The Conversation Has Changed
At the 2026 Pharmaceutical Management Science Association (PMSA) Fall Symposium, Ankush and I presented “From Metadata to Meaning: Operationalizing Semantic Layers and GraphRAG for AI-Ready Life Sciences Analytics.” The conference was valuable for reconnecting with clients, peers, and other leaders across the industry. More importantly, it made one shift unmistakable: AI in life sciences is beginning to move from promise and experimentation toward tangible capabilities.
Most organizations are no longer asking whether they can build an AI assistant, a next-best-action agent, or a generative AI copilot. Many already have demos and point solutions. The harder question is whether those capabilities can be trusted, reused, and scaled across brands, functions, and geographies. Only 22% of life sciences leaders say they have successfully scaled AI [1], and across industries just 39% of organizations attribute any EBIT (earnings before interest and taxes) impact to it [2].
The model is rarely the limiting factor. The missing piece is usually shared business meaning: a governed foundation that tells an AI agent what the organization’s data, metrics, relationships, and rules actually mean.
Why Promising Pilots Stall
Most AI programs begin one use case at a time. A brand team commissions a forecasting assistant, sales operations pilots a call-planning agent, and an analytics team adds a copilot to a dashboard. That approach produces visible wins and lets functions move quickly. But each initiative often prepares its own data, defines its own metrics, and encodes its own business rules.
Commercial life sciences data makes this fragmentation particularly risky. NBRx (new-to-brand prescriptions) can be calculated from claims data or specialty pharmacy feeds, each with different coverage and lag. Territory alignments change. HCP (healthcare professional) specialty codes differ across systems and markets. In a 2025 Veeva survey of 116 commercial biopharma leaders, 95% reported remapping global specialties to local ones, while 72% said matching HCP data across sources can take up to 100 days [3].
When meaning is not shared, two agents can answer the same question with two different numbers, both with confidence. Trust erodes quickly. Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data [4].
The Real Differentiation Is Moving Beneath the Demo
One of my clearest observations from PMSA was how quickly AI-enabled point solutions are becoming democratized. Several competitors and industry participants demonstrated or discussed agents and AI-enabled tools for specific commercial problems. The strongest technical visions were not limited to the agent itself. They addressed the architecture, context layer, and engineering required to make many agents work together.
That is the strategic shift. A good demo is increasingly easy to replicate. What is difficult, and far more valuable, is the underlying “AI plumbing”: the semantic foundation, context, governance, tooling, and integration patterns that allow AI to scale reliably. Gartner now advises data and analytics leaders toestablish a context layer as a core component of their infrastructure [5] and predicts that universal semantic layers will be treated as critical infrastructure by 2030 [6].
Shared Meaning Is the Foundation for Scalable Agents
A semantic layer is a governed business dictionary written in code. It defines a metric once, including its calculation logic, approved data sources, owner, lineage, and access rules. A knowledge graph complements that layer by mapping how products, HCPs, accounts, territories, channels, and business rules relate to one another. GraphRAG (graph-based retrieval-augmented generation) then allows an AI model to retrieve those connected facts before it generates an answer.
Consider a commercial leader asking why NBRx for a brand declined among cardiologists in the Northeast. A governed agent should not guess what NBRx means or which source is authoritative. It should retrieve the certified definition, follow the knowledge graph from specialty to territories and aligned accounts, apply the governing access rules, generate the appropriate query, and explain which definition and source it used.
This distinction matters because enterprise questions are rarely simple lookups. On Spider 2.0, a benchmark of 632 real enterprise text-to-SQL tasks, the best agent framework at publication solved 21.3% of tasks, compared with 91.2% on an earlier academic benchmark [7]. Research also indicates that graph-based retrieval adds the most value for multi-step questions connecting several entities [8], which closely resembles many commercial “why” questions. Open standards such as the Model Context Protocol can provide a consistent way for agents to access context, while the semantic layer defines what that context means [9].
Context Governance Must Move to the Forefront
A particularly useful conversation at the symposium reinforced that governance cannot be an afterthought. As AI becomes embedded in workflows, governance must extend beyond data to the context, metadata, agents, prompts, tools, and outputs that shape a decision.
Governance cannot rely on a static policy document that will not keep pace with agents acting in real time. Controls need to live where agents obtain and use context. One practical mechanism is a context contract: a machine-readable specification that binds each metric or business entity to its physical source, calculation logic, owner, lineage, permitted uses, and access rules. Gartner predicts that by 2030, half of organizations will use AI agents to translate governance policies into machine-verifiable data contracts [6].
