How Merck and Mastercard built the foundations that make AI actually work

Merck’s foundational infrastructure drives AI success

Merck’s progress with agentic AI is built on the kind of discipline that turns good ideas into working systems. Before deploying a single AI agent, Merck first focused on the fundamentals: reliable, scalable infrastructure that could handle billions of data points and connect seamlessly across platforms. Sean Finnerty, Vice President of Digital Platforms at Merck, explained that the company applied lessons from the cloud computing era, when many organizations rushed into adoption without proper groundwork. This time, Merck decided to get the “plumbing” right first.

The company’s digital base spans 2,500 Amazon Web Services (AWS) accounts, multiple Microsoft Azure environments, new Google Cloud integrations, and 47 edge locations. Together, these systems manage many petabytes of structured and unstructured data across Oracle, SQL, Excel, and other repositories. Merck built scaffolding that ensures this data moves securely and retains the right context, whether it flows through Databricks, Redshift, or other analytic platforms.

This infrastructure is the foundation for sustainable innovation. Without it, AI efforts risk becoming scattered, costly, and hard to maintain. Finnerty’s message is simple but powerful: you can build hundreds of AI tools, but without structure, they’ll collapse under their own complexity. The focus on integration, security, and interoperability means Merck can scale AI initiatives faster while minimizing technical debt that usually slows large enterprises.

For executives, this approach is instructive. Laying down the groundwork first may seem slow, but it creates speed later on. Investing in scalable digital architecture early ensures that when AI scales, across departments or business units, it does so without friction.

AI accelerates drug discovery and streamlines regulatory compliance

Merck’s agentic AI programs are reshaping how the company develops and launches new medical treatments. In pharmaceutical research, timing often defines success. A single discovery cycle can take years, but with AI, certain cycles have been reduced by one-third. That’s an entire year shaved off the road from lab to patient. By training AI systems to analyze molecular structures and disease states more effectively, Merck has made early-stage discovery significantly faster and more targeted.

Regulatory compliance, one of the most stringent areas in pharma, is also being transformed. Merck’s AI models now generate marketing drafts that are 99% compliant before human review. These drafts move through approval cycles that once took months in a matter of days. Review and delivery times have improved by 70 to 80 percent. Rather than humans leading long, manual workflows, the system now operates on a “human-as-governor” model: AI drafts, and humans verify. This shift keeps oversight intact while clearing operational bottlenecks that limit speed and capacity.

For leaders, the implications are clear. AI is redefining the pace of regulated business. Efficiency gains at this scale convert directly into competitive advantage, especially in sectors where innovation speed affects both cost structures and patient outcomes.

Sean Finnerty emphasized that Merck’s goal isn’t to replace human judgment, but to amplify it. With compliance accuracy at near perfection and reduced cycle times, the company is proving that AI can maintain accountability while driving operational acceleration. For enterprises weighing AI adoption, the lesson is that infrastructure and governance are not constraints, they’re what allow AI to scale safely and deliver measurable impact.

AI-driven application modernization at merck

Merck is using agentic AI to modernize how it manages and updates its internal systems. Historically, enterprise software updates involved long cycles, large teams, and significant overhead. Today, AI agents can map architectures, catalog APIs, analyze data flows, conduct authentication and authorization checks, and even refactor code. Tasks that required weeks or months of manual effort now take significantly less time and cost a fraction of what they once did.

The company’s new approach has transformed application modernization from a complex technical project into a continuous and automated process. By using AI to understand dependencies and produce deployment scripts through platforms like Terraform, Merck ensures smoother system upgrades and less operational disruption. Engineers focus on design and integration while AI handles repetitive audit and optimization steps.

For executives, this signals a structural shift in how software management contributes to competitive performance. Modernization has evolved from maintenance to a driver of adaptability. When systems are continuously documented, secured, and updated through AI, organizations stay ready for emerging workloads and new technologies instead of reacting to them. This reduces operational drag and aligns IT capabilities directly with business agility.

Merck’s use of AI in legacy system renewal also improves reliability and compliance tracking, key priorities in heavily regulated environments. Each update cycle strengthens the system, ensuring that digital infrastructure doesn’t become the constraint but rather a catalyst for innovation.

Guardrails against AI hallucinations ensure reliable outcomes

Even the most advanced AI models can make inaccurate or speculative assumptions, often referred to as hallucinations. Sean Finnerty and his team at Merck encountered this issue when AI systems produced invalid functions and incorrect testing scenarios. Instead of scaling fast and accepting errors, Merck focused on building trust into automation. Their system now uses layered guardrails to cross-verify AI decisions.

To achieve reliability, the team applies a multi-AI validation method. For example, one AI model, such as Anthropic’s Claude, may generate an initial output, which is then independently reviewed by Microsoft Copilot. Confidence scores are assigned based on agreement or discrepancy between the two systems. This iterative review process filters out low-confidence results, reducing error rates and producing consistently accurate outputs.

