Merck is significantly accelerating its drug discovery timelines, reducing them by as much as one-third, and streamlining the delivery of compliant marketing materials by up to 80%. According to Sean Finnerty, VP of Digital Platforms, these advancements are contingent upon the foundational infrastructure that Merck has meticulously established.
The pharmaceutical giant is already witnessing substantial benefits, with AI-generated marketing drafts demonstrating remarkable accuracy in compliance, ensuring “99% correctness.” This has drastically compressed review cycles from months down to mere days, expediting delivery by 70% to 80%. Concurrently, in the realm of medical research, a single AI-assisted drug discovery cycle saw a 33% reduction in duration.
However, Finnerty emphasized at a recent AI Impact Series event that the efficacy of agentic AI, which refers to AI systems capable of autonomous action and decision-making, is predicated on robust underlying infrastructure. He cautioned against ad-hoc implementations:
“If we do one-offs, we’re gonna end up with thousands and thousands of things that are ultimately just gonna be debt that we’ll have to deal with later,” Finnerty stated. “And that’s gonna be a drag on any further innovation.”
Establishing Foundational Infrastructure First
Merck’s strategic emphasis on building core infrastructure before implementing advanced AI solutions stems from lessons learned during the nascent stages of cloud computing in the early 2010s. Finnerty recalled the initial uncertainty surrounding cloud adoption:
“when nobody knew what the heck was going on.”
Establishing a proper cloud foundation involved a comprehensive, ground-up approach. At Merck, this infrastructure now underpins approximately 2,500 AWS accounts, numerous Microsoft Azure subscriptions, and emerging integrations with Google Cloud Platform (GCP).
“AI is gonna be the same exact thing,” Finnerty projected. “We’re going to have thousands and thousands of agents.” This anticipates a host of critical management questions: How will these agents be registered? How will they be secured? How can we ensure their seamless connection to appropriate tools and grant them access to the necessary data and context?
The seamless delivery of context is paramount, especially given Merck’s complex operational landscape, which involves three major cloud providers, forty-seven edge locations, and hundreds of diverse databases. Finnerty highlighted the vast data ecosystem:
“Many, many petabytes” of structured and unstructured data reside within Oracle databases, SQL databases, Excel spreadsheets, call transcripts, and various other repositories.
His team is actively developing frameworks to ensure meaningful context is available across a spectrum of applications. Finnerty explained that data must be meticulously organized and integrated into multiple platforms, acknowledging that “there’s no one solution to solve every single problem.” The technology stack includes platforms like Databricks and Amazon Redshift, alongside other specialized tools.
The overarching objective is to make data accessibility and utilization “easy and frictionless for people to do, and secure it, and make sure it’s well integrated with MCP [model context protocol], and A2A [Agent2Agent], and upstream compute,” Finnerty outlined. “If you wanna run stuff on GCP or you wanna run stuff on AWS, we’ve got the plumbing in place so you can run your adjacent workloads wherever you want.”
Merck’s Agentic AI Applications
As Merck solidifies its technological foundation, the company is actively exploring agentic AI applications across several critical domains: regulated enterprise operations, scientific discovery workflows, and application modernization.
In drug discovery, AI is proving to be a significant catalyst. Finnerty elaborated on the scientific process, where researchers analyze molecular structures and disease states to identify potential drug targets. While a disease state may be understood, developing a targeted therapy can span many years.
The integration of AI is yielding “very promising things,” including the reduction of a specific research cycle by a third. Finnerty remarked on the impact: “That’s a year off of the life of the discovery cycle. Which means, theoretically, we can get it to a patient who needs that therapy a year faster.”
Post-development and approval, pharmaceutical products are subject to stringent regulations, particularly concerning marketing materials. Finnerty explained the complexity:
“The way you communicate that information per market, per country, per state, per region, is all very carefully governed and regulated.”
This communication is highly variable; marketing campaigns for a vaccine, for instance, differ significantly based on region, such as between the state of Georgia and Canada.
Historically, ensuring compliance with these diverse legal frameworks required extensive human due diligence. Draft materials would undergo multiple review iterations, and any discovered discrepancy would necessitate restarting the entire process, consuming additional weeks or months.
