ZS’s Daniel Blessing and Satyam Shubham contributed to this article.

Pharma pipelines are under increasing pressure. The industry faces a roughly $350 billion patent cliff by 2030. Clinical organizations have to increase success rates, bring assets to market faster and lower costs—but the key metrics are going in the wrong direction. Success rates remain low, especially in phase 2, and costs per asset have increased as timelines to market have lengthened. Clinical organizations understand they have to fundamentally change.

Organizations are also recognizing incremental improvements like faster drafting tools, smarter dashboards and better analytics are not enough. What is required is a fundamental reimagining of how clinical development organizations operate. That is precisely why the industry finds itself in the middle of a wave of AI-driven transformation, which will change processes and workflows that have remained static for many years.

AI in clinical trials: Reflecting on the past and looking ahead

In recent years, pharma companies have used AI co-pilots for authoring, feasibility, identifying risks and data automation. Most efforts have remained at the proof-of-concept stage, meaning they’re promising in controlled settings, but are rarely scaled into production.

As we look ahead, orchestration is critical. Sponsors will need to think of clinical development as a system of decisions—from protocol design to trial execution to submission—where AI continuously senses, predicts and guides action. These six investments begin to enable that system:

1. Scale AI-powered clinical document authoring across the entire portfolio
2. Reimagine protocol design with AI-driven, persona-based simulation
3. Build a predictive trial and portfolio performance capability
4. Deploy clinical next best action at scale across every role in the field
5. Accelerate clinical data submission through AI-guided pipelines
6. Prioritize in silico modeling as a long-term strategic capability

Investment 1: Scale AI-powered clinical document authoring across the entire portfolio

Clinical documentation is one of the most time-intensive bottlenecks in drug development. A single trial can require more than 90 documents, with each taking days or months to reach a first draft. This translates into hundreds of days of authoring time annually.

The opportunity for improvement is significant: Early deployments of integrated authoring platforms show that 75%-90% of first drafts can be considered ready for review. This matches or exceeds human-written quality based on key accuracy and consistency measures.

And the value compounds over time. Each approved document strengthens the system, with the potential to drive 30%-60% faster authoring cycles and cut review timelines for complex documents from eight-to-14 weeks to roughly two-to-six weeks. The result is not just faster writing, but lower review burdens, accelerated trial timelines and more time focused on scientific rigor and patient impact.

What we’ve seen recently

By 2025, most large sponsors demonstrated that generative AI could substantially accelerate the drafting of individual clinical documents, including protocols, clinical study reports (CSRs), investigator’s brochures (IBs) and amendments.

But this approach has limitations:

• Most implementations remained document centric, with siloed authoring workflows and limited reuse
• Sponsor conventions were inconsistently applied and regulatory precedent often lived in disconnected repositories
• Changes made in one document rarely propagated cleanly to others

AI has improved speed in clinical document authoring, but not coherence.

What needs to happen now

Document authoring needs to evolve from isolated generation toward integrated authoring platforms. Clinical organizations must stand up AI authoring solutions at scale, spanning the entire portfolio and all document types, not just those that are high-priority or high-volume. The focus should shift from producing text to harmonizing clinical content across programs and portfolios.

This transition is now feasible because teams are getting better at turning unstructured content into standardized, machine-readable study definitions. By collaborating with generative and extractive AI—which understand context, dependency and lineage—these representations make cross-document consistency achievable at scale. Domain-specific ontologies, reusable templates and built-in compliance controls should become core capabilities, not incremental improvements.

Investment 2: Reimagine protocol design with AI-driven, persona-based simulation

Protocol design sits at the core of clinical development. But there are issues: For years it’s been slow, sequential and document driven. The cost is steep: Protocol amendments alone run roughly $500,000-$800,000 each and take an average of 6.5 months to complete. Siloed inputs from clinical, statistical and regulatory teams take months to translate into protocols, limiting early insight into enrollment, patient burden and execution risk.

But change is within reach. AI-driven systems are reducing protocol design cycles by 25%-40%, resulting in fewer late-stage changes and noticeably increasing the likelihood of technical success as trials stay more closely aligned with their original plans. As validated protocol components are reused across programs, each trial becomes faster, more consistent and more informed than the last, building decision confidence and momentum across the portfolio.

What we’ve seen recently

Many sponsors have adopted workflow‑driven platforms for protocol design, particularly for endpoints, eligibility criteria and schedules of assessments. These platforms required years of investment and extensive manual curation to create usable protocol data banks.

Approaches varied considerably across the industry. Some sponsors focused on using real-world data as a primary input to inform inclusion and exclusion criteria, while others leaned on similar historical trials as a reference base for design decisions. Still others concentrated primarily on workflow optimization without deeply integrating external data sources.

