Reliant Life Sciences’ Vice President Ben Locwin recently joined host Ross Katz on the Data in Biotech podcast to discuss how artificial intelligence is being applied across drug development, manufacturing and clinical research.
The conversation hit on a familiar pattern across the industry: While nearly every biotech organization now claims some level of AI adoption, far fewer are applying it in ways that significantly improve outcomes. The gap between aspiration and execution is where most teams are currently operating. And that’s a big gray space.
Below is the full episode, followed by key takeaways from the discussion.
The Gap Between AI Claims and Real Application
Locwin, who has spent more than 25 years working across the biotech landscape, framed the current moment as part of a recurring cycle. New technologies emerge, expectations rise quickly and organizations move to align themselves with what feels like the future.
Today, that pressure is centered on AI.
Many companies want to avoid being perceived as behind the curve, particularly in front of investors or partners. The result is broad adoption in name, but uneven execution in practice. Rather than defining clear use cases, some organizations are layering AI onto existing processes without an established objective.
A more effective approach, as Locwin shared, is to start with a defined problem and build toward a targeted application. That requires more discipline upfront, but it creates a path to measurable value.
Where AI Is Working Today
Not all applications are equal. According to Locwin, the most mature use cases are still concentrated in areas where the underlying systems are highly structured and quantifiable (such as molecular modeling, discovery, and screening)
Molecular modeling is a clear example: Advances in protein structure prediction and computational chemistry have demonstrated what AI can achieve when the inputs are well defined and the relationships are grounded in established science. These environments allow models to produce outputs with a high degree of precision.
As the industry moves further downstream, that clarity begins to fade.
In manufacturing and regulatory strategy, AI is often applied to more variable and less predictable systems. Many current implementations rely on simplified modeling of process parameters such as pH, dissolved oxygen, nitrogen, or fluid and gas flow rates, and other manufacturing parameters. While useful, these approaches are closer to advanced analytics than transformative AI.
This distinction matters. Labeling something as “AI” does not inherently make it more effective. The underlying methodology still determines the value.
The Challenge of Applying AI in Regulated Environments
One of the most persistent challenges discussed in the episode is the role of AI in regulatory processes.
By design, regulatory frameworks leave room for interpretation. They define boundaries of compliance, but within those boundaries there are multiple acceptable approaches. That flexibility is essential for scientific progress, but it creates ambiguity for AI systems that rely on clear rules and consistent patterns.
As a result, applying AI to regulatory strategy or submission processes is more complex than many vendors suggest.
The idea of “one-click” regulatory submissions continues to gain attention, but the reality is much more nuanced. Even as tools improve, regulatory agencies themselves are still determining how to evaluate AI-assisted outputs. And they are themselves using AI tools to review and evaluate the AI submissions. Until those frameworks mature, automation in this space will remain limited by uncertainty on both sides (the ‘Jagged Frontier’)
Clinical Trials and the Problem of Variability
Clinical research introduces another layer of complexity.
Unlike molecular systems, clinical trials involve human subjects with significant variability across demographics, physiology and external factors. This creates a level of biological noise that is difficult for models to interpret with precision.
As Locwin explains, AI models in this environment can quickly move from narrow, well-defined outputs to broader and less predictable ranges of results. Managing that variability becomes the central challenge.
One path forward is better trial design. Expanding or refining patient populations, incorporating decentralized trial models and applying design-of-experiments principles can improve how data is generated and structured. When done well, this gives AI systems more data to learn from.
However, more data alone is not the answer. The quality of the data matters just as much as the quantity. Poor inputs will produce unreliable outputs, regardless of the sophistication of the model.
Why More Detail Doesn’t Always Improve Outcomes (The Paradox of Added Detail Worsening Outcomes)
A common assumption in AI development is that adding more variables will lead to better predictions. In practice, this often leads to diminishing returns.
Overly complex models can become difficult to interpret and may introduce noise that obscures the underlying relationships. In some cases, increasing the level of detail actually reduces the usefulness of the output.
Locwin points to the importance of focusing on causality rather than correlation. Identifying whether a drug produces a specific effect in a defined population is ultimately the goal. Variables that do not contribute to that understanding may not improve the model, even if they increase its complexity. In fact, they often worsen the model’s results.
This principle is well established in statistics and remains relevant in modern AI applications. More information (an ‘overfit model’) is not inherently better. The right information is what matters.
An Industry Still Defining Its Approach
Across the biotech landscape, organizations are actively experimenting with AI, but there is no single playbook emerging yet.
Different teams are exploring applications across preclinical research, clinical trial optimization, manufacturing and quality systems. In many cases, these efforts remain disconnected, with separate models operating at different stages of the development lifecycle.
There is also a growing recognition that not every process needs to be optimized with AI. In some cases, existing systems are already effective, and the cost of replacing them may outweigh the benefits.
This is where a more measured approach becomes important. Applying AI selectively, based on clear value and feasibility, is more likely to produce tangible results than broad, unfocused adoption.
Moving Forward with Clarity
The current wave of AI adoption in biotech is not defined by a lack of innovation. It is defined by a lack of clarity.
Organizations that take the time to define their objectives, understand their data and align AI applications with real operational needs are more likely to see progress. Those that adopt AI as a broad mandate risk adding complexity without improving outcomes.
As the technology continues to evolve, the gap between potential and execution will narrow. In the meantime, the most effective teams will be the ones that remain disciplined in how they apply it.
