While advances in discovery technologies and computational tools have accelerated molecule design, the more formidable challenge in biopharma is establishing causality in human biology, and thereby improving the probability of success downstream in clinical development. Facing this hurdle, organizations are confronting the reality that productivity in late-stage clinical development has plateaued: Roughly 70 percent of Phase II and Phase III trials run an average of 16 months late.1
A leading contributor to this combination of static productivity and longer development is poor decision quality. Decision quality is defined by how well a choice is framed, how many genuine alternatives are considered, how reliable the supporting evidence is, and how clearly the reasoning is articulated. Clinical development currently falls short on each dimension. Teams make decisions without consistently accessing all the relevant data, which may be fragmented among clinical, operational, and financial systems. High-stakes clinical decisions also remain vulnerable to known judgment biases: teams contend with systemic optimism bias, regularly overestimating Phase III probability of success, while anchoring, champion bias, and sunk-cost fallacies delay critical pivots. The financial toll compounds across portfolio tiers: Suboptimal portfolio governance and delayed program pivots cost top 20 pharmaceutical R&D organizations $200 million to $400 million in lost expected net present value (eNPV) annually.2 Compounded over a decade, our research shows that differences in decision quality of this kind track with divergences in enterprise value exceeding $1 trillion across large-cap biopharma.
Policymakers are taking note. In late September 2026, the US Department of Health and Human Services launched SURPASS, a program aimed at supporting phaseless trial designs. These designs integrate digital twins and predictive models to simulate clinical and operational outcomes before launch, along with the evidence needed to support regulatory confidence in those approaches.3 Pharmaceutical companies are now strongly incentivised to apply such analytical models to drive accelerated clinical development programs with higher likelihood of technical success, but few have yet systematically embedded such approaches.
The solution: A connected, causal decision system
Connecting internal clinical, financial, and operational systems with external sources—such as registries, real-world data, and competitive intelligence feeds—creates an integrated data layer. Yet data integration alone merely reproduces historical precedent. Most existing digital and AI initiatives struggle here because they predict correlations: They identify patterns in past trials but cannot evaluate what would happen if a team changed key trial parameters. Choosing a superior trial design requires counterfactual reasoning—understanding the direct causal link between an intervention and its effect. Causal inference separates what is likely to happen based on current practice or historical precedent, and estimating what will happen under a new set of decisions or conditions the world has not yet seen (see sidebar, “Key decisions at the study level are informed by a range of modeling approaches allowing for causal reasoning”). Randomized clinical trials are themselves considered the gold standard in causal inference, because random assignment enables a comparison between groups that are (on average) equivalent, and pretrial assumptions and estimates can be verified against completed outcomes. Modern causal methods allow for the extension of causal inference beyond strictly randomized settings to answer questions at scale.
A causal, simulation-based decision system is defined by the nested clinical development decisions it connects. Leading R&D organizations are moving beyond isolated spreadsheets, data products, models, and dashboards toward an integrated decision engine that operates on three levels of the enterprise (exhibit).
To demonstrate the interconnectivity of the layers and feedback loop, consider an illustrative oncology asset. At the portfolio level, capital is allocated to the asset alongside a target risk and value profile, and a desired launch plan based on the competitive landscape. The program layer takes in the capital and timeline constraints, optimizes the choice of indications, decides the optimal sequence of indications to pursue, generates the dynamic product profile per indication, and the required end-to-end evidence package and corresponding clinical development plan. The study layer then inherits the indication, its dynamic product profile, a general approval timeline, budget, and other evidence requirements, and simulates to select a design that optimizes within these constraints. If no design reaches the required probability of success (POS) within the budget and the date, that finding travels back up the governance chain to determine which constraints may be relaxed (such as raising the budget, accepting a later readout, or redirecting the capital). Similarly, if new internal or external evidence emerges, this can impact the optimal design or estimations on the trial level, the dynamic product profile to reach, or the necessary clinical development plan. This information ladders back up to the portfolio level, with an updated portfolio profile and need for capital reallocation.