Despite their promise, adoption has remained uneven:

• Workflows have relied more on precedents than predictions to guide study design
• Curation of real-world data and historical clinical trials data has not scaled
• Cross‑functional alignment has added months to final protocolization

What needs to happen now

Protocol design should shift from precedent-driven workflows to AI systems that actively simulate what a trial will encounter before it begins. The core mechanism is straightforward: AI agents representing clinical, statistical, regulatory and patient perspectives evaluate the proposed design simultaneously, with each surfacing their own opinion and critique on endpoints, eligibility criteria and execution feasibility.

Each agent also draws on the data sources most relevant to its role:

• The statistical agent stress-tests power assumptions against real-world data.
• The patient agent flags recruitment risk and burden using advocacy and outcomes data.
• The clinical agent identifies precedent from internal and public protocol repositories.
• The regulatory agent flags exposure and compliance risk against current guidance.

In this approach, tradeoffs are surfaced transparently so teams can resolve disagreements in design rather than in amendment cycles. And design and authoring become a single workflow, with each decision flowing immediately into the protocol document so clinicians and writers are always working from the same draft.

Further value will come from simulating patient-level data. Real-world data will enable AI to model survival curves, adverse events and population diversity during the design phase, helping define the right trial population and improving the chances of trial success.

Investment 3: Build a predictive trial and portfolio performance capability

Executives today rarely have a clear, integrated view into how much capital is at risk or the probability of delays across the portfolio. The consequences of this limited visibility are significant:

• Fewer than 20% of trials meet their original enrollment goals
• 70% of trials face startup delays from contracting and regulatory bottlenecks
• More than 30% of patients drop out, with each replacement costing three times more to recruit

These are not isolated execution problems. They are systemic signals that go undetected until it’s too late to act. The core issue is that by the time problems surface through existing dashboards and reports, the window for low-cost intervention has already closed.

Predictive visibility at the portfolio level changes that equation entirely, enabling leadership to identify risk early, act preemptively and protect both timelines and capital before delays compound.

What we’ve seen recently

Sponsors have invested heavily in clinical data systems, such as clinical trial management systems, electronic data capture, interactive response technology and electronic trial master file, and many have layered AI on top to generate operational insights. But these insights have remained fragmented by system and by role. Executives get high-level summaries while study teams get operational data, and rarely do the two connect into a coherent picture of where the portfolio is heading. Predictive capabilities exist in pockets but haven’t been stitched together into a unified forward-looking view.

What needs to happen now

Enterprise data harmonization should enable a unified trial and portfolio view spanning clinical, quality, safety and operational performance, creating a single source of truth for decision making. More importantly, this view helps leadership shift from a retrospective understanding of performance to a preemptive perspective on emerging risks and issues.

AI-driven signals can highlight patterns such as early signs of site underperformance—for example, declining screening rates or rising screen failures—well before they translate into enrollment delays at the study level. This allows executives to anticipate downstream impact across timelines and portfolio milestones, enabling earlier planning, prioritization and risk mitigation. Leaders don’t have to wait for an issue to materialize to act.

Investment 4: Deploy clinical next best action at scale across every role in the field

Better trial and portfolio visibility enables organizations to deliver enhanced intelligence to trial execution teams. For these teams, knowing a site is at risk is only half the equation—the other half is ensuring the right person gets the right recommendation at the right moment.

Today that last mile remains broken, with study leads, CRAs and country managers spending the bulk of their time on manual coordination and reactive problem solving. Thankfully, next best clinical action can:

• Reduce manual coordination effort by 20%-30%
• Accelerate issue resolution cycles by 30%-50%
• Materially reduce downstream quality events

With organizations losing roughly $1 million for every month of delay, even modest improvements in how quickly field teams act on emerging risks will translate into protected timelines and capital.

What we’ve seen recently

Clinical roles in trial execution have remained largely static for nearly three decades. The core responsibilities of study leads, CRAs and country managers have changed little, even as trial complexity has grown dramatically. While sponsors have run AI pilots to bring more intelligence to these roles, widespread adoption has remained elusive.

These AI pilots tended to be small in scale, confined to a single function or role, and disconnected from the broader operational workflow. Recommendations were generic rather than role-specific, and without cross-functional orchestration, even good signals rarely translated into coordinated action.

What needs to happen now

With the unified trial and portfolio view from Investment 3 as the backbone, the focus shifts to the last mile: Delivering the right recommendation to the right person at the right moment.

Persona-aware intelligence layers deliver context-specific, impact-ranked actions to study leads, CRAs and country managers, moving beyond generic alerts to precise, decision-ready guidance. Site-facing platforms such as site CRM will evolve into intelligent execution hubs, dynamically orchestrating site-level interventions through proactive nudges, tailored support and automated follow-ups.

In parallel, action orchestration frameworks will codify mitigation patterns, simulate downstream impact on timelines and costs and dynamically prioritize the next best action. Over time, closed-loop learning systems will continuously refine these actions based on outcomes, enabling clinical ops to shift from reactive oversight to autonomous and self-optimizing trial execution at scale.

