As the pharmaceutical industry accelerates its adoption of artificial intelligence, a critical bottleneck has emerged: the “Sensor-to-Decision” gap. While AI tools can analyze vast datasets and generate insights at unprecedented speeds, the FDA and other global regulatory bodies are demanding more than just results—they are demanding explainability.
At the Industrial AI Summit 2026, industry leaders tackled this challenge head-on during the session, “You Closed the Sensor-to-Decision Gap. FDA Will Ask You to Explain It.” The panel featured Adam Procopio, Scientific Associate Vice President at Merck, and Kristen Sauter, President and General Manager of Life Sciences at Adlib Software. The discussion centered on the evolving requirements for defensible AI, the sanctity of source data, and the future of human-in-the-loop oversight.
The Strategic Dilemma: Build vs. Buy in AI Infrastructure
One of the most pressing questions facing pharmaceutical leaders is whether to develop AI explainability frameworks in-house or rely on vendor-provided solutions. This is not merely a budgetary question; it is a fundamental architectural decision that impacts long-term compliance and scientific integrity.
The Hybrid Architecture Approach
Adam Procopio of Merck argues that there is no “one-size-fits-all” answer. Instead, he advocates for a hybrid architecture. "We generally prefer leveraging vendor platforms for core AI infrastructure," Procopio explained. However, he emphasized that explainability, evidence traceability, and scientific defensibility are non-negotiable areas that require significant internal customization.
The core of the strategy, according to Procopio, is to use commercial capabilities as a robust foundation, while internal teams build the domain-specific “reasoning, traceability, and governance layers.” By maintaining control over these layers, Merck ensures that their AI outputs align with specific regulatory histories and scientific standards that a generic vendor might miss.
The Vendor Perspective: Accelerating Pattern Recognition
Kristen Sauter of Adlib Software provides the vendor’s vantage point, highlighting where organizations often waste resources. She argues that companies should not attempt to rebuild the “document accuracy layer”—the infrastructure that normalizes, structures, and validates documents before they reach an AI model.
"The rules for what makes a batch record processable, what a supplier Certificate of Analysis (CoA) must contain, or how a deviation log should be structured are not proprietary," Sauter noted. "These are industry patterns."
Sauter suggests that vendors who have worked across dozens of pharma organizations bring years of "pattern recognition" that would take an internal team eighteen months to rediscover. The real value for internal teams, she argues, lies in the "scientific judgment layer"—the interpretation of evidence relative to a company’s specific regulatory posture, risk tolerance, and program history.
Auditability and the Challenge of Source Integrity
The regulatory requirement for auditability is straightforward: if an AI model makes a recommendation, the organization must be able to trace that recommendation back to the original source. But a deeper, more technical challenge persists: How can you prove that the source document has not been altered or degraded during the AI’s ingestion process?
The "Decision Lineage" Gap
Procopio highlighted that while citation is necessary, it is insufficient for total compliance. "Citation alone is not the same as proving authenticity," he noted. For Merck, the gold standard remains the underlying quality and document management systems (DMS) that maintain version control and audit trails.
The emerging frontier is what Procopio calls "decision lineage." It is no longer enough to know which document informed an AI’s answer. Pharma must now track which downstream analyses, regulatory submissions, and internal decisions were influenced by that specific document. If the source record changes, organizations need automated systems that can trigger a reassessment of every downstream decision that relied on the outdated information.
Rendering Fidelity: The "What the AI Saw" Problem
Sauter expanded on this by introducing the concept of "rendering fidelity." When an AI processes a document, it must be looking at a faithful representation of the approved record, not a reconstructed approximation created by a low-fidelity ingestion engine.
"When we process a document, we lock a version-specific rendering at the point of intake," Sauter explained. "That rendering is what the AI sees, and it is what’s preserved in the audit record." This approach ensures that if a regulatory auditor asks three months later what the model saw when it made a specific recommendation, the organization can produce the exact, locked input used during that moment of decision. This creates a technical foundation for the authenticity claims that regulatory agencies now demand.
Scaling Human Oversight: From Proofreading to Reasoning
A major anxiety among pharma leaders is the volume of output generated by AI. If a model can produce a draft regulatory report in seconds, how can a human reviewer possibly keep pace without becoming a bottleneck?
Moving Beyond the "AI-Supported Editor"
Procopio argued against the notion that humans should review every piece of AI-generated content. "I don’t think the long-term answer is having humans review every piece of AI-generated content," he stated. "AI can create information much faster than people can read it."
Instead, he proposed a risk-based oversight model:
- Routine Tasks: Let AI handle the generation and normalization of low-risk content.
- High-Risk Decisions: Focus human expertise on exceptions, conflicting evidence, and areas requiring deep scientific judgment.
Procopio believes the industry needs to pivot from using humans as “editors” to using them as “reasoners.” The goal is not for the scientist to proofread the AI’s grammar, but for the scientist to evaluate the AI’s logic, confidence levels, and the underlying evidence used to reach its conclusions.
Implications for the Future of Pharma AI
The Industrial AI Summit 2026 underscored that we have moved past the "hype" phase of AI in pharmaceutical manufacturing and R&D. The focus has shifted toward the "governance" phase.
Key Takeaways for Industry Leaders:
- Stop Rebuilding Infrastructure: Use vendors for document ingestion, normalization, and pattern recognition. Reserve your internal talent for the domain-specific scientific logic that differentiates your products.
- Invest in "Decision Lineage": Auditability is not a static report; it is a dynamic map of how data flows from source to decision. Automated reassessment of downstream impacts when source data changes will be the next major regulatory requirement.
- Redefine the Human Role: The future of quality assurance in AI is not manual review of every output. It is the architectural design of oversight systems that prioritize human review for high-risk, high-judgment decisions.
Chronology of the Regulatory Shift
- 2023–2024: The industry experimented with Large Language Models (LLMs) for content generation.
- 2025: Regulatory agencies began issuing warnings regarding the "black box" nature of AI decisions, leading to a surge in internal explainability initiatives.
- 2026 (Current): The focus has matured into establishing formal "Decision Lineage" and "Rendering Fidelity" standards, ensuring that AI-generated insights are as defensible as those produced by human teams.
As Procopio concluded during the session, "The goal isn’t AI-supported editors. It’s AI-supported decision makers." For the pharmaceutical industry, the path forward is clear: success will be defined by those who can best harmonize the speed of AI with the rigorous, evidence-based scrutiny required to ensure patient safety and regulatory compliance.
This report is based on the session “You Closed the Sensor-to-Decision Gap. FDA Will Ask You to Explain It,” held at the Industrial AI Summit 2026. Panelists included Adam Procopio (Merck), Kristen Sauter (Adlib Software), and Maya Schushan-Orgad (Teva Pharmaceuticals), with moderation by Rick Franzosa (Tech-Clarity). The session was sponsored by Adlib Software.
