Why AI Fails in Drug Development and How to Build Tools That Actually Deliver Real Value

Artificial intelligence (AI) is increasingly being applied in drug development, and regulators are addressing its use as more regulatory submissions incorporate AI. Broader conversations often focus on value: improving development timelines, reducing costs, enabling better-informed decisions, and lowering the risk of late-stage failures; however, these benefits are not guaranteed by technology alone.

AI delivers the most value when paired with deep drug development expertise and designed to operate within real-world constraints, supporting the decisions scientists already have to make and fitting within workflows they trust. This requires designing and validating tools in close partnership with domain experts.

Without grounding AI in these operational realities, implementation can become tool-driven, with progress measured by models and outputs rather than outcomes that matter in practice. This perspective reflects lessons learned while building Intrepid Labs, an AI-driven drug formulation company, and collaborating with Quotient Sciences to embed AI within an established development program.

Challenges in the identification and optimization of drug formulations

Formulation provides a useful lens for the potential value and challenges of AI in drug development. A successful formulation balances multiple performance parameters, including storage stability, dose consistency, drug efficacy, and toxicity.

Because of these competing goals, formulation becomes a multi-objective, real-world optimization problem with unavoidable constraints. The difficulty of predicting the optimum composition means that formulation development programs are typically exercises in trial and error.

Many teams default to familiar solutions that exist for good reasons. However, these defaults also narrow the search space and can leave better options unexplored—this is where AI can add value. By complementing and extending how formulation teams work, AI enables them to better navigate higher-dimensional spaces, learn efficiently from sparse and noisy data, and adapt as new data arrives.

The key question is: How do we design AI tools for formulation that operate within real-world constraints and deliver value inside established development workflows?Answering this requires the active participation of experts who have been doing this work for decades.

Practical takeaways from an AI integration

A good test of whether AI can deliver value under constraints is to embed it within a workflow that already operates at clinical speed. Intrepid collaborated with Quotient Sciences to adapt an AI tool to operate within the workflow of an integrated drug development platform and guide decisions under the platform's constraints.

The hardest work was not the model, but integration: defining the specific decisions the system should support, encoding the constraints that determine feasibility, capturing the right data and metadata to make learning reliable, and delivering outputs that scientists and clinicians can actually use. That integration work is where domain expertise is essential.

What goes wrong with AI-first or AI-only approaches

When AI projects fail in complex, high-stakes domains, it’s often because the tool was not designed around real-world constraints from the outset. Consider IBM Watson for oncology. The intent was commendable: help clinicians navigate a massive and growing cancer literature and support treatment decisions.

However, this clinical recommendation model was trained using a small number of synthetic, hypothetical cases rather than real patient data, and its outputs were influenced by the expertise of a limited set of specialists. By prioritizing textbook scenarios over the diversity of real-world patient histories, the model failed to generalize reliably across the complexity of clinical care, leading to unsafe or incorrect treatment recommendations.

Drug development faces similar challenges. One example is generative chemistry—while many generative models can propose novel structures, chemical synthesizability is often deprioritized rather than made a first-class design constraint. Consequently, the system produces molecules that look great on screen but are not practical to make at the bench.

The pattern is consistent. AI-first approaches fail when they are not anchored in the constraints and judgment that domain experts already bring to the table.

Guidelines for successful integration of AI into established workflows

Based on our experience, we’ve compiled practical guidelines for integrating AI into drug development programs in a way that delivers real value.

• Start with the workflow, not the model: Map the decisions the team needs to make, who makes them, what information they rely on, and what constraints shape what is feasible.
• Define “good” with the people who own outcomes: Define success metrics and trade-offs with domain experts, so that the tool optimizes what the program actually needs.
• Treat constraints as high-priority design inputs: Build feasibility constraints into the optimization loop from the start. This includes drug and material availability, stability windows, manufacturability, assay throughput, timelines, and regulatory requirements.
• Use real-world data, and capture metadata that makes learning transferable: Incomplete metadata and inconsistent experimental context can quietly undermine performance and limit transferability. Furthermore, models must be trained on positive and negative data to strengthen learning.
• Evaluate as if you plan to use it: Validate the tool under conditions that resemble prospective use. Avoid evaluation setups that look strong on paper, but do not reflect deployment reality.
• Design for adoption, not just performance:Teams are more likely to adopt tools when expectations are clear, training is built in, and early use cases show that the tool reduces low-value work and strengthens expert decision-making.
• Keep humans in the loop, and make accountability explicit: AI should support decisions, not obscure responsibility. Be clear about who owns the decision, how outputs are reviewed, and how models are monitored and updated over time.
• Partner early with domain operators: Co-design with domain experts from the start. Partnerships with groups that run real programs under real constraints are one of the fastest paths to building tools that actually work.

Conclusion

AI is not replacing people. The opportunity in formulation and drug development is enormous; however, the path forward is not AI in isolation.

It is collaboration: building tools that strengthen expert judgement and help teams make better decisions faster, in the environments in which those decisions actually matter. Over time, once tools have earned trust and demonstrated value, teams can build on that foundation to evolve workflows together.

Done well, that evolution brings everyone along and converts decades of development experience into scalable, AI-enabled ways of working.

About the Authors

Christine Allen, PhD, is co-founder and CEO of Intrepid Labs.

Andrew Lewis, PhD, is Chief Scientific Officer of Quotient Sciences.