Investment 5: Accelerate clinical data submission through AI-guided pipelines

Timely, high-quality clinical data submission is one of the most persistent bottlenecks in clinical development. The journey from last patient, last visit to final regulatory submission can take 12 months or more. This includes data cleaning, study data tabulation model (SDTM) and analysis data model (ADaM) conversion, statistical programming and clinical study report (CSR) authoring, with database lock alone taking four-to-eight weeks.

Every week lost in this pipeline delays regulatory review, pushes back potential market entry and defers the point at which a therapy reaches patients. Shortening the timeline even modestly can greatly boost commercial success and patient outcomes for assets that have billion-dollar sales potential.

What we’ve seen recently

Data submission today remains heavily manual and fragile. Study specifications are written in narrative text that humans can interpret but machines cannot reliably process. A limited number of sponsors have initiated agentic AI pilots—but despite promising results in controlled environments, these sponsors often found the pilots had limited scalability across programs.

Automated SAS code generation has remained uneven, with models hallucinating logic or overlooking nuanced statistical constraints. As a result, programmers have been acting as a critical bottleneck for quality control, traceability and data lineage. This has added weeks of incremental cycle time per study.

What needs to happen now

Clinical organizations should move toward standardized, machine-readable specifications and structured metadata to remove ambiguity at the point of data definition. Once this foundation is in place, AI models trained on sponsor-specific data can automate large parts of code generation, while built-in validation workflows can continuously check outputs against expected results to ensure accuracy and traceability. Together, these capabilities streamline data preparation and analysis, accelerating submission timelines by 30%-50% and improving traceability by design.

A key enabler is the growing adoption of R and Python alongside SAS, which opens up significantly more possibilities for deploying AI agents across the programming workflow. Companies need to accelerate their clinical data repository (CDR) and statistical computing environment (SCE) implementations, while also layering AI agents to drive the journey from raw data through SDTM and ADaM to final tables, listings and figures (TLF).

This transforms a historically manual, sequential process into a largely automated pipeline, freeing programmers to focus on quality and interpretation.

Investment 6: Prioritize in silico modeling as a long-term strategic capability

In silico simulation is positioned to become a major driver of change in clinical development, as it shifts organizations from reactive execution to a more predictive, model-driven strategy.

Simulations in virtual environments will provide early visibility into outcomes and probability of success before any patient is enrolled, enabling the rapid evaluation of protocol variations across design, population and site strategy. While current applications remain limited and experimental, early evidence signals their strong potential, with in silico poised to fundamentally transform clinical organizations in the next three-to-four years.

What happened recently
In silico remains a frontier area for most organizations, with efforts largely concentrated in experimental pilots. Momentum is accelerating as traditional quantitative modeling approaches across dosing, disease progression and trial simulation are increasingly combined with AI-driven prediction models to stress test development assumptions and trial strategies.

Early experimental models have demonstrated the ability to reproduce historical trials and generate forward-looking predictions for studies in flight. A fast-growing vendor/startup ecosystem is productizing these capabilities, lowering the barrier to adoption without building everything in-house.

Regulatory clarity is also improving: ICH M15 (2026) sets expectations for planning, evaluating and documenting model-informed evidence across development.

Still, the impact of in silico approaches and digital twins remains limited by data scarcity, narrow feature sets and siloed deployment outside of core trial workflows.

What needs to happen next

In silico modeling is continuing to mature but remains largely experimental for most sponsors. It’s not yet at the core of how trials are designed and run. In just three years, however, leading clinical organizations will have in silico models front and center across their development programs.

To get there, sponsors need to take three key steps:

• Step 1: Continue refining modeling capabilities, building both in-house capabilities and using the specific strengths of the modeling partner ecosystem.
• Step 2: Keep building AI-ready data foundations that make models credible and useful, particularly real-world data and EMR assets, which remain the critical input layer for simulation quality.
• Step 3: Define clear organizational ownership of in silico-modelling and establish the processes for when and how models are used to inform asset and portfolio decisions. This critical step is sometimes overlooked.

Scaling AI in clinical development to reduce costs and accelerate trials

Nearly every clinical development program has demonstrated that AI can drive meaningful change. Now they must scale that impact by making the six investments that will transform workflows at scale and embed AI into new ways of working.

These investments don’t exist independently of each other. Protocol design simulation feeds better-defined trial populations into execution. Portfolio visibility creates the signal layer that drives next best action in the field. Authoring platforms compound value with every document approved. Data submission pipelines accelerate the journey from last patient to regulatory filing. And in silico modeling, which is still maturing today, will ultimately enable simulation across the full arc from protocol design to trial execution, changing how leading organizations derisk assets by the end of the decade.

Turning AI into decision systems for faster clinical development

Taken together, this shift is less about deploying isolated AI capabilities and more about redesigning how decisions and execution work together across clinical development. When high-impact decisions operate as connected, continuously learning systems, intelligence can scale across the portfolio with greater speed, control and confidence.

Sponsors that connect these investments and build the data foundations and ownership needed to make them work together could reduce development costs by up to 60% and shorten trial cycle times by up to 40%.

The question is no longer whether AI will transform clinical development. It already is. The only question is which organizations will lead that transformation and which will follow.