Regulatory thinking offers a useful design pattern even when it does not directly govern commercial analytics. FDA draft guidance evaluates model credibility against a defined context of use [10], while joint EMA and FDA principles emphasize traceability, data provenance, and documented processing steps [11]. The transferable lesson is straightforward: define where an agent will be used, what evidence it relies on, who owns the meaning, and what controls apply before the agent scales.
The Shift From AI-Assisted to AI-Native Work
The symposium also reinforced a broader change in the nature of work. Historically, considerable manual effort went into finding, preparing, reconciling, and curating information before a team could solve the actual business problem. AI is making portions of that work easier. The source of value is therefore shifting from assembling information toward interpreting it, designing solutions, exercising judgment, and making decisions.
Don’t simply add AI to your existing processes. Redesign workflows so AI becomes part of the operating model. McKinsey found that AI high performers are nearly three times as likely as other organizations to have fundamentally redesigned individual workflows, 55% compared with 20% [2]. Adding an agent to an unchanged process caps the value it can create.
For consulting and analytics teams, this means institutional knowledge must become something teams actively use, not something that sits stored. The collective context, experience, and judgment of teams should become a connected organizational capability that improves how people prepare, analyze, create, and decide. Internally, we are building a practical environment for capturing context, testing governance, and rethinking work around AI, rather than waiting for a perfect end-state architecture.
Four Priorities for Analytics and Technology Leaders
Fund the shared foundation, not only the next use case-
Continue to deliver point solutions, but use each one to strengthen a reusable context layer. The goal is for every new AI capability to inherit certified definitions, relationships, controls, and integration patterns rather than rebuilding them.Certify the highest-value business meaning first-
Start with the metrics and entities that drive the most consequential decisions, such as NBRx, market share, territory performance, HCP and account relationships, patient journey stages, and targeting rules. Give each definition a named business owner and an explicit source of truth.Build and learn before the architecture is complete-
A working context layer for one brand, workflow, or commercial question will expose conflicting definitions and governance gaps faster than a prolonged design exercise. Use the first implementation as a proving ground, then generalize what works.Design governance and workflow change into the first release-
Lineage, access, ownership, observability, and human review should be part of the architecture from day one. At the same time, redesign the surrounding workflow so the agent is accountable for a meaningful step in the process, not merely an additional interface.
How ProcDNA Is Approaching the Opportunity
Our PMSA paper proposed an architecture that unifies a domain knowledge graph and ontology with a semantic and calculation layer, then exposes that governed meaning to agents through context contracts. Our approach pairs the design of knowledge graphs and agent workflows with the data platforms, integrations, and governance controls that keep the context accurate as business definitions, source systems, and organizations change.
The objective is not a single agent. It is a foundation where every new agent starts with consistent business meaning and can work with other agents and workflows reliably. That foundation also becomes organizational memory: the definitions, relationships, decisions, and rules that allow the business to operate with continuity as data and teams evolve.
The Question Leaders Should Ask Now
The promise of AI is beginning to materialize, but the next phase will not be decided by who launches the most demos. The organizations that create the shared meaning, governance, and operating model to let AI work at scale will decide it.
The question is no longer simply, “What agent should we build next?” It is: “Are we building the context and institutional memory that every future agent will need?”
Organizations that invest in that foundation now will be able to add new use cases faster and govern them more confidently, moving from isolated AI assistance toward an AI-native way of working.
References
1. Deloitte Center for Health Solutions. 2026 life sciences outlook. Deloitte Insights; 2025 Dec 9.
2. Singla A, Sukharevsky A, Hall B, Yee L, Chui M. The state of AI in 2025: agents, innovation, and transformation. McKinsey & Company; 2025 Nov.
3. Veeva Systems. The state of data, analytics, and AI in commercial biopharma. 2025 Oct 7.
4. Gartner. Lack of AI-ready data puts AI projects at risk. 2025 Feb 26.
5. Gartner. Lack of semantics causes inaccurate AI agents and wasted spending. 2026 May 11.
6. Gartner. Top predictions for data and analytics in 2026. 2026 Mar 11.
7. Lei F, et al. Spider 2.0: evaluating language models on real-world enterprise text-to-SQL workflows. ICLR 2025.
8. Han H, et al. RAG vs. GraphRAG: a systematic evaluation and key insights. arXiv:2502.11371.
9. Anthropic. Donating the Model Context Protocol and establishing the Agentic AI Foundation. 2025 Dec 9.
10. U.S. Food and Drug Administration. Considerations for the use of AI to support regulatory decision-making for drug and biological products: draft guidance. 2025 Jan.
11. European Medicines Agency and U.S. Food and Drug Administration. Guiding principles of good AI practice in drug development. 2026 Jan 14.