel at every day.”
The gap between the demo and reality shows up in the broader market, but Bristol Myers Squibb’s Meyers offered a contrarian view. He rejected the idea that AI is already reshaping drug discovery itself.
“I think there’s too much hype on the idea that an agent running in the background can solve big problems in disease,” Meyers says. “We know that AI is not good at extrapolating outside of data that has already been sampled, and there’s an awful lot about biology that hasn’t been sampled.” He also flagged “a lot of agent-washing going on in the market”—tools that automate a subsection of a task being marketed as fully agentic systems.
Meyers called the gap between experimenting with AI and putting it into everyday use “prototype purgatory.” It’s easy to experiment with AI because the stakes feel low, but “infusing it back into core processes is terrifying for a lot of companies,” he said. Bristol Myers Squibb tackled this by forming AI accelerator teams of six to eight people, working in six two-week sprints. The company treats “even failure” as “a success,” so employees can test ideas without staking their reputations on whether a project succeeds.
Bayer’s Jasti framed it as a mindset shift rather than a technology one. “It’s no longer about being opportunistic, which is how we often approached these capabilities in the past. Now, we are committing to these capabilities.”
Being honest about what’s real matters just as much as embedding AI day to day. Diehl pointed to predictive safety modeling, generative chemistry, AI-assisted literature review, and protocol writing as areas already delivering. Jasti cited a 50 percent speed-up in early target discovery and a shift from four-to-five-month regulatory dossiers to under four weeks using generative AI, all “without compromising quality.” AstraZeneca’s Bob Li described AI turning research that used to take one to two years into same-day analysis. “You ask a question in the morning, and you have a draft analysis by afternoon.”
Stanford’s James Zou offered a counterpoint: AI agents in his “virtual lab” designed nanobodies against COVID-19 variants that outperformed some earlier human-designed candidates in lab tests.
McKinsey’s research on AI drug discovery supports Meyers’s caution. Investment has focused on molecule design, where structured data and mature tools make AI easier to apply. But designing better molecules won’t resolve two major causes of failure: targeting the wrong disease mechanism and finding that results from preclinical studies don’t hold up in humans.
2. Better models are helping address the data bottleneck
Preparing data for AI has historically required years of cleaning and organizing. Newer foundation models and agentic capabilities are making some of that work more feasible, helping teams curate and structure data that previously was difficult to use. These advances can ease the preparation burden, though they do not eliminate the need for reliable biological evidence.
Diehl’s team spent many years building data42, a clinical trial data lake now underpinning predictive safety work. But he was candid that the harder, ongoing task is “cleaning, curating, and harmonizing data to make it AI-ready. Having a lot of data is one thing; having AI-ready data you can actually leverage … is a very different thing.”
Jasti described a parallel effort at Bayer to prepare instrument and partner data for AI analysis, freeing scientists to explore ideas rather than spend time “getting data ready across ten different systems.”
Meyers offered the cautionary version of the same lesson. “Applying technology to a broken business process just makes the broken process move faster,” he said. Even as AI helps prepare data, infrastructure and process discipline are essential to producing reliable results.
3. Agentic AI, a new kind of workforce, needs guardrails
Several guests compared AI agents to employees who need defined access permissions and oversight. Agents “need to be governed almost like associates,” pointing to the open question of how to build “a replicable audit trail” defensible to health authorities, said Diehl. Zou’s team built a “virtual lab” that mirrors his physical Stanford lab. It includes an AI professor, AI students with different specialties, group meetings, and even a “critic agent,” whose job is to poke holes in the other agents’ reasoning to catch hallucinations before they compound.
Some organizations are moving more slowly. Jasti said Bayer was still in the early phases of adopting agentic AI; the company uses “co-scientist” tools to mine scientific literature. Bayer is also building AI agents for medical writing, and Jasti expects fuller agentic systems “maybe in a year or two.”
Meyers said AI agents are most advanced in code development. But even leading reasoning models, he said, “are not great at taking that leap two or three levels beyond the core of what your research question is.”
Rather than a binary human-in, human-out choice, guests described a spectrum. Diehl framed it as a range running from moments where a human must make the final call to full automation once a process is validated. Stanford’s Zou described the complementary split he sees in practice: AI agents are good at breadth, creatively recombining existing tools; humans are better at depth, building genuinely new methods from scratch. Meyers said Bristol Myers Squibb requires every employee to read its AI policy up front. Employees must keep humans in the loop and check outputs for accuracy. People remain accountable for decisions.
