What happens when AI stops guessing and starts reasoning? Agentic AI is bringing scientific logic into the heart of drug discovery.
Artificial intelligence is now a routine presence in biomedical research, from literature mining to image analysis. Yet most systems remain limited to pattern recognition or text generation.
In drug discovery, where decisions depend on verifiable evidence, these tools often fall short. The core question is not how fast AI can produce an answer, but whether its reasoning can be examined, reproduced and trusted.
This is where agentic AI differs. Rather than generating isolated responses, it performs goal-driven reasoning: setting objectives, planning actions, gathering and checking evidence and verifying results before presenting them to a researcher. In life sciences R&D, this approach is beginning to influence how scientists interrogate data and test hypotheses.
In this article, we speak with Yiannis Kiachopoulos, co-founder and CEO of Causaly, who describes how this form of AI is being applied to the practical demands of biomedical research.
From automation to orchestration
Before founding Causaly in 2018, Kiachopoulos worked in consulting, helping enterprises use technology to accelerate knowledge work. His background in computer science and data systems design shaped a view that AI could support scientific reasoning by connecting different research processes, rather than simply automating individual tasks.
“We started Causaly in 2018 to help scientists interpret biomedical knowledge faster without losing the rigour of their work demands.”
In this model, multiple AI ‘agents’ operate together, each performing a distinct research function. One agent might collect and analyse data, another might validate literature evidence, and a third – acting as a Principal Investigator – reviews the results to ensure quality before returning any output.
“We built our technology with the needs of scientists in mind. Causaly Agentic Research runs on an agentic orchestration layer that coordinates a multitude of specialised agents – including a Planner, Executor, Principal Investigator and other agents that are experts in a particular workflow or tool – under the oversight of a human-in-the-loop.”
This approach moves beyond the idea of AI as a single tool and towards an environment where systems and researchers collaborate through structured, iterative reasoning.
What makes agentic AI different
Most people are now familiar with generative AI, which can draft, summarise or paraphrase large volumes of existing information. While useful, it is often unsuitable for scientific research, where the credibility of every statement depends on its underlying evidence. Agentic AI, by contrast, is designed to perform goal-driven reasoning. It does not simply produce an answer but explains how that answer was reached, allowing scientists to review the process.
Each agent performs a distinct research function, and the Principal Investigator verifies quality before any output is returned. If standards are not met, the process repeats.
“Each agent performs a distinct research function, and the Principal Investigator verifies quality before any output is returned. If standards are not met, the process repeats.”
In practice, this means that an agentic system behaves more like a research assistant who knows how to plan experiments, gather data and cross-check findings against prior studies. Every stage is documented and traceable, ensuring that the result can be reviewed and reproduced.
“That’s how we deliver what we call science-grade AI. It’s transparent, traceable and reproducible.”
Applying agentic AI to target identification
Target identification remains one of the most complex stages of drug discovery. Research teams must consolidate evidence from vast sources of literature, experimental data and competitor activity before deciding which biological targets are most promising. Agentic AI aims to make this process faster and more reliable.
“In target identification and prioritisation, teams traditionally spend weeks consolidating literature, experimental data and competitive intelligence. With Causaly these steps are orchestrated automatically.”
A large biopharma partner completed a target-prioritisation exercise in five days that would have required four weeks previously, while preserving full traceability.
Through an internal Scientific Information Retrieval System (SIRS), Causaly’s platform integrates public and proprietary data into a unified ‘Data Fabric’. This structure enables AI agents to reason across multiple evidence types, while automatically documenting every citation and assumption in their reports.
“A large biopharma partner completed a target-prioritisation exercise in five days that would have required four weeks previously, while preserving full traceability.”
The educational value of this approach lies in its transparency. Instead of delivering black-box results, the AI shows scientists how evidence was weighted, what limitations exist and which sources support each claim. This helps ensure that the reasoning process remains aligned with scientific standards of documentation and reproducibility.
Maintaining scientific rigour
In biomedical research, accuracy is not only a quality measure but a regulatory necessity. The use of AI within R&D must therefore meet the same standards of scrutiny that apply to human-generated analysis. For Kiachopoulos, this principle is fundamental.
Rigour is non-negotiable. Every statement Causaly makes is backed by citations and our deep research reports are transparent about any assumptions and limitations, enabling scientists to make informed and defensible decisions.
“Rigour is non-negotiable. Every statement Causaly makes is backed by citations and our deep research reports are transparent about any assumptions and limitations, enabling scientists to make informed and defensible decisions.”
