Strategic Advisor | Ex-FDA | Helping Leaders Translate Life Science and Technology Innovation into Adoption and Impact | AI-Enabled Development | Strategy, Systems & Leadership
What AI changes in drug development
Why the next generation of model-informed development will depend on context, credibility, connected evidence and accountable workflows
Before artificial intelligence became the dominant language of drug development, the industry was already asking a more fundamental question:
How can models help us improve processes and make better decisions under uncertainty?
We can all agree that drug development has never suffered from a shortage of data. The challenge has always been about turning incomplete data and uneven evidence into decisions about dose, population, trial design, manufacturing, investment and whether a program should continue at all.
While AI expands what models can see and do, it does not remove the need to define the decision, understand the evidence, test the assumptions or remain accountable for the outcome.
The unit of value in AI-enabled drug development is therefore not the prediction, analysis or generated output. It is the decision that the output improves, together with the evidence, oversight and accountability surrounding that decision.
The question predates AI
In their 2015 article, Model-based clinical drug development in the past, present and future, Holly Kimko and José Pinheiro described clinical development as a largely empirical and costly enterprise in which decisions often relied on qualitative assessments of risk without fully using the information generated across a program.
Model-based drug development offered a different approach. Models could integrate evidence from studies and external sources, quantify important relationships and simulate outcomes that had not yet been observed. PK/PD, PBPK, disease-progression and clinical-trial models could inform first-in-human dosing, dose selection, patient populations, endpoints and study design.
The strategic contribution was not prediction for its own sake. It was the ability to explore assumptions before committing patients, capital and time. A model could help teams ask how a proposed trial might perform under different dropout rates, treatment effects, dose-response relationships or design choices. It made uncertainty more visible and allowed alternatives to be examined before a decision became difficult to reverse.
That principle remains central to model-informed drug development today. The final ICH M15 guideline, issued in June 2026, now provides a harmonized framework for planning, evaluating and documenting model-informed evidence. The terminology has evolved from model-based to model-informed development, but the core idea is consistent: models contribute evidence to a decision; they do not make the decision independently.
From analytical tool to connected operating system
The role of models also expanded beyond clinical development.
In Industry 4.0 for Pharmaceutical Manufacturing: Preparing for the Smart Factories of the Future, published in 2021, my FDA colleagues and I examined how connected sensors, process analytical technologies, artificial intelligence, robotics, advanced computing and real-time data integration could reshape pharmaceutical manufacturing.
In this environment, the model is no longer only an analytical tool used to review a completed experiment. It can become part of a cyber-physical system that continuously receives process data, interprets conditions, supports predictions and informs control. Manufacturing can become more responsive, with the potential for deeper process understanding, improved quality assurance, greater flexibility and more efficient production.
But embedding models inside an operating system changes the consequence of error. A weak prediction in an exploratory analysis may prompt another experiment. A weak prediction influencing process control, batch disposition or product quality may affect supply and patients. The closer a model sits to a consequential action, the more important its boundaries, data, validation, monitoring and governance become.
This is one reason AI cannot be treated as a separate digital initiative. When models influence clinical, regulatory or manufacturing decisions, they become part of the development system itself.
What AI changes
AI significantly broadens the computational canvas. Models can work across chemical structures, multi-omics, imaging, clinical data, scientific literature, real-world evidence and manufacturing-process data. They can identify patterns across data volumes and dimensions that traditional approaches may struggle to integrate.
Across the product lifecycle, AI may help to:
- identify or prioritize targets and candidate molecules
- connect biological signals across modalities and datasets
- support patient stratification and response prediction
- optimize protocols, recruitment and clinical operations
- analyse real-world data and digital endpoints
- monitor processes and support the selection of manufacturing conditions
- detect safety or quality signals across complex information streams
Generative systems add another capability: they can propose new structures, hypotheses, analyses and text. This can compress the time required to explore a larger decision space.
Yet faster exploration is not the same as better development. A model may find a strong statistical pattern without establishing biological causality. It may perform well in training but fail in a new population, site, assay, process or operating environment. It may encode biases, depend on data that are not representative or change in performance as inputs drift.
AI can accelerate the production of answers. Development strategy still has to determine which questions matter.
