Thomas Celerier
Pharmaceutical companies are not short of artificial intelligence initiatives. Most have discovery models, generative AI copilots, data-platform programs, and dozens of proofs of concept underway.
What the industry lacks is something more difficult: AI systems that survive contact with scientific uncertainty, fragmented enterprise data, regulatory scrutiny, and the realities of daily work. This distinction matters.
The next era of pharmaceutical innovation will not be won by the company with the most models. It will be won by the company that can turn complex evidence into trusted decisions faster than its competitors, without compromising patient safety, scientific integrity, or regulatory confidence.
The economic opportunity is significant. It is estimated that generative AI could create between $60 billion and $110 billion in annual value across pharmaceuticals and medical products (McKinsey). But capturing that value requires far more than deploying a large language model or accelerating an isolated task. It requires redesigning how scientific and operational decisions move through the enterprise.
The real bottleneck is shifting
AI is already demonstrating that it can contribute meaningfully to drug discovery.
In 2025, a randomized Phase 2a trial of an AI-discovered drug target combination for idiopathic pulmonary fibrosis demonstrated safety and early signs of efficacy (Nature). It was an important milestone because it moved the conversation beyond computational promise and into clinical translation.
But it would be a mistake to conclude that the discovery challenge has been solved.
AI can generate hypotheses, identify patterns, prioritize targets, and propose molecules at unprecedented speed. It cannot eliminate the need for biological validation, representative data, carefully designed experiments, or clinical judgment.
In fact, faster hypothesis generation can expose a new bottleneck: the organization’s ability to validate, contextualize, and operationalize what AI produces.
A model may identify a promising biomarker. But can the organization trace the underlying datasets? Can researchers reproduce the finding? Can the insight be connected to imaging, multi-omics, clinical, and real-world evidence? Can it be translated into a protocol, submission, or treatment decision? Can the system explain where the model is confident, where it is uncertain, and when a human must intervene?
This is why the most valuable unit of AI in pharma is not a prediction. It is a decision that can be defended.
Data must evolve from an asset into an evidence system
The industry frequently describes its data problem as fragmentation. It is accurate, but incomplete.
Pharma data is not simply distributed across systems, but across scientific contexts. The same patient, compound, or therapeutic hypothesis may appear differently in discovery platforms, laboratory systems, imaging repositories, clinical databases, safety systems, manufacturing applications, and real-world datasets.
Each environment has its own terminology, metadata, quality thresholds, access rules, and regulatory significance. Moving everything into a cloud platform does not automatically resolve those differences. A data lake can centralize information while preserving the confusion that existed before it.
AI-native pharma therefore requires more than a data platform. It requires an evidence architecture: a governed layer (including permissioned sources and a knowledge graph/ontology to reconcile terminology across contexts) and an evaluation dataset that signals when an output can be trusted and connects data with meaning, provenance, context of use, and decision rights.
This architecture should answer several questions continuously:
Where did this information originate? How has it been transformed? Which population does it represent? What assumptions were made? Which model used it? How has that model been validated? What decisions may the output support? Who remains accountable for the result?
Without these answers, AI may accelerate analysis but weaken institutional trust.
This is why data platforms, governance, and organizational change cannot be treated as separate programs. They have to progress together across discovery, clinical development, regulatory, manufacturing, and commercial operations.
In regulated AI, observability is a competitive advantage
Regulatory requirements are sometimes portrayed as barriers to AI adoption. A better perspective is that regulation is helping define what enterprise-grade AI must become.
The FDA’s draft guidance for AI supporting drug and biological product decisions proposes a risk-based credibility framework tied to a model’s specific context of use. The European Medicines Agency has similarly emphasized data representativeness, bias mitigation, generalizability, performance monitoring, and the management of model drift. In early 2026, the FDA and EMA announced common principles intended to support good AI practice in medicine development.
The implication for pharma leaders is clear: governance cannot be added after the model has been built; it must be engineered into the system.
This means maintaining data lineage, versioning models and prompts, recording human interventions, defining acceptable performance thresholds, monitoring degradation, and establishing clear escalation paths. It also means separating probabilistic AI recommendations from deterministic business rules wherever quality, safety, or regulatory commitments require certainty. Running evaluations against representative datasets and capturing agent traces are equally important.
A well-designed AI system should not be less transparent than the manual process it replaces. It should be significantly more observable. Every important output should leave behind an evidence trail: what the system reviewed, how it reached its recommendation, which controls were applied, what changed, and who approved the final action.
Such a level of observability does more than satisfy compliance teams. It reduces organizational friction. Scientists, clinicians, quality leaders, and regulators are more likely to use AI when they can interrogate it rather than merely trust it.
