Leading corporations are proving that agentic AI delivers tangible business value—but the transformation relies on a critical foundation. Merck and Mastercard executives revealed at a recent industry event that their early experiments with autonomous AI agents are already yielding measurable results, from accelerating drug discovery to streamlining regulatory marketing. However, both companies emphasize that these gains were only possible because they prioritized building the underlying technical infrastructure first.
The infrastructure-first foundation for agentic AI
Merck’s vice president of digital platforms, Sean Finnerty, explained that the company’s agentic AI success stems from years of deliberate infrastructure development. “If we create one-offs, we’ll end up with thousands of isolated systems that become technical debt and slow future innovation,” he warned during a recent panel discussion. The company’s digital backbone now spans 2,500 Amazon Web Services accounts, multiple Microsoft Azure subscriptions, and newly integrated Google Cloud Platform environments—an architecture built during the company’s early cloud migration in the 2010s.
Agentic AI introduces new complexity: how do you register, secure, and connect thousands of autonomous agents to the right tools and data sources? Merck’s approach addresses this by treating AI adoption like cloud adoption—requiring standardized protocols. The company works across three hyperscale cloud providers and maintains 47 edge locations while managing hundreds of databases. These repositories include petabytes of structured and unstructured data stored in Oracle systems, SQL databases, spreadsheets, and even phone call transcripts.
Finnerty described the team’s focus on building “scaffolding” that delivers precise context to AI agents based on specific scenarios. “There’s no single solution that solves every problem,” he noted, pointing to varied tools like Databricks for some workflows and Amazon Redshift for others. The goal is creating a unified environment where agents can access the right data regardless of where it resides, supported by protocols like Model Context Protocol (MCP) and Agent-to-Agent communication standards.
Pharma breakthroughs: Faster discovery and compliant marketing
Merck’s agentic AI initiatives span three critical areas: regulated enterprise operations, scientific discovery workflows, and application modernization. In pharmaceutical research, AI is making substantial progress in shortening drug development timelines.
One particularly promising case involved molecular structure analysis for a potential therapy. Finnerty reported that AI assistance reduced a specific research cycle by 33%, effectively shaving a year off the discovery process. “That translates to getting potentially life-saving treatments to patients one year sooner,” he explained. While human expertise remains essential for validation and ethical considerations, the AI acceleration enables researchers to explore more hypotheses faster.
Beyond discovery, AI is revolutionizing how Merck approaches regulatory marketing—a historically labor-intensive process. Pharmaceutical advertising must comply with varying laws across states, countries, and regions, with even minor wording changes requiring extensive legal review. Traditionally, human teams would draft materials, submit them for approval, and face repeated cycles of revisions that could stretch for months.
Now, Merck’s AI systems generate compliant drafts that are “99% accurate” from the first iteration, reducing review cycles from months to days and accelerating marketing material delivery by 70% to 80%. The evolution is moving from human-in-the-loop oversight to a “human-as-governor” model, where AI handles the heavy lifting while humans focus on strategic governance and final validation.
Modernizing legacy systems through AI agents
The company is also leveraging agentic AI to modernize outdated applications that were previously too costly and time-consuming to update. These legacy systems often lack proper documentation, making updates complex and expensive. Merck’s AI agents are now capable of discovering system architectures, documenting data flows, identifying APIs and network paths, and conducting authentication and authorization checks.
The agents can even generate deployment scripts in Terraform and refactor code between languages like JavaScript and Python. Finnerty highlighted that what once required weeks of manual work and hundreds of thousands of dollars per application now happens through simple prompts. This automation not only reduces costs but also minimizes human error in complex modernization projects.
Navigating the pitfalls of autonomous AI
Despite the progress, Merck’s team has encountered significant challenges in their agentic AI journey. Finnerty admitted encountering “wackiness” during automated code testing and scenario generation. In some cases, AI systems fabricated non-existent functions or scenarios, either due to incorrect context or over-creativity.
“We expected these later models to have moved past hallucination issues, but we still see them,” Finnerty shared. To mitigate this, the team implemented multi-layered validation: when one AI model generates output, another evaluates it for accuracy. They’ve also introduced confidence scoring systems that improve with each review cycle.
Finnerty described their approach as “using AI to supervise AI.” For example, if Claude generates an initial draft, Microsoft Copilot assesses it before human review. “Each iteration increases confidence and reduces the garbage created in early runs,” he explained. This layered validation is becoming essential as companies deploy more autonomous systems in critical business processes.
The path forward for enterprise agentic AI
The experiences of Merck and Mastercard demonstrate that agentic AI delivers real business value—but only when built on solid infrastructure. As companies scale from pilot projects to production environments, the focus must shift from experimentation to robust governance and integration.
The next phase will likely involve standardizing how enterprises deploy, monitor, and secure thousands of AI agents while maintaining human oversight. Those who get the infrastructure right today will be best positioned to capture the full potential of agentic AI tomorrow.
AI summary
Merck ve Mastercard, AI ajanslarını kullanarak süreçleri hızlandırırken altyapıyı önce inşa etmenin önemini vurguluyor. Peki bu strateji nasıl uygulanıyor?