Study-level design optimization
This layer optimizes the design of an individual clinical trial to achieve specific evidence-gathering goals, as specified in the target product profile. Without unified simulation, teams often design protocols through sequential, reactive compromises, requiring iterations over multiple months: For instance, a team may design a trial for optimal biological efficacy that proves too expensive, forcing late-stage trade-offs—such as fewer sites, smaller sample sizes, or overly broad eligibility criteria. Such disconnected planning slows decision-making, white space and trial timelines, and increases failure risk: More than 80 percent of Phase II and Phase III trials undergo at least one significant protocol amendment, averaging three to four per trial across large pharmaceutical sponsors.
The study-level simulation engine addresses this by letting teams systematically evaluate design trade-offs, even live during governance, before locking protocols. Teams can vary key input parameters—including target populations, endpoints, trial arms, statistical properties, operational footprints, and assessment schedules—and simulate innovative alternatives such as synthetic control arms, causal treatment effect estimation, and disease-progression modeling using real-world data. Analytics rely on two complementary methodologies: patient-level treatment effect estimation using causal machine learning, and trial-level treatment effect estimation using data-driven priors and conditional probability. By evaluating complex trade-offs among trial success, speed, and cost, this layer can help teams increase POS for Phase IIb and Phase III trials by five to ten percentage points while shortening trial timelines by three to six months.
In practice, teams apply these causal models across several critical design choices—from optimizing dose-exposure response curves to evaluating head-to-head competitive positioning. For instance, when refining inclusion and exclusion criteria, simulation lets teams evaluate the trade-off between effect size and commercial label breadth across patient subtypes, using prognostic and causal treatment effect models to protect efficacy without unnecessarily restricting recruitment. Similarly, teams can model indirect comparisons against competitor trial readouts using synthetic patient cohorts, testing whether anticipated efficacy margins will hold before committing capital to a head-to-head design.
Optimization does not end when the study protocol is approved. As operational data accumulate and interim analyses report out, simulations can be rerun against observed data rather than design assumptions. Within prespecified adaptive designs, sample size, arm allocation, and site operations may be adjusted based on the live evidence. Each completed study also contributes to the historical performance database.
Many players are already adopting such approaches and delivering tangible benefit. For example, a top ten global pharmaceutical company progressively established patient-level treatment effect estimation capabilities across more than half of its indication portfolio, derisking late-stage trials and making causal simulation a mandatory gate in technical reviews.
Another large biopharma developed a study-level clinical trial simulator minimum viable product (MVP) in ten weeks by integrating existing statistical, enrollment, and financial logic into a single scenario-comparison workflow. Using this unified tool, the asset team evaluated more than 100 simulation scenarios and identified $1 to 2 billion in potential eNPV prior to their final governance gate.
End-to-end asset strategy and program-level sequencing
This layer focuses on the end-to-end clinical development plan for an asset, including the choice and sequencing of indications, selection of combination therapies, and strategic timing for gathering clinical evidence relative to cost—automatically generating dynamic product profiles and corresponding clinical development plans. Evidence here extends beyond registrational trials to real-world evidence, Phase IV studies, patient-reported outcomes, and investigator-initiated studies. The system links individual study outputs directly to asset valuation, enabling an integrated evidence strategy. Teams refine end-to-end evidence planning throughout multistudy programs, ensuring that every trial advances the broader objective of maximizing an asset’s clinical and commercial profile while maintaining an optimal benefit/risk profile.
To evaluate combination regimens, the engine, for example, simulates how a novel mechanism of action interacts with standard-of-care backbones—such as PD-1/PD-L1 inhibitors—enabling teams to assess biological synergy, toxicity risks, and competitive differentiation before committing capital to complex codevelopment programs.4
A midsize specialty oncology biopharma upgraded its indication and combination sequencing processes by synthesizing orthogonal evidence—including scientific literature, real-world data, and biomedical knowledge graphs—alongside automated agents to analyze business dimensions such as unmet need, competition, and developability. This accelerated the end-to-end process by 50 percent and surfaced high-potential indications and regimens not previously considered.
Portfolio-level capital allocation
Scenario modeling is where the decision engine intervenes on portfolio strategy rather than only accelerating it. Efficient frontier analysis narrows thousands of feasible configurations to the three to five most attractive portfolio compositions. Each can then be tested against the events that might affect it (such as clinical readouts, transactions and partnerships, payer/reimbursement changes, and regulatory shifts). The output is a range of POS, revenue, and cost outcomes for every potential capital allocation scenario. A scenario with corresponding asset pipeline can then be selected, with an outlook on the triggers that would change the preferred configuration. In addition to modeling internal portfolio options, business development scenarios can also be included (for example, via simulating target product profiles and clinical development programs for external assets).