Agentic AI must earn its autonomy
Agentic AI is likely to become an important part of the pharmaceutical operating model. Potential applications include scientific literature synthesis, protocol development, site feasibility, regulatory intelligence, safety case processing, quality investigations, deviation triage, supply planning, and medical content workflows.
Several analyses suggest that 75-85% of pharmaceutical workflows contain tasks that could potentially be enhanced or automated by AI agents (McKinsey; ILO). Yet the operative word is tasks, not entire functions.
Pharma should not race toward unrestricted autonomy. Autonomy should be earned in stages.
An AI system may first operate as a copilot that prepares information for expert review. It may then become a constrained agent authorized to complete bounded activities within predefined rules. Only after demonstrating reliability, traceability, and consistent performance should it execute selected low-risk and reversible actions automatically.
The question should never be, “Can this agent perform the task?”
Rather, “Under what conditions should it perform the task? What evidence does it need? What could go wrong? How will we know? And where must a human remain accountable?”
The future is not human versus AI. It is a deliberately engineered collaboration between human expertise, deterministic controls, and probabilistic intelligence.
What production experience teaches us
At Coditas, we have seen that the greatest value rarely comes from a model operating in isolation. It comes from integrating the model into a complete data and workflow environment.
In one engagement for a research organization working in discovery and translational imaging, multimodal MRI, PET, and SPECT analysis depended heavily on manual interpretation. The process faced motion artifacts, variability, limited specialist capacity, and difficulty scaling research across therapeutic programs.
Coditas developed machine-learning-driven image alignment, deep-learning models for isolating brain regions, AI-based denoising, and automated processing pipelines. The solution achieved 87% accuracy in automated image analysis, increased workflow efficiency by 60-70%, reduced cost per scan by 80%, and enabled more than 40 concurrent research projects.
The lesson was not simply that an algorithm could analyze an image. The value came from making heterogeneous imaging data consistent, improving its quality, automating repetitive steps, and delivering outputs that researchers could incorporate into real scientific work.
A second Coditas engagement involved an integrated clinical-genomics platform. The solution combined AI-supported patient screening, education, test ordering, results analysis, post-test counseling, bidirectional HL7 integration, and automated documentation. According to the case study, it increased the identification of at-risk patients fivefold, reduced genetic-testing turnaround time from six weeks to 1.5 weeks, and contributed to more than $50 million in additional revenue within a year.
Again, AI was only part of the answer.
The business and clinical impact came from connecting patient experience, interoperability, decision support, documentation, and operational workflows into one coherent platform.
These examples reinforce a fundamental principle:
AI creates value when the entire decision pathway is engineered, not merely the intelligence at its center.
Five leadership choices will separate the leaders from the experimenters
First, fund business outcomes rather than disconnected use cases. A portfolio of pilots may demonstrate activity without producing transformation. Select decisions and workflows where improved speed, quality, or accuracy can materially affect development timelines, patient outcomes, regulatory performance, or operating cost.
Second, treat data products and AI controls as reusable enterprise capabilities. Identity, consent, terminology, lineage, validation, observability, and human-review mechanisms should not be rebuilt independently for every project.
Third, redesign workflows before automating them. Adding AI to a fragmented or inefficient process can create a faster, fragmented process. Leaders should challenge handoffs, decision rights, duplicated reviews, and unnecessary documentation before introducing agents.
Fourth, measure adoption and trust alongside technical performance. A model with excellent benchmark accuracy but low user adoption creates little value. Organizations should monitor override rates, time saved, decision quality, error recovery, user confidence, and downstream outcomes.
Finally, build co-innovation models that unite domain experts with product, data, and engineering teams. Pharmaceutical transformation cannot be delivered through a sequence of technology handoffs. Scientists, clinicians, quality leaders, regulatory professionals, experience designers, and engineers must build and learn together.
The destination is a learning enterprise
The AI-enabled pharmaceutical enterprise will not be defined by a single platform, model, or agent, but by its ability to learn across the full product lifecycle.
Discovery insights will inform clinical strategy. Clinical and real-world evidence will refine scientific hypotheses. Manufacturing and quality data will improve process understanding. Patient and provider experiences will feed back into development and commercialization. The feedback loop is the true strategic opportunity.
The organizations that succeed will not merely perform today’s work faster. They will create an operating system in which evidence moves continuously, decisions become more informed, and every interaction improves the next one.
AI will play a central role. But data discipline, human-centered design, responsible governance, and strong digital engineering will determine whether it becomes another layer of experimentation or a durable source of scientific and business advantage.
The future of pharma will not belong to the company with the most AI. It will belong to the company that can convert intelligence into trusted action, repeatedly and at scale.
