Merck and Mastercard are seeing real agentic AI results. Both say the plumbing came first.

Pharmaceutical giant Merck and payment technology company Mastercard have both begun deploying autonomous artificial intelligence agents across their enterprise operations with tangible results, yet both organizations emphasize that their technological success stems from foundational infrastructure work rather than rushing to implement the latest AI capabilities. Merck's Vice President of Digital Platforms Sean Finnerty reported that the company has achieved a one-third reduction in drug discovery cycles and accelerated the delivery of compliant marketing materials by seventy to eighty percent through strategic agent deployment. Similarly, Mastercard's Chief Data Officer Andrew Reiskind is directing his team's efforts toward automating complex transaction and dispute workflows using agentic AI. These achievements underscore a critical lesson that has resonated across both organizations: the underlying technical infrastructure and data architecture must be deliberately constructed before introducing autonomous agents into production environments. Understanding why these organizations prioritize foundational infrastructure requires examining the broader trajectory of enterprise technology adoption. The transition to cloud computing during the 2010s presented similar challenges, as Finnerty recalled, when organizations struggled to determine the optimal approaches for cloud deployment and integration. Merck learned from that experience and built its cloud infrastructure methodically from the ground up, now supporting two thousand five hundred Amazon Web Services accounts alongside numerous Microsoft Azure subscriptions and emerging Google Cloud Platform integrations.

That foundation now positions the company to handle what Finnerty predicts will be thousands of autonomous agents operating across different departments and functions. The lesson applies directly to agentic AI deployment: without proper foundational architecture, organizations risk creating technical debt that will ultimately impede future innovation and scalability. Mastercard faces comparable complexity given the nature of its business, where transaction disputes and chargebacks trigger multifaceted processes involving consumers, merchants, internal decision-making systems, and regulatory requirements that must all coordinate seamlessly. Merck's practical applications of agentic AI reveal both substantial achievements and concrete technical details about how the company operationalizes this technology. In pharmaceutical research, AI agents are helping scientists evaluate molecular structures and disease states to determine drugability, with one discovery cycle already reduced by approximately one year through AI acceleration. Within marketing compliance, a domain historically requiring months of human review cycles, AI systems now generate first drafts that are ninety-nine percent compliant with regulatory requirements on the first attempt, enabling the company to compress review timelines from months to mere days. Finnerty acknowledged that managing these agents requires solving infrastructure puzzles: how to register thousands of agents, secure them appropriately, ensure they access the correct tools and data, and deliver meaningful context within their operational environment.

Merck manages this complexity across three major cloud providers and forty-seven edge locations storing petabytes of structured and unstructured data across Oracle databases, SQL systems, Excel spreadsheets, telephone transcripts, and other repositories. His team is deliberately building scaffolding that makes context delivery frictionless while maintaining security and integration with emerging protocols like model context protocol and Agent-to-Agent communication. The implications of this infrastructure-first approach extend beyond Merck's immediate operational gains to suggest a broader industry pattern in how mature enterprises should approach transformative technologies. Finnerty cautioned against one-off implementations, warning that ad-hoc agent deployments would eventually become problematic technical debt requiring remediation. Instead, organizations should view agentic AI as the enterprise computing paradigm of the coming decade, requiring the same level of architectural rigor applied to previous technology transitions. This perspective aligns with how Mastercard conceptualizes its challenge: transaction disputes represent highly orchestrated workflows where multiple actors operate within structured rules and unstructured data simultaneously. When consumers dispute charges, they initiate backend processes involving merchant investigations, network rules regarding timing and information submission, and decisioning systems that must balance deterministic logic with probabilistic assessments.

Automating these workflows with agents promises significant efficiency gains, but Reiskind emphasized that the financial and reputational stakes demand careful risk assessment from the initial stages of solution design. Both organizations acknowledged encountering unforeseen challenges during their agentic implementation phases, reflecting the still-emerging nature of this technology domain. Merck's teams encountered what Finnerty described as "wackiness" in automated code testing scenarios, where AI agents fabricated test functions that did not exist in the actual codebase, suggesting either contextual failures or unexpected creativity on the part of the AI models. He expressed surprise that hallucination challenges persisted in newer language model iterations, having expected the technology to have matured beyond such limitations. To mitigate these risks, Merck engineered guardrails using what amounts to AI supervision of AI, employing confidence scoring mechanisms where a second AI system validates outputs from the initial agent before implementation. This cascading verification approach, where multiple independent AI systems assess the same problem, progressively increases confidence and reduces erroneous outputs from early iterations. Mastercard faces different but equally consequential risks: incorrectly processing a consumer's legitimate dispute could damage customer trust, while simultaneously manual processing of all disputes perpetuates inefficiency.

Reiskind advocated for cost-benefit analysis that distinguishes between minor inconveniences like serving the wrong sandwich versus serious harms like exposing people with celiac disease to gluten, establishing explicit thresholds for acceptable error rates before proceeding to the next implementation phase. Looking forward, organizations implementing agentic AI must monitor two specific critical dimensions that will determine success or failure in production environments. First, enterprises should track the evolution of autonomous agent registration, security, and governance frameworks, as standardized approaches emerge for managing thousands of agents operating across heterogeneous infrastructure. Merck's approach to ensuring agents connect to proper tools, access appropriate data, and maintain security across multiple cloud providers and edge locations will serve as a reference model, and early evidence of whether similar frameworks gain industry adoption will indicate whether the technological ecosystem is maturing appropriately. Second, organizations should observe how risk assessment methodologies develop for agentic systems in highly regulated environments, particularly in financial services and pharmaceuticals where errors carry substantial consequences. Mastercard's experience managing probabilistic and deterministic decision-making in dispute workflows, combined with Merck's regulatory compliance automation, will illuminate whether enterprises can establish defensible frameworks for deploying AI agents when some error rate is unavoidable. The coming months will reveal whether the infrastructure-first philosophy both organizations champion becomes industry standard practice or remains a differentiating competitive advantage for early movers.