The most aggregate level of decision-making optimizes the full portfolio of assets along an efficient frontier of risk versus reward under capital constraints. The engine replaces periodic quarterly reviews with continuous, event-driven governance. Live signals, such as interim trial readouts, competitor failures, or regulatory shifts, trigger automatic recomputation of eNPV. Threshold-based triggers can convert live signals into governance action. For example, when an event moves an asset’s expected value past a given threshold, the system can flag it for a decision instead of holding it for the next quarterly review. Committees can then respond to opportunities and risks worth $100 million or more within days. The same mechanism works on the downside: When conditional power on a Phase III trial falls below a set level, the system flags it for a termination decision, freeing budget months before a final readout. The engine also makes capital availability visible and explicit. Most of a portfolio’s capital is already committed to ongoing trials, making it difficult to see what capital is genuinely available for allocation.
A multinational biopharma implemented portfolio-wide efficient frontier analysis, enabling leadership to optimize risk-return trade-offs and free up more than $2 billion in capital from low-yield development programs. A top ten diversified biopharmaceutical sponsor deployed an event-driven decision capability focused on competitor dynamics, enabling governance committees to reevaluate asset timelines and capital allocation within days of external clinical readouts.
Together, the three layers of study, program, and portfolio levels provide the scaffold and decision ontology for AI agents to prepare these decisions as part of an integrated technical review for human-led governance. This paves the way for simulation-based clinical development, a live technical record, and a push to document the decision-maker’s rationale whenever deviating from the recommendations.
Implementation choices to achieve scale rapidly: Tailoring the system to R&D maturity
When building a simulation capability, leadership teams face three foundational strategic choices that determine implementation speed, organizational buy-in, and long-term impact.
1. Scope and focus: Decision-layer-first versus asset-first rollout
Organizations will first want to decide where to concentrate their initial deployment resources:
- Decision layer-first: Select a single core capability within a given decision layer—such as trial-level POS or enrollment-speed forecasting—and deploy it across the entire enterprise portfolio. This quickly builds broad organizational familiarity, standardizes specific decision inputs, and establishes an enterprise-wide baseline.
- Asset-first: Focus on a single high-stakes disease area or priority asset, building a fully integrated suite of simulation capabilities across all three levels (study, program, and portfolio). This focused approach delivers rapid, high-value impact for high-priority programs and demonstrates the end-to-end value proposition before broader expansion.
The most appropriate choice depends largely on where impact and business value can be demonstrated quickly, organizational capability—whether shortcomings have recently held trials back—and two portfolio features: the degree of concentration around higher-priority asset or disease areas, and the maturity of the highest-priority assets.
A critical subchoice in scope selection is identifying the primary end user. Tailoring the interface and outputs to biostatisticians and clinical scientists emphasizes statistical rigor and protocol parameters, whereas catering to portfolio managers and governance committees requires synthesizing trade-offs into commercial eNPV, risk profiles, and resource-allocation scenarios.
2. Delivery model: Embedded enterprise product versus expert-delivered services
Institutions choose a delivery model aligned with their technical infrastructure and immediate strategic urgency. Embedded products serve as a continuous enterprise system of record for assumptions and scenarios, enabling broad portfolio scalability, though they require upfront system integration. Alternatively, expert-delivered services provide targeted, ad hoc analysis for high-stakes trials or governance gates, offering immediate time to value with minimal IT overhead. Leading organizations often begin with expert-delivered services on high-priority trials to demonstrate value before migrating to a fully embedded platform.
No single vendor can supply a decision system entirely off the shelf; several sponsors have returned to building internally after a licensed product failed to change how decisions were made. A system’s strategic value derives from a sponsor’s proprietary evidence, historical decision records, operational benchmarks, and financial targets. Because underlying analytical methods are advancing rapidly, sponsors will want to avoid locking into a single proprietary vendor—instead, maintaining ownership of their core decision-making backbone while operating a plug-and-play architecture that integrates best-in-class modeling and infrastructure components as they evolve.