AI now addresses this challenge with far greater efficiency. The process is evolving from a “human-in-the-loop” model to one where humans act as “governors.” With human oversight, AI can produce an initial draft in a day or week that is approximately 99% compliant, enabling marketing teams to release materials up to 80% faster.
In application modernization, AI agents are identifying architecture, documenting data interactions, APIs, and network paths, along with performing authentication and authorization checks. They can also generate code for deployment using Terraform and refactor existing JavaScript into Python.
Finnerty noted that tasks previously requiring weeks, months, and substantial investment to update a single application are now being managed by AI agents through strategic prompting.
Navigating AI’s “Wackiness”
Despite these advancements, significant challenges remain. Finnerty acknowledged encountering instances of AI “wackiness,” particularly in automated code and scenario testing. This has manifested as AI fabricating test scenarios, potentially due to incorrect context, infrastructure issues, or what he described as the AI “getting creative with, ‘You should be testing these three functions that don’t even exist in the code that you’re trying to test.’”
“That surprised me a little bit because I thought we were further past some of the hallucination challenges in these later models,” he admitted.
To mitigate these issues, Merck’s team has implemented robust guardrails to minimize AI hallucinations. This approach involves using AI to supervise AI outputs and employing confidence scoring mechanisms. For example, if one AI model like Claude generates an initial output, Microsoft Copilot might be tasked with evaluating it.
“So if you ask something once, have AI check it, then ask it a third time, the confidence increases every time, and it minimizes some of the garbage that gets created in the early runs,” Finnerty explained.
Agentic AI Use Cases in Financial Services
Meanwhile, at Mastercard, Chief Data Officer Andrew Reiskind and his team are concentrating their agentic AI experimentation on highly structured transaction and dispute resolution workflows. Reiskind pointed out that handling a chargeback or fraud dispute is far from a singular event:
When a consumer disputes a charge, typically initiated online, “that kicks off an entire other process on the back-end that tends to be very labor-intensive.”
Mastercard must meticulously gather dispute specifics. Subsequently, the merchant conducts its own investigations, such as verifying if the card was reported lost or stolen or if the consumer frequently disputes charges. The network intermediary also adheres to its own set of rules regarding timing and information submission.
“You have each and every one of these steps, many of which are unstructured, but there are also structured data elements to this,” Reiskind observed. While information like whether a card was lost or stolen is typically structured, consumer complaints represent “unstructured data of questionable reliability.”
Consequently, “you’re sitting there with a decisioning system that has deterministic decisions, but also probabilistic decisions.”
AI agents hold the potential to accelerate and potentially resolve these complex processes. However, the implementation involves careful consideration of task delegation to agents, determining when human representatives should intervene, optimizing the number of agents utilized, and analyzing the associated cost implications.
Furthermore, reputational risks and associated costs must be evaluated. Reiskind questioned the potential fallout from mistakenly accusing a consumer:
“It’s an exact problem where you want to, as a bank, maintain trust with your consumer. But you also wanna make this efficient and take costs out of the system.”
Assessing Acceptable Risk: The PB&J vs. Turkey Analogy
Reiskind stressed that inherent risks are unavoidable with AI, and enterprises must conduct thorough risk assessments from the outset of product design. A critical aspect of this assessment involves determining the level of acceptable risk.
He illustrated this with an analogy: distinguishing between a minor inconvenience, such as mistakenly serving a peanut butter and jelly sandwich instead of a turkey sandwich, and a severe issue, like serving gluten to a person with celiac disease.
“Is it an acceptable risk if one percent of the time it makes the mistake? If it is, let’s go to the next stage of how you’re mitigating that risk,” Reiskind advised.
Business leaders are urged to conduct rigorous cost-benefit analyses, breaking down complex problems into their fundamental components and estimating the cost for each. However, Reiskind cautioned that these are projections, as accurately forecasting real-world usage remains exceptionally challenging. “It is not a simple process to get to the cost. But it is doable.”
Business Style Takeaway: The successful deployment of advanced AI, like agentic systems, hinges on robust foundational infrastructure, not just innovative algorithms. Companies must prioritize building scalable, secure, and integrated data and cloud platforms to avoid technical debt and unlock AI’s true potential for efficiency gains and accelerated innovation across R&D and operations.
Based on materials from : venturebeat.com