Insights

Agentic AI in Drug Discovery: Why We Must Start from Scratch

Agentic AI in Drug Discovery: Why We Must Start from Scratch

By Yair Benita, CTO, AION Labs

AI has achieved remarkable progress in biomedical research, yet its impact on pharmaceutical R&D has been marginal. The reason is structural: pharma workflows, built over decades for manual processes and human decision-making, are incompatible with automation and AI. Agentic AI in a modular team-based architecture of interacting software agents and humans is the future of building robust processes for drug discovery and development. But to realize its full potential, the pharmaceutical industry will not be able to simply digitize the existing processes and workflows; the entire system needs to be redesigned for a new era.

A Technological Advance Meets an Architectural Wall

AI is transforming biology across the spectrum, including protein folding, molecule design, gene expression modeling, and clinical trial optimization. But these breakthroughs are slow to translate into faster or better drug discovery. In practice, AI often sits in an isolated box, powerful but disconnected from the workflows that drive drug candidate selection, regulatory documentation, or clinical trial design.

The issue isn’t the science, but the inability to integrate AI tools into the organizational infrastructure. Consider Microsoft and OpenAI’s CoPilot. As a standalone tool, it’s a capable AI assistant for writing and search, like outsourcing tasks to an external expert. But to be useful in a pharma company, one needs to understand how things are done: internal standards, terminology, and expectations. That requires deep access to organizational documentation and context. While this may be feasible for text, it becomes highly complex when applied to scientific data and workflows.

Pharma’s infrastructure was built for a different era, one of benchwork, paper trails, and human-led decision-making. No matter how powerful the model, if the surrounding workflow is complex and manual, AI won’t deliver its full value.

Retrofitting Doesn’t Work

Digital transformation efforts in drug discovery are progressing incrementally, digitizing documents, unifying data warehouses, and layering predictive tools. But results have been underwhelming. Most processes remain non-modular, heavily manual, with non-standardized data. The cost of rebuilding the process properly is typically too high to justify.

Take antibody optimization as an example, a workflow that we at AION Labs identified as suitable for AI. This process occurs after antibody hits have been discovered and affinity has been improved. Building AI into just this step in the process keeps the rest of the legacy workflow intact. However, in reality, once a process is triggered, it is difficult to plug in a new technology, which was seen as a disruption of a small component that could not be justified.

The proper approach is not to replace a single component in a process but rather rebuild the entire process from scratch, generating optimized, high-affinity antibodies from a selected target of interest. Rebuilding such a workflow is possible and requires many different types of AI tools, de-novo antibody generation models to make hits, optimizers to improve the sequences, and refining models coupled to experimental testing. Such an approach would allow parallel execution of the process and direct comparison of the outcomes.

Agentic AI: Rebuilding the Team

Agentic AI offers a new operating model. Rather than relying on a monolithic single black-box model trained for a specific task, agentic systems are composed of specialized agents that function like a digital team. Each agent performs a well-defined task, analyzing data, writing regulatory text, creating visualizations, or formatting documents. These agents are orchestrated, monitored, and validated, just like human experts.

Consider a familiar, well-defined task: writing a clinical study report. The traditional process includes:

  • A statistician reviewing the data and generating an analysis,
  • A clinician  interpreting the results,
  • A regulatory expert drafting the document,
  • A project manager ensures timelines and quality.

Agentic AI replicates this structure, with the added benefits of speed, reproducibility, and scalability. It also allows the workflow itself, not just the output, to be optimized. However, agents are not autonomous systems. They require human oversight, validation, and intervention throughout.

To make this shift, we must let go of a common misconception: AI agents are not traditional software. Classic software is deterministic; it behaves the same way every time. Generative AI does not. It produces different outputs depending on context. It learns, generalizes, improvises, and makes mistakes. This isn’t a bug; it’s a core feature of how it works.

As such, we must treat AI agents like digital team members, not code. They require:

  • Training, like new hires
  • Supervision and validation, like junior scientists
  • Auditing and reviewing, like any expert’s output

Our corporate systems are already designed for this. Rather than treating AI as software, we should treat it as a junior employee and provide onboarding, guidance, and feedback, which are essential for continuous improvement. Managing agents are more like managing people than managing code.

The Case for Starting From Scratch

Studying different industries, the pattern for starting from scratch provides a clear benefit. Tesla, for instance, didn’t just swap in an electric motor; it reimagined mobility with a vision for autonomous driving. Similarly, ESH Bank is not just a digital bank; it’s a cloud-native banking platform designed to become the operating system of modern banks. Dana Farber Cancer Institute (DFCI) is building a new AI hospital from the ground up, allowing it to design how medicine will be practiced in the age of AI. These organizations have a better chance to succeed not because they integrate AI, but because they build a new paradigm that leverages the advances in technology.

Pharma should do the same by creating independent AI-native entities and empowered teams tasked with reimagining every step of drug discovery. These entities should be small, fast-moving, and self-contained, focused on building robust platforms that make drug development faster, cheaper, and smarter.

AI-native organizations require rethinking roles. Scientists become curators and reviewers of AI-generated output. Engineers must understand scientific logic and constraints. Teams jointly own agent development, validation, and improvement. The goal isn’t automation but a true human–machine collaboration.

Culture, Not Just Code

One of the biggest barriers to change is cultural. Pharma organizations are structured around document-centric workflows, formal hierarchies, and risk reduction, all of which result in long timelines. AI thrives under very different conditions: feedback loops, rapid iteration, shared ownership, and modular systems.

AI-native organizations require rethinking roles. Scientists become curators and reviewers of AI-generated output. Engineers must understand scientific logic and constraints. Teams jointly own agent development, validation, and improvement. The goal isn’t automation but a true human–machine collaboration.

How Do We Start?

The case we make here points to a bold but necessary approach: creating new companies from scratch. However, because this requires significant effort and commitment, it makes sense to first test core principles before fully scaling the model.

At AION Labs, for instance, we learned that challenges we define for AI should not target narrow tasks embedded within larger pipelines. Instead, they should focus on complete workflows. For example, rather than isolating antibody optimization, we should define the entire antibody discovery process, from target selection to therapeutic lead. This represents a departure from current challenge framing and would require rethinking our definitions. Yet it would allow us to test not only individual models, but also the integration of a full, AI-native workflow, making successful outcomes more impactful and easier to scale. Meaning, we need to focus on delivering an agentic platform and not only a novel AI technology.

A parallel approach within a pharmaceutical company could involve building independent internal teams tasked with reinventing specific, well-defined workflows, such as writing clinical study reports. These teams should operate autonomously, with access to the data, talent, and resources required for success. They should be freed from traditional process constraints to iterate, adapt, and co-develop agentic systems alongside domain experts.

While these efforts may demand more time and investment upfront, they offer the best chance at meaningful and lasting change. They are not pilots; they are a path to a new paradigm.

Conclusion: Rebuilding the process

Agentic AI is not a plug-in for existing processes; it’s an opportunity for a new kind of drug discovery organization. One where humans and machines collaborate in teams. AI agents run the operational and repeatable tasks while humans review and make decisions. Realizing this vision means walking away from legacy systems and processes. We have to start from scratch.