3. Modeling architecture: Patient-level versus trial-level dynamics
At the technical layer, institutions select the appropriate analytical granularity. A mature strategy integrates both patient-level and trial-level data, feeding granular patient-level outputs into higher-level trial and portfolio engines. Most organizations are in a mixed position, with patient-level data from their own internal clinical trials and some licensed real-world data, and have aggregate data across their own trials or those of external sponsors. Patient-level modeling uses individual patient data to build granular prognostic and treatment-effect models, using causal machine learning and biomedical knowledge graphs. Trial-level modeling uses Bayesian meta-analytic priors, aggregating historical trial data sets to refine sample-size calculations and simulate operational benchmarks under effect-size uncertainty. The FDA’s January 2026 draft guidance5 signals openness to Bayesian methods and their use in registrational trials, but many of the implementation details (such as how to set priors) remain unsettled. Additionally, organizations have to balance model granularity with scalability: Patient-level models offer greater precision for complex subpopulation effects, while trial-level models scale rapidly across large portfolios.
Operationalizing the causal decision system: Core enablers for enterprise adoption
To deliver sustained financial and strategic returns, treat causal simulation as an enterprise-wide capability rather than an isolated analytical exercise. An AI-powered simulator layered on a flawed decision process amplifies and accelerates the same mistakes with more confidence. Successful implementation depends on six primary conditions:
-
Roadmap and strategy
A complex, multiyear program to build and maintain a layered decision system requires a clear, business-led road map. Sequencing the work as quarterly value releases facilitates tying the technology to real business value, with each quarter tied to a live decision on a priority asset. Deploying against visible assets and decisions makes the impact visible across the organization.
-
Data
An integrated decision engine requires a unified data layer that ingests internal historical trial data and external intelligence feeds. It is founded on a disease-specific data strategy that evaluates internal data availability and granularity, and external opportunities for complementarity. Using large language models (LLMs), automated pipelines extract structured data from clinical trial publications, registries, and real-world data sources and organize it into disease-specific schemas. This structured data layer provides baseline parameters for trial-level priors, competitive landscape benchmarks, and real-time event signals.
-
Technology
The decision system’s backbone should be modular so analytic components can be replaced or upgraded as methods improve. User interface is also key; outputs need to reach clinical, quantitative, operational, and finance users in a form they can act on. Increasingly, this means agentic interfaces that accept plain language, rather than point solution dashboards.
-
Talent
Reasoning with probabilistic, multiobjective models is a different skill than the deterministic analytics that many clinical or operational teams have used for decades. Making that transition requires codevelopment rather than delivery of a new product. Successful deployments involve dedicated cross-functional pods comprising data and product engineers alongside biostatisticians, clinical scientists, and portfolio strategy leads. This combination also ensures that the algorithms remain scientifically rigorous, technically scalable, and directly aligned with the key questions facing executive decision-makers.
-
Operating model
The most sophisticated simulation model creates no value if it runs in parallel with actual decision-making. Simulation outputs should be formally embedded in existing technical review structures, governance gates, milestone reviews, and portfolio review meetings, with explicit decision rights over who owns assumptions, validates outputs, and signs off. Many of the technical analyses6 and modeling rely heavily on assumptions and need to be accompanied by upfront diagnostics to confirm or adjust the given approach before investing too much time or too many resources. As teams upskill and adopt, the decision process needs rewiring so inputs are leveraged together rather than through sequential, manual handoffs. This decision system should ultimately provide the integrated recommendations for study, program, and portfolio decision-making, against which governance leaders should stipulate a clear rationale for deviating as a record of control.
-
Adoption and scaling
Lighthouse deployments and early quarterly value-release milestones on visible assets build momentum for the decision system. Incentives that recognize teams for testing alternatives help change behavior. Structured learning cycles after each milestone or decision help improve the underlying models as well as the human-led process. Human-in-the-loop validation is essential throughout to ensure that quantitative insights enhance rather than replace expert judgment.
The future of clinical development belongs to organizations that treat decision agility as a core strategic advantage. By linking study-level design choices directly to program sequencing and portfolio-level capital allocation, simulation-based decision systems replace fragmented, subjective estimates with real-time, scenario-driven, evidence-based governance. Unifying these three layers into a real-time, agentic decision system fundamentally changes how the R&D budget is deployed, ensuring it flows to the highest-impact programs designed for success.
