Merck is using AI agents to reduce drug discovery cycles by a third and send tailored marketing materials 80% faster — but VP of digital platforms Sean Finnerty says the only reason it works is because they built the infrastructure first.
And the pharmaceutical manufacturer is seeing promising early results: AI is generating marketing drafts that are “99% accurate” in terms of compliance, reducing review cycles from months to days and speeding up delivery by 70% to 80%. Meanwhile, in the company’s medical research, an AI-assisted discovery cycle was reduced by 33%.
Still, agentic AI only works if companies build the underlying “plumbing” first, Finnerty said at a recent AI Impact Series event about digital platforms and services.
“If we do this all at once, we’ll end up with thousands and thousands of things that will eventually become debt that we’ll have to deal with later,” he said. “And that’s going to hamper any further innovation.”
start with plumbing
Merck’s plumbing-first strategy comes from lessons learned during the early days of cloud in the 2010s “when nobody knew what was going on,” Finnerty said.
Getting the cloud right meant building from the ground up; At Merck, that infrastructure now supports 2,500 AWS accounts, multiple Microsoft Azure subscriptions, and new Google Cloud Platform (GCP) integrations.
“AI will be exactly the same thing,” Finnerty said. “We will have thousands and thousands of agents.” Then questions pile up: How do you register them? How do you secure them? How do you ensure they’re connected to the right tools, and have access to the right data and the right context?
Reference distribution is also important; Merck works with three hyperscalers and has forty-seven edge locations and hundreds of databases. “Many, many petabytes” of structured and unstructured data are stored in Oracle databases, SQL databases, Excel spreadsheets, phone transcripts and other repositories, Finnerty said.
He explained that his team is building scaffolding to provide meaningful context in different situations. Data must be organized and fed into a variety of platforms, because “there is no single solution to every problem.” Sometimes it’s Databricks, sometimes it’s Amazon Redshift, “plus four other things.”
The goal is: “Let’s make it easy and frictionless for people, and make it secure, and make sure it integrates well with MCP [model context protocol]and A2A [Agent2Agent]And upstream compute,” Finnerty said. “If you want to run stuff on GCP or you want to run stuff on AWS, we’ve got the plumbing so you can run your adjacent workloads wherever you want.”
How Merck is using agents
As it builds out its technology pipeline, Merck is experimenting with agents in regulated enterprise operations, scientific discovery workflows and app modernization.
In particular, AI is accelerating drug discovery. Finnerty explained that scientists look at molecular structures and disease states to determine whether a given condition is druggable. But even if a disease state is known, it can take years to develop a drug to target it.
Now with AI, teams are starting to see “very promising things,” such as reducing a particular research cycle by one-third. “It’s a year off from the life of the discovery cycle,” Finnerty said. “Which means, theoretically, we can get it to the patient who needs that therapy a year sooner.”
Once developed and approved, these products are regulated and the marketing material around them must be clearly and unambiguously expressed. “The way you communicate that information per market, per country, per state, per region, all of that is very carefully controlled and regulated,” Finnerty said. It’s also variable: An advertising campaign for the vaccine in the state of Georgia looks very different from one launched in Canada.
Historically, humans performed due diligence to ensure that a company complied with various laws. Draft material undergoes iterations of reviews; When a mistake is discovered, it is “taken back to the beginning, and it happens again, and then it takes weeks and months,” Finnerty said.
But now, AI can do it “much more effectively,” and the process is increasingly evolving from essentially human-in-the-loop. "As human-governor." With human oversight, AI can deliver a first draft in a day or week that is 99% there, allowing teams to send out content 80% faster.
Meanwhile, when it comes to app modernization, AI can explore architecture, document data interactions, APIs, network paths, and perform authentication checks and authorizations; It can also write code for Terraform for deployment and refactor JavaScript in Python.
Where the company previously spent weeks, months and hundreds of thousands of dollars to update an application, agents are now handling the job through signals, Finnerty said.
run into "strangeness"
This does not mean that there are not significant challenges; Finnerty noted that his team was facing some “wackiness”; For example in automated code and scenario testing. AI has clearly created scenarios, whether due to the wrong context, infrastructure, “or if it was just being creative, ‘You should test these three functions that don’t even exist in the code you’re trying to test.'”
“It surprised me a little bit because I thought we were past some of the challenges of hallucinations in these later models,” he said.
To address this, his team has built guardrails to keep hallucinations to a minimum, essentially using AI to monitor people and apply confidence scores. So if the cloud created the first output, they would instruct Microsoft Copilot to evaluate it.
“So if you ask something once, have the AI check it out, then ask a third time, the confidence increases each time, and it reduces some of the waste that is created in the early stages,” Finnerty said.
Use cases for agentic AI in financial services
Meanwhile, at Mastercard, Chief Data Officer Andrew Reiskind and his team are focusing on agentic experimentation on highly streamlined transaction and dispute workflows. That said, a chargeback or fraud dispute is not a single incident.
When a consumer disputes a charge (usually online), “a whole other process starts on the back-end that is very labor-intensive,” Reiskind said.
Mastercard must collect details regarding the actual dispute; The merchant then does its own investigation (Was the card lost or stolen? Does the consumer frequently incur dispute fees?). Furthermore, the networks sitting in the middle have their own rules for timing and presenting information.
“You have each of these steps, many of which are unstructured, but also have structured data elements,” Reeskind said. Whether a card is lost or stolen, it is structured, but consumer complaints are about “unstructured data of questionable reliability”.
“So you’re sitting there with a decision system that has deterministic decisions, but also probabilistic decisions,” he said.
This problem can be accelerated and potentially solved by AI agents, but it can be a complex process: What tasks are you delegating to agents? When are they going to hand things back over to human representatives? How many agents are you ultimately using? What are the cost implications?
Then there are the reputational questions and costs: Did you potentially call a consumer a liar when they weren’t?
“This is an exact problem where you, as a bank, want to maintain trust with your consumer,” Reiskind said. “But you also want to make it efficient and take costs out of the system.”
PB&J vs. Turkey Mistake: Determine Which Risks Are Acceptable
There will always be risk with AI, Reeskind said, and enterprises should assess it from the beginning of product design. There is also a question of acceptable risk.
For example: Did you serve a customer a peanut butter jelly sandwich instead of a turkey sandwich (a minor inconvenience)? Or did you serve gluten to someone with celiac disease?
“Is it an acceptable risk if even one percent of the time he makes a mistake? If so, let’s move on to the next step about how you’re reducing that risk,” Reiskind said.
Leaders must conduct cost-benefit analyses, breaking down problems into their “component pieces” and calculating the costs for each. But these are estimates; Actual usage is almost impossible to predict, Reeskind said. “Getting the costs is not an easy process,” he said. “But it’s possible.”
<a href