ICON is developing AI agents to reduce administrative work and give clinical research associates more time for site relationships.
Clinical trials are among the pharma domains seeing the most rapid embrace of AI agents, thanks to their potential extensibility and ability to automate routine tasks. Cost and complexity have a lot to do with that interest. An average phase 3 trial collects close to 6 million data points, according to a Tufts CSDD and TransCelerate study published in December 2025. A separate 2024 Tufts analysis put daily direct costs in the ballpark of $55,000, expressed in 2023 dollars.
Given such high stakes, AI agents present a “huge opportunity,” said Tony Clarke, senior vice president of enterprise AI at ICON plc, the global contract research organization that announced collaborations with Microsoft in June and Anthropic in July.
With those partners in tow, ICON continues to develop and scale agents that can perform individual tasks or work together across a workflow. The architecture spans automation, orchestration and multi-agent coordination. “But what’s equally important is the data layer, and the semantics around definitions and understanding structures,” Clarke said in an interview.
The broader aim is to identify where process changes, whether driven by AI or another technology, could streamline a task. “When we look at an area where we feel we could provide AI capabilities, it’s very much a case of forensically mining out the current state and really understanding the problem, and taking that data and AI lens to it,” Clarke said. “Sometimes it’s not necessarily even AI. Just pure technology or integration can drive some acceleration as well, but the goal is really reimagining it, questioning how the processes run, and understanding the data underneath it.”
A foundation in Orbis
ICON’s Orbis platform provides the foundation for developing and running AI agents, with access controls, security and audit records. Agents supporting statistical code generation and site contracting are already in production, delivering shorter cycle times in each area, Clarke said.
Tony Clarke
Such efficiencies have drawn executives across industries to AI agents’ potential to scale productivity without proportional headcount increases, albeit, at times, with less reliability than human experts. In coding and finance, for instance, AI agents echo the early cloud in that they are cost-effective for bounded tasks, but prone to runaway costs and scope drift when work is open-ended. Still, it is difficult to make firm pronouncements with moving targets. New frontier models continue to ship at a rapid cadence, with steady improvements on tasks with clearly verifiable results, such as coding and mathematics.
Given those gains, a core question remains: How much will those increasing capabilities change what drugmakers pay? Asked whether agents could shift ICON’s billing model, Clarke said he foresaw limited change, while noting that sponsors are “interested in AI and the efficiencies that it brings.” “Our clients are sophisticated buyers,” he added, and price was a priority long before genAI became popular.
On the longer-term cost calculus in clinical trials, Clarke emphasized the continuing role of human decision-makers. “We are a regulated industry, so there’s always going to be a human that is making a decision,” he said. “I don’t see any time soon that it’s going to be an AI agent making a decision in lieu of a human, and the EMA and FDA principles around this are, number one, human centricity.” Those requirements, he said, will keep people at the core of ICON’s AI capabilities.
Human review remains central to ICON’s clinical-trial AI workflows, which combine structured data, task-specific agents and tools designed to produce repeatable results.
Use the protocol across multiple tasks
Clarke breaks clinical trials into four operational domains: initialization, where ICON collaborates with a client on protocol design, system setup, and initial documentation; study setup, which focuses on site activation; execution, covering field monitoring and filings into the electronic Trial Master File (eTMF); and closeout, which encompasses biostatistics and programming. “We see, quite frankly, [agentic] opportunities across all of those areas,” he said.
To connect these stages, ICON recently developed a digitization capability built on the Clinical Data Interchange Standards Consortium’s (CDISC) Unified Study Definitions Model. Clarke said this foundation accelerates system setup and allows AI, with human oversight, to draft downstream documentation such as statistical analysis plans. “We’re trying to use the right tool for the right job, and to ensure that interconnectivity accelerates digital data flow, so our documents and materials can be used throughout the lifecycle, and we don’t have teams doing duplicate work,” Clarke said. “The same protocol that’s digitized to set up our systems and help author some of the documents is also the same digitized document that our CRAs will query if they have a question about the protocol.”
On the 2026 innovators list
Anthropic and Microsoft are among R&D World’s 100 most innovative companies of 2026. The list examines what each company built, who is using it and the results it has reported. R&D World will publish the full comparison in the 2026 Global R&D Funding Forecast.
ICON is also reimagining how its Clinical Research Associates (CRAs) work. The role, the field staff who conduct monitoring visits, travel to study sites, and ensure trials run according to protocol, is “a very complex role,” Clarke said, due to the orchestration required across multiple platforms.
“We’re leveraging Anthropic models in combination with our Orbis platform to really reshape how those roles are performed, and allow our staff to spend less time on administrative activities and more time on site relationships, where that human element is really important,” he said.
The CRA capabilities are under active development and will run on Orbis, Clarke said. ICON also plans to use the Orbis platform to detect enrollment risks, trends, and operational signals across active studies in real time.
Making completion more predictable
Across the broader landscape, agentic adoption in clinical trials remains at a relatively early stage, but Orbis is already offering benefits, Clarke says. “We have a number of agents in production at the moment doing various activities,” he said, pointing to automated SDTM-to-SAS code generation that has driven “cycle reduction in the region of 30%.”
While speed is a major priority for clinical trials, “it’s really important that what we do is predictable from a sponsor perspective, and that the quality is there as well,” he added. Delivering that reliability means targeting “a high degree of accuracy and precision” and “leveraging deterministic approaches,” ensuring models do not infer their way through regulated steps.
Clarke noted that while frontier models continue to improve in accuracy and produce fewer hallucinations over time, foundation models remain inherently probabilistic. Optimizing their performance in life sciences requires clear guardrails and deterministic tools that guarantee repeatable outcomes.
“It’s really important to us that where possible we’re taking a deterministic approach, so that the model isn’t inferring or making an educated guess,” Clarke said. “Also, that responses are grounded in semantic and ontology structures, and that humans remain in the loop.” To minimize risk, ICON restricts individual agents to bounded steps rather than open-ended workflows, while building in explicit stop conditions to prevent agents from looping.
On Artificial Analysis’ AA-Omniscience benchmark, which tests factual knowledge across 6,000 questions, Anthropic’s Claude Fable 5.1 achieved 67% accuracy in its highest-scoring configuration. Furthermore, a Princeton University preprint evaluating 14 agentic models found that rising task accuracy across model releases produced modest gains in overall reliability; agents capable of completing a task often struggled to replicate that success consistently across multiple attempts.
To counter this variability in probabilistic genAI models, ICON builds deterministic tools and APIs that agents call upon to make their actions consistent, explainable and auditable. That architectural separation allows an agent to reason about a task while relying on hardcoded, validated steps to execute it. ICON also invests heavily in foundational data engineering so that routine steps never stumble. “If we’re all enamored by the complex piece of a task being done correctly, but the basic pieces aren’t done well, then that’s not a solution that’s going to get deployed or adopted,” Clarke said.
Beyond technical safeguards, user adoption introduces a different set of hurdles. “What we’re finding is that equally as important in terms of adoption is user experience, and the UX components of a solution have as much of a bearing on adoption as the accuracy does,” Clarke said.
To smooth the transition, ICON pairs rollouts with change-management initiatives, structured training and an internal network of AI champions who cultivate organizational buy-in and help staff adapt to redesigned workflows. The core focus is on operational efficiency while meeting regulatory guidelines, Clarke stressed. “What we ultimately want to do is accelerate time to market for our clients, and do that in a compliant way.”