4. Talent and geography are both being redrawn
The talent conversation is moving past hiring more data scientists and more in favor of hybrid roles. Jasti said Bayer is hiring change management specialists who understand human-machine interaction, as well as people who can combine scientific and technical expertise, including MDs who are also data scientists.
Diehl described Novartis’s AI Early Career Initiative, which rotates new graduates through biomedical research for two years to build breadth. He also invoked a famous idea in AI called Moravec’s paradox: Tasks humans find difficult can be easy for AI, while tasks humans perform instinctively can be hard for machines. He described AI as a “copilot” whose skills complement a human scientist’s. Zou suggested that many companies could adopt AI agents faster by working with partners, given how quickly the underlying technology is moving.
Geography is reshaping where that talent and R&D activity concentrate. McKinsey partner Fangning Zhang pointed to China’s evolution from a commercial market into a genuine innovation engine. More than 30 percent of the industry’s innovative pipeline now originates there, driven by an intense focus on speed and cost efficiency. Zhang called this the industry’s “ten and ten” problem: Developing a drug takes ten years, with a success rate below 10 percent and costs running into billions per approval. Those pressures demand faster, more efficient development.
Zhang added that China hasn’t yet cracked novel biology at scale. Additionally, she noted that industry AI experts see China as largely in “catch-up mode” on AI-driven discovery and lab automation. But the trajectory is steep. The country’s pipeline now has roughly 40 to 50 first-in-class molecules, up from just one more than five years ago. That estimate now has hard cost data behind it. McKinsey Global Institute research puts a Chinese biopharma company’s levelized R&D cost at roughly 2.7 times lower than a multinational running the same fast-follower program, driven mainly by faster development timelines and lower wages rather than any single advantage.
For multinational R&D organizations, Zhang’s advice focused on operating-model design: building genuine local decision-making capacity without losing global coherence.
5. AI should improve patient access, not just speed
AstraZeneca’s Li offered the clearest reminder that infrastructure and speed are means, not ends. His starting point was stark: Fewer than 7 percent of cancer patients get access to a clinical trial, even though trials represent guideline-endorsed best care. “The bottleneck isn’t budget,” he said. “It’s a very long, tortuous process of drug development.”
His Medical Affairs Revolution Spirit (MARS) mission spans next-generation evidence generation, education, “Building Cancer Networks,” and international collaboration. It uses AI and decentralized trial infrastructure to close that access gap, not just to speed up drug development. It’s a useful check on the other four lessons. The data platforms, agent governance, and talent strategies only matter insofar as they get treatments to more of the patients who need them.
Facing the risks
Every podcast guest working inside a regulated company pointed to a similar set of concerns. Diehl addressed privacy directly, saying, “We’re actively addressing patient data privacy and genetic data.” He also highlighted the risk of agents inadvertently surfacing proprietary intellectual property. Zou described how his team reduces hallucinations by using critic agents, running parallel meetings to reach a consensus, and keeping agents “within bounds” in a sandboxed environment.
Jasti was unequivocal about prioritizing quality, saying, “Compliance and regulatory issues are always front and center … We never compromise on quality.” Meyers said real-world experimentation reveals where a system breaks down and what happens when you give it too much control.
The road ahead
The six conversations distinguish organizations reimagining R&D by building the underlying capabilities—data infrastructure, governance frameworks, hybrid talent, and regional operating models—from those layering AI onto processes that haven’t changed. Call them “builders” and “bolt-ons.” As redesign becomes the stated intent almost everywhere, what increasingly separates the two is follow-through. The builders are willing to improve data quality through timebound sprints, to redesign organizational structure around agents the way they would around new hires, and to treat failed experiments as information rather than embarrassment.
Zhang’s closing framework for what separates leaders applies well beyond the China question she was asked. It comes down to a clear ambition, an operating model wired to enable that ambition, and the willingness to keep learning and adjusting as the ground shifts. Diehl put forth a more compact version of the same idea: “We’re chasing impact, not novelty, though sometimes the latest novelty does have the biggest impact.”
Li brought the discussion back to patients. He described AstraZeneca’s ambition to eliminate cancer as a cause of death in his lifetime “so that our grandkids can live in a much better world, one that isn’t fearful of cancer.” AI in life sciences R&D will be judged, ultimately, not by how sophisticated the agents get, but by how many more patients gain access to clinical trials.
Explore the latest episodes of the Eureka! podcast here or on your favorite podcast platform.