Causaly’s framework enforces multiple layers of verification. Agents are restricted to approved data sources, a central audit layer enforces compliance and continuous benchmarking measures both accuracy and reproducibility.
“Our latest benchmarking shows that 99 percent of statements are backed by evidence. In a separate evaluation using the LitQA2 dataset, the system reached 88 percent precision on expert biomedical questions – slightly above the human baseline reported by FutureHouse (87.9 percent) and ahead of other evaluated AI models, while preserving full traceability of sources.”
However, the company also evaluates the quality of reasoning, not just numerical accuracy.
“We’ve developed a framework to measure foundational accuracy as well as qualitative depth of analysis and argument structure – and transparency of assumptions and limitations. We mirror scientific peer review criteria to evaluate AI agent output as we would that of a human scientist.”
This educational perspective reflects a broader trend: AI is no longer judged solely on performance metrics but on its ability to meet the epistemological standards of science itself.
Where agentic AI adds most value
Although agentic AI can, in theory, support any research process, its early successes have appeared in areas where the evidence base is especially large and fragmented.
“Particularly strong results appear in what we call high-impact discovery use cases: target identification, biomarker discovery and drug repurposing. These are areas where scientists handle enormous volumes of evidence and where every week saved can change program timelines.”
Particularly strong results appear in what we call high-impact discovery use cases: target identification, biomarker discovery and drug repurposing.
Such tasks require the synthesis of thousands of papers, datasets and reports – an undertaking that can overwhelm human teams. By orchestrating these steps automatically, AI can reduce cycle times while retaining traceability.
“Causaly contributes across nearly every stage of the R&D pipeline from early research through clinical development to post-market activities like pharmacovigilance and medical information management.”
These examples illustrate a broader educational point: the effectiveness of AI in science depends not only on algorithmic sophistication but on how well it fits within the scientific method. The ability to produce reproducible reasoning chains that can be reviewed by peers is what ultimately distinguishes agentic AI from generic automation.
The wider context
The rise of agentic AI reflects broader shifts in enterprise R&D. According to a recent PwC survey, 79 percent of senior executives report adopting agentic AI, with 88 percent planning to increase their budgets. This suggests that the technology is moving rapidly from experimental use to operational integration.
In life sciences, where datasets are growing exponentially, the attraction lies in AI’s capacity to act as connective tissue across teams and disciplines. Instead of each department using isolated tools, agentic systems can coordinate work from discovery to clinical development in a single environment.
“We’re seeing teams use Agentic Research to surface insights that link pathway biology to clinical outcomes, connect competitive-intelligence findings with portfolio strategy and ensure that every conclusion is backed by verifiable evidence.”
Such integration has implications not only for efficiency but also for the quality of reasoning that underpins regulatory and strategic decisions.
Towards continuous discovery
For Kiachopoulos, the coming years will see agentic AI evolve from supporting specific tasks to managing entire research lifecycles.
“Over the next few years, I expect the biggest gains to come from end-to-end orchestration of research workflows, where multiple specialised agents plan, execute and validate scientific tasks in coordination with human experts.”
He anticipates that scientists will spend less time harmonising data and more time interpreting results. In this model, discovery becomes a continuous process, constantly updated by feedback from models, experiments and simulations.
“Scientists will spend less time collecting and harmonising data and more time interpreting results and testing new hypotheses, which will accelerate the rate of validated discoveries across therapeutic areas.”
Further integration of predictive modelling and experimental data could create earlier feedback loops, improving decision-making at every stage.
“Ultimately, the opportunity isn’t just faster research. It’s a new operating system for scientific reasoning that increases reproducibility, reduces development risk and gets effective therapies to patients sooner.”
Educating the next generation of researchers
For scientists and research leaders, understanding agentic AI is increasingly part of professional literacy. The technology challenges established assumptions about how evidence is generated, validated and communicated. Learning how to collaborate effectively with such systems – interpreting their reasoning as well as their results – will become a key skill for modern biomedical science.
As Kiachopoulos observes, the transformation is still in its early stages.
“We’re only at the beginning of what agentic AI can do for life sciences.”
Meet the expert
Yiannis Kiachopoulos, co-founder and CEO of Causaly
Yiannis Kiachopoulos is an 18-year veteran of driving strategy, transformation and innovation at global enterprises, with a background in computer science and a vision to radically transform life science R&D with artificial intelligence. For several years, he was a consultant at Accenture, advising some of the world’s largest companies, before co-founding Causaly in 2018.
He holds a bachelor’s in Japanese language studies, a master’s in computer science and an MBA from Hong Kong University of Science and Technology.