AI in Biopharma Strategy: Opportunities, Challenges, and What Separates Success from Failure

Graham:

Large language models (LLMs) are genuinely powerful at one thing: processing large volumes of unstructured data efficiently, which the biopharma industry generates in abundance.

With LLMs, we can now access unstructured data in a way we simply couldn't before. Previously, for instance, as a business development professional, if you had 1,000 potential in-licensing opportunities in your database, you might realistically be able to work through only 200 or 300. Now it's possible to interrogate the full 1,000, and that can increase the likelihood of identifying the right opportunity with the right team, rather than just the best opportunity you could find with available capacity.

That said, LLMs' analytical capability isn't very high; they can't connect ideas and make the inferential leaps in the same way as experienced strategists. As they continue to improve, though, they have the potential to monitor option spaces in real-time, continuously updating the picture as new data emerges. In a sector where things can shift quickly, that kind of continuous intelligence is invaluable.

Jonathan:

LLMs can't really drive strategy; you still very much need people with deep industry knowledge and experience for higher-order strategic thinking, for unveiling those 3–5 insights that drive most of the decision-making value. And there aren't even that many individuals in the industry who can do that effectively. What AI can do well, though, is generate hypotheses for experienced strategists to validate and to further complement their analyses.

If you can use AI to complement hypothesis generation, you can reduce human error in the groundwork. Combining the two means you can make your strategic output both higher quality and more robust.

Wignand:

I see an LLM, on its own, similar to a fresh junior hire — someone who is highly motivated but lacks significant context and experience. They'll work tirelessly with whatever you give them, sure. But the output depends entirely on how well you instructed them on what data and evaluation approach to conduct during onboarding. To that point, Gartner predicted that 60% of AI projects in 2026 would be abandoned due to a lack of AI-ready data, which really emphasises the need to provide relevant context.

At the moment, I see their value in rapidly going through, and bringing semantic understanding to, unstructured data such as news and trial updates. This data is messy and variable, and otherwise can be somewhat impenetrable at scale. The opportunity, then, is in unlocking these huge datasets to better inform strategic decisions.

Graham:

One thing is that, in biopharma, there are infinite possible combinations of ideas and concepts, and the bad combinations won't necessarily have been written down or are subject to publication bias. For example, people in our industry just know that using cytotoxic antibody-drug conjugates in Alzheimer's Disease (AD) is generally a bad idea, but an LLM won't necessarily make that inference independently, because it likely hasn't been documented. The lack of clear causal logic in the text leads to haphazard risk identification. Building in the kind of expert reasoning — the unspoken knowledge of what works, what doesn't, and why — is pretty painstaking work that can only be done by a world-leading team.

Then comes cycle times, which are the principal challenge of our industry. Going from idea to marketed drug typically takes 8–15 years. So if you're trying to train an AI system to identify the best drug development candidates, it has to wait a huge amount of time for feedback. Sure, you can optimise for nearer-term goals, like clinical go/no-go decisions, but if all the candidates fail in the clinic, the optimisation was pointless. Compare this to software development, where LLMs can write code, test it against outcomes, and iterate thousands of times in hours. In biopharma, it's just very challenging.

Jonathan:

In other industries like engineering or marketing, you also have more straightforward processes and hard binary readouts. In biopharma strategy, it's much more complex, and non-linear readouts are on a spectrum. Therefore, experienced interpretation of this complexity, often in a discussion with peers, is critical to aligning on the most likely correct interpretation. This, combined with long cycle times, can make it almost impossible for many companies to optimise LLMs towards ground truth.

Biopharma also has a bias problem, which LLMs can struggle with. Because LLMs consume text rather than raw or processed data, when they analyse research papers, they're internalising what the authors say about the data, rather than what the data in graphs actually says. The difference is rarely black and white, but subtle shades of grey. In addition, the scientific literature is knowingly biased towards positive results; sometimes, the absence of information is also informative. An LLM consuming that literature, especially summaries such as abstracts, without any human checks or input, just absorbs those biases, which could have profound implications for the overall strategy.

Wignand:

The bias problem is getting worse, too. The largest uncontrolled data source is the internet, and an increasing proportion of content on it is now LLM-generated. So an uninstructed LLM searching the web absorbs both human bias and LLM bias. It becomes an echo chamber, and one that's very difficult to escape. Taken together, this strongly highlights that context and data management is critical for an effective biopharma LLM strategy agent.

Wignand:

The thing that's often underestimated is the importance of starting from a strategy process that already works. Companies may try to invent both the strategy methodology and the AI system at the same time, without knowing if either will work, often thinking that 'better/newer' models directly yield more accurate results. If you're translating a validated process into an LLM rather than building it from scratch or relying on the model for the ground truth, you've already de-risked a huge part of the task.

Graham:

You also need structured, well-documented records of past decisions to serve as a benchmark for results validation. But most organisations don't have this. Information is often scattered across emails, meeting minutes, and presentations, and there's no clear record of why decisions were made. Having that structured data in place, alongside the industry-specific expertise to recognise a good output, is what makes validation possible.

Beyond that, it's a lot of trial and error. You build, you test, and you find failures you never would have anticipated. For example, I asked an LLM to match two lists of company names, controlling for abbreviations, synonyms, ownership, etc., and it deleted 140 entries from the output, without telling me, because it didn't know what to do with them. A human would never do that. It's like asking a 5-year-old to put on a jumper; they might put it on their leg, or backward, or inside out. Except with an LLM, there's an unlimited number of ways they can get it wrong. You need a human watching every step of the process.

Jonathan:

And, if you can't validate results yourself, you need someone with the right knowledge and experience to do so. People often think that you can have one person infuse all their knowledge into an LLM and that it will work, but that isn't the case. You have to pressure-test everything that comes out to refine it and deliver the right quality. And these validation cycles are complex and labour-intensive, so you need a whole team of people with industry-specific expertise to make them work.

Wignand:

Identifying the right problem to solve is just as important as how, and how successfully, you solve it, too. A lot of organisations dive into AI deployment without a clear sense of where AI adds value. Once you've identified that, you then have to focus on that specific task, commit to it, and see the testing cycle through properly, rather than spreading effort too thinly across lots of different AI applications and deployments.

Graham:

What excites me most for the future is access to data that was effectively unreachable. Every drug in clinical development is surrounded by thousands of pages of documentation, regulatory submissions, and competitive analyses. The ability to work through that material intelligently, finding what's actually relevant rather than just matching keywords, will be so, so valuable.

Jonathan:

The challenge that will remain, and that I think will become more important as other technical barriers fall away, is the underlying data bias. The highly biased nature of published drug data isn't going to change any time soon. So remaining critical, being conscious of where data is coming from and what might be shaping it, will be a critical long-term skill.

Wignand:

The companies that will get the most out of AI in this space are the ones that can provide the appropriate context, not the ones with the most performant model. This means providing them with the right data and the right instructions, which is only possible with deep expertise in how to approach these problems.

Whatever happens with the technology, keeping the human in the loop will remain non-negotiable. AI in biopharma strategy is ultimately a supportive system — it can take on a huge amount of the heavy-lifting, but the expertise, the critical thinking, and the accountability have to stay with the people. That's not necessarily a limitation of the technology, though. It's just the reality of working in an industry this complex.

Want to find out more about how AI is helping support better biopharma business development decisions? Read our original research article, "A biopharma-specific AI agent for high-stakes strategic decision-making."