As the global life sciences industry embraces the latest advances in science, and adapts to evolving regulatory demands, as well as pressures on price and margins, artificial intelligence (AI) has risen in perceived value — both as a practical solution to processing soaring workloads, and as a means of transforming the work that companies do.
Today, the technology is being deployed with increasing momentum, as well as rising ambition, as organisations come to understand its full potential. Generative AI (GenAI), which came to mainstream attention with the launch of ChatGPT in 2022, has proved a disruptive catalyst for whole industries, offering the ability to take what has gone before, amalgamate knowledge, distil key facts including fresh insights, and present these in new ways.
In life sciences R&D, across functions such as regulatory affairs and drug safety/ pharmacovigilance, AI has paved the way for intelligent automation of highly labor-intensive routine processes. The technology is being actively used in marketing authorisation, application preparation, product change control/regulatory impact assessment management, adverse event case processing, and safety reporting. Through these proliferating applications, AI has contributed to tangible improvements in process cost-efficiency.
Gains have included accelerated task execution, honed accuracy and consistency in process output, and large-scale resource optimisation, as teams of professionals have recovered capacity for more strategic and challenging tasks. In deploying AI across these discrete use cases, pharma organisations and their functional leads have learned a lot about the technology’s potential and how best to leverage it to get good and trusted results. All of this has paved the way for an even bigger wave of AI advancement — in the form of agentic AI.
Defining agentic AI
Agentic artificial intelligence (agentic AI) involves the autonomous coordination of goal-driven AI “agents”. OpenAI, ChatGPT’s creator, talks about agentic AI “bridging research and action”; about GenAI platforms being able to “proactively choose from a toolbox of agentic skills to complete tasks for [the user]”.
The term “tasks” understates the potential, however; agentic AI’s undertakings could be large and complex. ChatGPT agent, for instance, integrates earlier agentic capabilities such as Operator and Deep Research into a unified system that is able to “think and act” with its own “virtual computer”. The vision is of a system that sets its own goals, decomposes tasks, acts across tools, and adapts without human oversight.
The arrival of agentic AI is the most significant leap in AI technology’s development since the launch of ChatGPT three years ago. This is because of the inherent potential to redefine the way organisations operate and the value they deliver — based on agentic AI’s ability to apply its own reasoning. It marks a step change from previous incarnations of AI which involved the automation of predefined processes according to given rules.
Agentic AI systems operate with significantly enhanced independence in both their actions and methodologies. When provided with a target objective, specialised agents apply their distinct capabilities, accumulated knowledge, and analytical processes to execute their designated functions optimally. An orchestration layer coordinates and oversees these activities. Beyond simply optimising outcome delivery, this orchestration component synthesises collective intelligence to identify novel opportunities for value generation. Put differently, this emerging AI paradigm transcends mere efficiency gains; it fundamentally enables innovative value-creation models where independent, collaborative agent reasoning generates fresh insights and informs decision-making processes.
What could this mean for pharma?
The ability to reason, anticipate, generate insight and knowledge and make better decisions is ideal for pharma, as an industry that is data-rich, process-heavy, and outcome-critical. Agentic AI is not just about doing the same things more efficiently and more accurately. It can help to challenge current processes and determine what else might be possible; what other opportunities might be leveraged.
So where is agentic AI now in pharma? Pre-agentic AI is already enabling new cost-efficiency in R&D functions such as regulatory affairs and drug safety/pharmacovigilance. To date, this has tended to be in discrete areas such as marketing authorisation application preparation, product change control/regulatory impact assessment management, adverse event case processing, and safety reporting.
Agentic AI’s vision is more ambitious, potentially enabling step changes in the role played and value contributed by safety, regulatory and adjacent teams.
Rendering pharmacovigilance more responsive
AI has already demonstrated its capacity to improve both the speed and precision of adverse event data classification, while enabling additional validation checks and accelerating subsequent workflows. Pairing automated MedDRA coding with intelligent signal detection could remove manual processing delays. For instance, when AI agents identify an atypical pattern among coded terms, they could automatically generate a preliminary signal warning, create a draft signal report complete with relevant case selections, chronological analysis, and supporting documentation, while assigning an appropriate urgency level for human safety specialists to review.
This approach would reduce the time required to identify credible signals while allowing experts to concentrate their efforts on complex, uncertain, or unprecedented cases and developing investigation protocols. Concurrently, the system could automatically direct high-risk patterns to epidemiology and medical affairs teams and propose immediate protective measures—such as targeted stakeholder notifications, product holds, or intensified surveillance - to support human decision-making processes.
Tightening regulatory conformance
Within the regulatory landscape, agentic AI could fundamentally transform how organisations manage worldwide product compliance obligations. The technology now exists to enable autonomous systems that understand regulatory frameworks to assemble dossiers and coordinate submissions automatically.