The key insight for leaders is that control systems for AI must evolve alongside performance systems. As AI takes on more decision-making roles, organizations must ensure accountability is embedded in the workflow. Guardrails don’t slow progress, they maintain operational credibility. A single flawed output in areas like drug formulation or regulatory documentation could create major setbacks, and layered oversight helps prevent that.

Finnerty’s team has shown that pairing speed with verification leads to higher trust and better outcomes. For executives advancing AI adoption, this experience underscores the necessity of governance frameworks that are both technical and procedural. It’s not about eliminating risk entirely but about detecting and managing it with precision.

Mastercard leverages AI to streamline dispute and transaction workflows

At Mastercard, agentic AI is being deployed to simplify one of the most complex and labor-intensive parts of financial services, the transaction dispute process. Andrew Reiskind, Chief Data Officer, explained that a chargeback or fraud dispute touches multiple data streams and business entities, each governed by different rules and timeframes. These interactions generate both structured data, such as transaction codes, and unstructured information, such as consumer complaints or merchant statements.

AI agents are helping Mastercard automate the orchestration of these processes. They assist in categorizing data, identifying missing details, and prompting human agents when needed. The result is greater speed, improved accuracy, and reduced operational costs. The system handles deterministic data, where outcomes are clearly defined, while also managing probabilistic scenarios that require contextual assessment. This combination allows Mastercard to accelerate resolutions without compromising compliance or customer trust.

For executives and decision-makers, this shows how automation and data intelligence can reshape high-stakes service operations. Streamlined workflows not only lower overhead costs but also reduce resolution times for consumers and merchants. However, Reiskind emphasized the importance of balance: full automation without human oversight can risk fairness and accuracy, which are crucial in maintaining consumer confidence. Mastercard’s approach ensures that efficiency is achieved without eroding trust, each process is faster, but still governed by measurable accountability.

This model demonstrates how large-scale enterprises can implement agentic AI in regulated environments while keeping human decision-making at the center of oversight. It represents a deliberate move from reactive case management toward proactive and structured intelligence.

Balancing AI risk through strategic assessment and governance in financial services

Mastercard’s next focus is understanding and managing AI risk with precision. Andrew Reiskind noted that AI inevitably carries a margin of error, and organizations must determine what level of risk is acceptable before deployment. He suggested that both minor and major outcome variances need to be defined clearly, supported by a framework that calculates not just the likelihood of errors but their potential business and reputational costs.

This type of governance requires analyzing each function AI performs and quantifying its impact in measurable terms. Reiskind pointed out that while it is difficult to forecast actual usage and costs due to real-world variability, structured analysis can still guide responsible decision-making. The process includes cost-benefit assessments, performance monitoring, and validation steps to ensure the AI operates within approved thresholds of reliability.

For C-suite leaders, this approach emphasizes the necessity of embedding governance into AI strategy rather than treating it as a compliance add-on. Strategic AI use is not only about performance, it’s about predictability and controlled exposure to risk. Setting defined boundaries for acceptable error rates helps prevent operational surprises and builds confidence with regulators, partners, and customers.

Reiskind’s message aligns with a broader shift across financial services: AI success depends as much on disciplined governance as on technical capability. With the right oversight and risk definition in place, AI becomes a tool that strengthens brand credibility, operational integrity, and customer trust.

Key highlights

• Build infrastructure before scaling AI: Leaders should invest in robust, interoperable digital infrastructure early to ensure scalable and secure AI deployment. Merck’s success shows that strong data pipelines and cross-platform integration prevent technical debt and enable rapid innovation.
• Use AI to shorten critical cycles without losing compliance: AI can dramatically reduce R&D and regulatory approval times when paired with strong human oversight. Merck’s 33% faster drug discovery and 80% faster compliance approvals demonstrate that governance and speed can coexist effectively.
• Automate modernization to unlock agility: Executives should apply AI to automate legacy software management for greater operational efficiency. Merck’s use of AI in architecture mapping and code refactoring cuts costs, reduces manual workload, and boosts organizational adaptability.
• Establish layered AI validation frameworks: Companies must create multi-tiered guardrails to manage AI errors and hallucinations. Merck’s use of multiple models to cross-check outputs increases accuracy and builds confidence in mission-critical AI systems.
• Apply AI to optimize complex, data-heavy processes: Mastercard shows that AI can streamline dispute handling by automating structured and unstructured workflows. Leaders in service-heavy industries can replicate this to reduce labor intensity and improve response times while preserving customer trust.
• Define and manage acceptable AI risk from day one: Executives should quantify acceptable error margins and integrate governance into AI design. Mastercard’s approach highlights that strategic risk tolerance and strong oversight turn AI into a reliable asset instead of a liability.