From model output to development workflow
AI adoption is not the same as AI-enabled transformation. A model may perform well while the process surrounding it remains poorly defined: the wrong question enters the system, important sources are omitted, review responsibilities are unclear or the output cannot be traced to the decision it influenced.
The implementation challenge is not simply to place AI inside an existing process. It is to redesign the workflow so that the right evidence reaches the right decision, with proportionate human review, traceability and accountability. That requires process and decision mapping, defined source and data requirements, review and escalation points, exception handling, documentation, change management and measures of whether the workflow actually improves decision quality.
This distinction matters because speed, adoption and usage are incomplete measures of value. The more consequential test is whether the AI-enabled workflow reduces a decision-critical uncertainty, improves consistency, preserves important options or helps the organization act with greater confidence.
Credibility belongs to a context, not to a model
FDA’s 2025 draft guidance on AI used to support regulatory decision-making makes an important distinction. The credibility of an AI model is not treated as a universal property. It is assessed for a particular context of use: the model’s specific role and scope in addressing a defined question of interest.
The proposed framework begins with the decision, not the algorithm. It asks sponsors to define the question, specify the context of use, assess model risk, establish a credibility plan, execute and document that plan, and determine whether the model is adequate for that use. The level of evidence and oversight should be commensurate with the consequences of an incorrect model output and the extent to which the decision relies on it.
This decision-centred logic is also visible in the joint FDA-EMA Guiding Principles of Good AI Practice in Drug Development published in January 2026. The principles emphasize human-centric design, a risk-based approach, clear context of use, multidisciplinary expertise, data governance, performance assessment and lifecycle management.
Together with ICH M15, these developments point toward a broader convergence. Whether a model is mechanistic, statistical or AI-enabled, confidence must be built through a relationship between purpose, evidence, performance, risk and decision consequence.
A highly accurate model used for an ill-defined purpose may have limited strategic value. A simpler model, transparently bounded and supported by relevant evidence, may be far more useful for a specific decision.
From model performance to decision architecture
This shifts the central question away from whether an organization has adopted AI.
The more useful question is whether the organization can use AI to improve consequential decisions without weakening scientific credibility or accountability.
Before relying on an AI-enabled model, development teams should be able to answer seven questions:
1 Decision What decision will the model inform, and who owns it?
2 Context and workflow Where does the model sit in the process, who uses it and within what boundaries?
3 Consequence and reversibility What happens if the output is wrong, and how difficult would the resulting commitment be to reverse?
4 Evidence Are the data relevant, representative, traceable and fit for the intended use?
5 Credibility What validation, performance assessment and uncertainty characterization are proportionate to the risk?
6 Oversight and traceability Who can interpret, challenge or reject the output, and can the decision be reconstructed?
7 Lifecycle and value How will performance, change and the effect on decision quality be monitored?
These questions connect model governance to development strategy. They also expose dependencies that may otherwise remain hidden. A patient-selection model may depend on assay performance and population representativeness. A manufacturing model may depend on sensor reliability, process understanding and change control. A clinical prediction may depend on data generated under conditions that do not match the future setting.
Seen through a Pathway Intelligence lens, the model is one component within a wider system of evidence, assumptions, dependencies, decision rights and downstream consequences.
Decisions create commitments
Every development decision commits some combination of patients, capital, time, technical resources and organizational attention. Inevitably it will also close or preserve future pathways. AI can help teams explore more scenarios before making that commitment, but its strategic value depends on whether it helps distinguish between the decisions that are reversible and those that are difficult or impossible to unwind. All while keeping an end goal of patient access and adoption in view.
UNCERTAINTY > EVIDENCE > DECISION > COMMITMENT > OPTIONALITY
A model supporting an exploratory experiment may tolerate greater uncertainty. A model influencing process control, pivotal-trial design, patient selection or a regulatory conclusion requires a much higher level of evidence and oversight. The objective is not simply to identify the apparently optimal answer. It is to generate evidence proportionate to the consequence of the decision while avoiding the premature closure of valuable future options.
A hypothetical CMC example
Consider a monoclonal antibody program in which an AI model identifies a cell culture configuration predicted to increase yield and reduce variability. The immediate technical question is whether experimental evidence confirms the prediction. The development decision is wider.