AI agent networks can be configured to continuously process incoming materials—including clinical data sets, study documentation, chemistry manufacturing and controls (CMC) files, electronic trial master file (eTMF) references, and historical submission materials. These agentic platforms are also capable of conducting automated compliance assessments against jurisdiction-specific requirements, generating region-tailored Common Technical Document (CTD) and electronic Common Technical Document (eCTD) sections with full source attribution and document traceability, while managing the technical formatting requirements such as file nomenclature and directory organisation. Including human expertise remains important, but progression towards greater AI autonomy is about routing suggestions for human review when potentially ambiguous scenarios arise and an expert check is needed.
By extension, such an agentic platform could produce a concise justification for its decisions alongside a curated list of points requiring human verification, while executing validation protocols to ensure file integrity, verify cross-references, and check region-specific appendices. These outputs would enable the system to automatically direct materials to appropriate specialists — including CMC, clinical, and labelling experts—accompanied by proposed modifications and risk rankings, ultimately delivering reviewers a dossier requiring minimal additional preparation before submission.
From a strategic perspective, reducing the time taken to execute these steps offers the potential to expedite portfolio decisions and accelerate patient access to therapies, while enabling sponsors to refine study protocols with greater agility. The compliance assessment insights generated by these agents, meanwhile, could be channeled back to clinical operations and protocol development teams, facilitating trial designs that pre-emptively address regulatory requirements and minimise the need for subsequent clarifications.
How to harness the technology
Developing a cohesive plan
Organisations contemplating agentic AI deployment must adopt a comprehensive, long-term perspective on AI implementation rather than viewing it as a series of isolated initiatives. This means recognising how the technology’s value is compounded when applied across multiple interconnected applications, which necessitates a more integrated and methodical deployment framework.
The fundamental premise of agentic AI systems lies in their ability to achieve designated objectives through optimal pathways of their own determination— leveraging and building upon their entire accessible knowledge base. These systems offer autonomous analytical and decision-making capabilities, along with ongoing learning and refinement, in pursuit of specified targets. The aggregate value proposition intensifies as individual agents progressively improve their performance through self-directed inference and emerging knowledge.
Because of this, organisations will require a cohesive governance framework and suitable control mechanisms to extract maximum reliable value as traditional workflow boundaries dissolve and agents engage in innovative crossfunctional collaboration to optimise outcomes. But what form should such measures take?
Building confidence quickly
Agentic AI’s potential reinforces the importance of companies’ underlying data assets — their quality and completeness, as well as how well that data can be combined and leveraged in different contexts. Importantly, this includes consideration of how organisations’ own data agents safely connect into a wider multi-agent fabric outside their walls - both data-to-data and agent-to-agent.
It also challenges “trust” around AI reasoning. Where AI systems are being afforded new autonomy across extended workflows, the risks go well beyond incorrect outputs. They now include potential for unintended data movement, loss of operational control, misaligned decision-making and blurred lines of accountability.
Having guardrails to mitigate unintended behavior is critical. This is sometimes referred to as “bounded autonomy.”At the same time, companies need to avoid being too prescriptive and limiting in their attempts to pin down good governance — because of the need to allow for future scenarios.
Although multi-agent frameworks are emerging to support the creation of AI systems, these do not inherently manage trust, context-sensitive decision making or risk-aware governance. Such provisions need to be both designed-in from the start, and able to adapt to evolving needs. That is, governance needs to be viewed as a facilitator as well as a mitigator of risk.
Allowing for innovation within risk management approaches
Finally, if accounting for all eventualities and risks incurs too much complexity, companies risk undermining any economic benefits from the technology. Taking a “principles-based” approach, rather than one that is hard-wired around specifics, supports process stakeholders in defining scenarios and goals that agentic AI can help solve.
This principles-based approach — as advocated by the Council for International Organisations of Medical Sciences (CIOMS) Working Group XIV on AI in Pharmacovigilance1 — aims to create a common foundation for regulators, industry and technology providers that can keep pace with the unprecedented rate of technological advancement.
In practice, organisations may complement such principles with their own systemic-thinking or service-design methods — for example, developing journey maps to plot how agentic workflows trigger, interact and evolve. These tools help translate high-level principles into operational governance models, including the degrees of autonomy afforded to individual agents. The idea is that trust emerges from the infrastructure itself, rather than relying on static checklists that date quickly.
Once fully understood these considerations could be built into adaptable provisions for human involvement, allowing companies to move at their own pace towards trusted use of agentic AI — and, by extension, the maximum associated benefits.
Footnotes:
1 Artificial intelligence in pharmacovigilance, CIOMS Working Group report, Draft, 1 May 2025: https://cioms. ch/wp-content/uploads/2022/05/CIOMS-WG-XIV_Draftreport-for-Public-Consultation_1May2025.pdf
