Adding AI to broken workflows just burns cash. Insiders at Sanofi and Merck reveal the key to getting your organization AI-ready.
Sydney Scott
Editorial Strategist, AI
Workday
In the 1980s, General Motors spent billions to build a cutting-edge factory, deploying robotics to replace human assembly workers. But GM made a fatal mistake by simply bolting high-tech automation directly onto disjointed, inconsistent assembly lines without fixing the underlying workflows. The result was high-speed chaos. Robots famously spray-painted each other, dropped engines onto concrete floors, and welded car doors shut. Instead of modernizing its operations, GM had simply mechanized dysfunction.
Today, companies adopting AI are making some of the same mistakes. Modern software manages administrative tasks, staff, and money well, but employees still bridge the gaps by hand. Many spend their days passing requests and chasing approvals or hunting down policies. They act as human APIs, stitching together broken workflows.
This is where AI transformations usually lose steam because bolting an agent onto a broken workflow fixes nothing. The agent simply learns to make the same broken handoffs much faster.
To see value from AI, companies must clean up messy processes and structural debt before handing tasks to agents.
“We went after the foundation of work. It’s organizing the data to be sustainable.”
—Danielle Bushen, Global Head of People & Culture Technology, Sanofi
Report
The hype around AI made it easy to believe the technology would solve all problems. Just turn on the model, point it at your data, and watch the magic happen. Practitioners who have actually done this at scale tell a different story.
Sanofi, the healthcare biopharmaceutical company, did some major cleanup in order to empower employees with AI.
“We went after the foundation of work. Seventy percent of this project is fixing process engineering. It’s organizing the data to be sustainable,” Danielle Bushen, Sanofi’s global head of people and culture technology said. “It’s actually doing decision management the way it was intended, so that we can automate.”
Fidelity National Information Services (FIS) arrived at the same conclusion. During a keynote at Workday Rising 2026, Steve Bugley, FIS senior vice president of people operations, dsaid his team made a deliberate choice to reengineer workflows before adding AI to the equation.
“What we didn't want to do was just lay AI on top of our old processes. We wanted to say, ‘Okay, let's reengineer everything,’” he explained. “We didn't want people to think, ‘This is the same process I've always had. I still need to go to three systems. It's not integrated.’ …So we stopped and said, let's think about this differently.”
Few organizations have been as candid about the extent of what's actually buried in a mature HR system as Sanofi.
While implementing AI, the company discovered what Bushman called "technical debt we inherited.” The debt wasn’t from negligence, but from years of local teams working in silos. Sanofi's audit turned up 23,000 conditional rules standing in for proper configuration, 50,000 cost centers that no one actually used, and more than 10,000 empty job profiles.
This kind of localized knowledge can be one of the richest assets in a company.. But it doesn’t scale: employees learn which workaround keeps a broken process running and that knowledge never becomes a data model an AI agent can actually use.
"AI is only going to respond to what you feed it and if we're meeting at 5,900 different answers to the same thing, it's going to be very confused all the time,” Bushen said.
Sanofi's fix required revisiting job architecture, compensation plans, benefits administration—and roughly 380 downstream integrations hanging off its core HR platform. It’s the kind of foundational, unglamorous work that rarely makes it into an AI keynote demo.
“AI is only going to respond to what you feed it and if we're meeting at 5,900 different answers to the same thing, it's going to be very confused all the time.”
—Danielle Bushen
Merck faced a different problem. With a looming patent cliff and a mandate to reallocate funds to support the project, the company's HR leaders dug into what was actually driving volume at its HR service centers.
They found two out of every three calls were not requests to do something, but employees asking where to go to find what they need.
"The content was there, the system worked as it should, but nobody knew where to go to find these things,” said Eric Ervin, Merck’s head of HR digital solutions and strategy. Merck had built roughly 50 different entry points into HR, which left employees struggling to find anything.
That insight reframed Merck's entire AI roadmap. Rather than starting with the most sophisticated agent they could build, Merck prioritized a self-service agent aimed squarely at navigation and low-complexity tasks.
It paid off immediately. In an early demo to senior leadership, an AI-guided path to start a routine task was enough to convince the room. This early win was just the first step. Ervin explained that the team at Merck are not just layering agents onto processes as is. Instead, they’re asking: “If we are running zero-based, if we are running AI-first, how should these things actually operate in the future?"
“If we are running zero-based, if we are running AI-first, how should these things actually operate in the future?"—Eric Ervin, Head of HR Digital Solutions & Strategy, Merck
Report
Leaders at Sanofi and Merck learned three clear lessons that others can build on:
1. Audit before you automate. Count conditional rules, empty cost centers, orphaned job profiles, and redundant entry points. Teams cannot fix broken workflows without measuring the mess first. Executives often assume the problem stays small, but the hidden debt runs far deeper.
2. Fund the groundwork. Shiny new demos look great in board meetings, but unglamorous groundwork makes the software actually run. Leaders must back this invisible labor before rolling out any new tools.
3. Target the most common headaches first. Merck started where staff lost the most hours, which meant fixing simple navigational issues. That move built trust across the company and freed up time so workers could tackle tougher overhauls later.
Modern tools cannot rescue broken systems on their own. Skipping the dirty cleanup only burns more cash.
Clean house first, so enterprise AI agents tackle real jobs instead of copying messy habits.
A gap between effort and impact is the real ROI problem. Discover the solution.