The team must determine whether the proposed configuration is robust at commercial scale; whether it introduces dependence on specialized equipment, raw materials or data infrastructure; whether it could affect critical quality attributes or the control strategy; and what evidence would be required before incorporating it into regulatory submissions. The team must also consider technology transfer, comparability, COGS and whether the choice can be reversed if later evidence challenges the model.
The value of AI is not merely that it identifies a potentially better process condition. Its value lies in helping the team explore the decision space. Experimental evidence, process understanding and human review still determine whether the recommendation is credible enough to act upon and whether acting now preserves or constrains future pathways.
Connected evidence matters more than more data
AI can integrate information at a scale that was difficult to imagine when model-based development first emerged. But integration is valuable only when the underlying evidence has been designed and governed well.
This is where AI-enabled development connects directly with evidence architecture. Evidence should be generated not simply because it can be collected, but because it reduces a decision-critical uncertainty. Its provenance, quality and limitations must remain understandable as it moves across models and teams. The sequence still matters: analytical validity may precede clinical usefulness; process understanding may precede autonomous control; external validation may be necessary before a prediction is generalized.
The danger is that AI creates the appearance of coherence by combining fragmented data into a confident output. A polished prediction can conceal weak evidence, mismatched populations or unresolved assumptions. The quality of the interface does not determine the quality of the decision.
This makes traceability more than a documentation exercise. Teams need to reconstruct how an output was produced, what data and model version were used, what human interventions occurred, what assumptions were made and how the output influenced the final decision. Responsibility cannot migrate to the algorithm or to the technology vendor.
The next generation of model-informed development
The future is unlikely to be a choice between traditional models and AI. The strongest development systems will combine them.
Mechanistic models can encode biological and physiological understanding. Statistical and pharmacometric models can quantify variability and exposure-response relationships. Machine learning can detect complex patterns across high-dimensional data. Generative AI can accelerate exploration and synthesis. Experiments and clinical studies can test whether predictions hold in the real world. Manufacturing knowledge can anchor models in the conditions required to produce a consistent product.
Human judgement remains responsible for connecting those forms of evidence to the purpose, constraints and consequences of the decision.
This requires more than technical capability. It requires multidisciplinary teams, shared definitions, process ownership, governance that follows the risk of the use case, and leaders who understand when a model should influence a decision and when it should not. It also requires organizational capability: people must be able to interrogate outputs, recognize failure modes, document judgement and learn from outcomes rather than simply accept recommendations.
Organizations that build these capabilities will not simply run more AI pilots. They will create learning systems in which evidence, models, experiments and decisions continually inform one another across development and manufacturing.
The strategic objective has not changed
The technology has changed considerably since the early development of model-based approaches. The strategic objective has not.
Drug development still requires teams to integrate evidence, make uncertainty visible, compare plausible pathways and improve decisions before they become expensive or irreversible.
AI can generate more possibilities, more predictions and more apparently confident answers. It cannot determine which uncertainty matters most, which evidence is sufficient for the decision or which future options should remain open. Those remain development-strategy questions.
The measure of an AI-enabled model is therefore not simply how accurately or quickly it produces an output. It is whether the model, embedded within a credible and accountable workflow, leads to a better, more defensible decision for the program and ultimately for patients.
Where could AI most improve decision quality in your development pathway, and what evidence, oversight and organizational capability would you need before trusting it there?
Selected sources
- Kimko H, Pinheiro J. Model-based clinical drug development in the past, present and future: a commentary. British Journal of Clinical Pharmacology. 2015;79(1):108-116. Link
- Arden NS, Fisher AC, Tyner K, Yu LX, Lee SL, Kopcha M. Industry 4.0 for pharmaceutical manufacturing: Preparing for the smart factories of the future. International Journal of Pharmaceutics. 2021;602:120554. Link
- US Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Draft Guidance. January 2025. Link
- International Council for Harmonisation. M15 General Principles for Model-Informed Drug Development. Final Guidance. June 2026. Link
- FDA and EMA. Guiding Principles of Good AI Practice in Drug Development. January 2026. Link