BLOG
In May 2026, ONTOFORCE convened a panel of industry leaders from organizations like Roche, Sanofi, and others, for a discussion focused on a practical question: what is actually standing between today's AI capabilities and the workflows life sciences organizations need them to power?
The promise of agentic AI in life sciences is well established: autonomous systems that reason across compound libraries, synthesize evidence at scale, and surface insights that would take a researcher months to reach manually. The technology exists. The investments are being made. And yet, across the industry, the gap between controlled-environment performance and reliable production deployment remains significant.
In May 2026, ONTOFORCE convened a panel of industry leaders from organizations like Roche, Sanofi, and others, for a discussion focused on a practical question: what is actually standing between today's AI capabilities and the workflows life sciences organizations need them to power?
The answer, consistently, pointed not to the models but to the data infrastructure beneath them.
Key takeaways
Life sciences organizations are realizing genuine value from AI, but predominantly within a narrow class of use cases. Literature synthesis, document summarization, and regulatory reporting are all areas where significant efficiency gains have been reported. When a task that once took a scientist several hours of literature review drops to under an hour, the gain is not marginal. Compounded across hundreds of researchers, it represents meaningful recovered capacity.
The limitation is that these wins are largely vertical. They occur within single data domains, operated by specialists who already possess the contextual knowledge the system depends on. When organizations attempt to extend AI into more genuinely agentic territory, where a system must reason horizontally across multiple data sources and domains, the same foundations that support narrow use cases become obstacles.
"If you want to go truly agentic, it is about horizontal connectivity. Combining data from different sources together. Inside organizations, we see the data layer is not ready. Every silo still has their semantic model, their ontology, but there is no horizontal connectivity layer available that connects this as tissue. Without that layer, you are not building an agentic workflow. You are just building something that does the existing work faster."
— Lieven Poelman, Panelist, VP R&D, ONTOFORCE
The root cause is structural. Individual data silos within most large pharma organizations have their own semantic models and ontologies, built to serve the systems they were designed for. What is absent is a horizontal integration layer: a common semantic infrastructure that harmonizes terminology, resolves entity relationships across domains, and enables an agent to move coherently between sources without losing provenance or context.
"AI-ready data" has become a frequently used term, but the concept is often reduced to data quality. A more useful definition is this: AI-ready data is data that an agent can access and reason over with little or no human intervention. Quality is a prerequisite for that, but it is not the distinguishing factor. The more consequential and underestimated challenge is context.
Enterprise data assets, particularly in large pharma organizations, are largely legible to the humans and systems that were built alongside them. Field names, identifiers, and taxonomies carry implicit meaning that practitioners understand without documentation. AI agents do not inherit that understanding. Without sufficient context, they either produce inaccurate outputs, fail to retrieve relevant information, or consume disproportionate compute attempting to compensate.
The FAIR Data principles, which have guided data governance investment across the life sciences for years, offer a useful lens here. Progress on findability and accessibility has been real. Interoperability and reusability, the properties most relevant to agentic systems, have lagged, in part because achieving them at enterprise scale has historically required sustained human coordination that does not scale.
What is emerging in response is a more targeted approach. Rather than attempting to build comprehensive, organization-wide ontologies, leading organizations are constructing narrow, fit-for-purpose semantic layers scoped to specific agent workflows. These structures are rich enough to be machine-readable and contextually sufficient for the task at hand, without the overhead of maintaining a monolithic knowledge model.
Part of the rationale is practical as much as architectural: a smaller, scoped semantic layer means less context for an agent to plough through at inference time, and that translates directly into lower token costs and faster, more reliable reasoning. A monolithic ontology might be more "complete," but most of it is irrelevant to any single task — and irrelevant context isn't free; it's something the agent has to read, weigh, and risk being distracted by. The goal, as one panelist framed it, is to build small, well-reasoned semantic structures that can be composed incrementally into something more capable over time.
A related discipline that featured prominently in the discussion is context engineering: the practice of deliberately managing what information is made available to a model during inference, and in what form. The challenge it addresses is well understood by practitioners. A large language model without grounding will produce confident outputs regardless of whether its knowledge is sufficient for the query at hand. In a domain like pharmaceutical R&D, where a single term can carry different meanings across chemistry, manufacturing, and preclinical functions, uncontrolled inference is a meaningful risk.
"Context becomes the new compute. If you are able to build a semantic infrastructure on top of your data layer, that is what allows you to have efficient, optimized compute to get the right decisions through agentic operations. Otherwise you are going to have a very expensive infrastructure, going through language models without being able to reason accurately in a cost-balanced manner."
— Sabya Dasgupta, VP Global Head of Data and AI Platforms, Sanofi
Context engineering treats this as an infrastructure problem rather than a prompt design problem. It involves building the semantic layer, managing ontology alignment, resolving identifiers, and maintaining data provenance so that agents have a reliable, bounded knowledge base to reason over. It is the approach panelists consistently pointed to: investing in the underlying data infrastructure as a prerequisite rather than an afterthought.
The panel also examined what a more autonomous AI system in scientific research would actually require. The question of the "virtual researcher," an agent capable of independent scientific reasoning rather than assisted task execution, surfaced a clear set of preconditions. Three emerged as foundational:
The consensus view was that first-generation scientific agents function as capable copilots, not autonomous researchers. The organizational and technical infrastructure required for genuine autonomy, including agentic observability, responsible AI governance, and the kind of semantic data layer described above, is still being built.
The full panel recording covers additional ground, including concrete examples of where agentic AI has already changed the pace of work in early drug discovery, how organizations are approaching governance and observability for agentic systems, and a substantive discussion on whether advances in context window capacity will eventually reduce the need for structured knowledge graphs.
You can watch the full panel discussion recording for these insights and more.
© 2026 ONTOFORCE All right reserved
