CEOs want AI fast, but leaders who deliver it say fundamentals decide the pace. At COLLIDE 2024, panelists from Elder Research, Norfolk Southern, Equifax and Baptist Health covered model drift, AI governance councils and talent, plus results from AI train inspections and ambient listening.
Key takeaways
- Start with the business problem. Amouie said the answer might be generative AI, traditional AI or a dashboard, not a chatbot by default.
- Production is where trust is lost. Model and data drift can undo a successful prototype, so MLOps and DataOps matter.
- Governance councils review AI end to end. Baptist Health's council approves ideas and vendor products, reviews pilots for bias and audits production systems.
- Talent is a mix of strategies. Interesting problems, modern tools, upskilling and consultants all play a role.
- Real results exist. Norfolk Southern runs about 60 algorithms on trackside camera images to flag rail car defects within minutes.
Generative AI has CEOs asking how fast their companies can move, but the leaders who have to deliver say speed depends on fundamentals: a clear business problem, good data, governance and talent. That was the message of "From Hype to Reality: Practical Implications of AI Transformation," a panel at COLLIDE Data & AI Conference 2024 in Atlanta, in a session sponsored by Elder Research.
The panelists were Rick Hinton, Vice President of Analytics Strategy & Transformation at Elder Research; Mabby Amouie, AVP Enterprise Data & Analytics at Norfolk Southern; Vijay Antonyraj, Vice President, Consumer Business Insights & Analytics at Equifax; and Rosalia Ungaraza, AVP, Artificial Intelligence at Baptist Health, who leads the health system's AI program.
Where does the AI hype diverge from reality?
Antonyraj said the first gap is data quality. Clean, labeled, available data is the foundation for any AI implementation, and at Equifax ethics, privacy and compliance limits apply from pilot to production. Ungaraza said many people misunderstand what AI can and cannot do, and she recommended AI literacy as the fix.
Hinton said Elder Research's clients range from complete skeptics in one business unit to enthusiasts in another. Generative AI is a general-purpose technology, he said, so value comes from integrating it into specific tasks within business processes. "This is the first time in my career I've ever seen the CEOs of organizations saying, how fast can we move?" he said.
Amouie said he often gets asked where a chatbot could be embedded in the business. His response is to ask whether "you have a specific business problem that you're trying to solve or you really want to jump on the bandwagon." The answer might be generative AI, traditional AI or a simple dashboard, he said.
Antonyraj added that companies should be ready to stop projects that don't work. "You need to have the courage to abandon it," he said.
What challenges come after a successful AI pilot?
Scaling a prototype brings new problems, Amouie said. Norfolk Southern has seen models succeed as prototypes and then degrade in production through model drift and data drift. "You lose your business partner's confidence very quickly if you put something in production that's good for 10 days," he said. His answer is skilled talent plus solid MLOps and DataOps practices.
Antonyraj said models must integrate with existing software and keep adapting as customer and market behavior shifts. Ungaraza said regulated industries such as healthcare must balance fast-changing AI rules with room to experiment. Hinton said teams should anticipate resistance, such as from a general counsel's office worried about company data leaving the firewall.
How do AI governance councils work?
Equifax has formed an AI compliance and ethics council that reviews use cases across business units to prevent bias and discrimination, Antonyraj said. He recommended that every company have one, even if a small company has limited budget.
Baptist Health's enterprise-wide AI governance council includes members from supply chain, finance, clinical and other functions, Ungaraza said. It approves AI ideas and vendor products that contain AI, reviews internal pilots for bias and fairness, and follows the NIST AI playbook. Once products are in production, the council audits them to decide whether they should be improved or decommissioned for failing to meet expected ROI.
Hinton said generative AI multiplies governance risk because new vendors and models are largely untested. He advised honest prioritization before the flood of ideas overwhelms risk committees.
How can companies compete for AI talent?
Amouie listed exciting problems, an inclusive environment, current technology that attracts graduates from schools such as Georgia Tech and GSU, and competitive pay. Ungaraza described upskilling employees from outside the AI team and bringing in consultants on short-term contracts while internal staff learn alongside them. Hinton said AI programs must address both "mindsets and skill sets," with learning built into everyday work.
What AI use cases are delivering results?
Amouie described Norfolk Southern's digital train inspections. Traditionally, two inspectors walk each side of a two- to three-mile train, checking each rail car for about 350 possible issues. Trackside cameras now capture 360-degree images of trains passing at 70 mph, about a thousand pictures per rail car. Around 60 algorithms in production flag defects within minutes and automatically generate work orders. He said the system is "turning the finders of those defects, our car inspectors, to actual fixers," and is being rolled out to inspect 95% of traffic annually.
Ungaraza described ambient listening at Baptist Health, where generative AI drafts doctors' notes from recorded patient conversations for the doctor to review. Adoption has been about 50/50, she said, because the notes aren't always as accurate as doctors want. Antonyraj cited fraud and risk monitoring at Equifax.
Hinton described a Microsoft Copilot program at a consumer packaged goods client with about 300 users. Marketing showed the highest adoption, with brand managers using it for creative briefs and ad copy.
See more AI transformation stories at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center; registration is coming soon. Data Science Connect's free live webinars continue the conversation year-round.
Frequently asked questions
Why do AI projects fail after a successful pilot?
Panelists cited model and data drift, integration with existing systems and changing customer behavior. Amouie said a model that works for only 10 days quickly costs business trust.
What does an AI governance council do?
At Baptist Health, the council approves AI ideas and vendor products, reviews pilots for bias and fairness using the NIST AI playbook, and audits production AI to improve or decommission it.
How does Norfolk Southern use AI to inspect trains?
Trackside cameras capture 360-degree images of trains moving at 70 mph, and about 60 algorithms flag defects and automatically generate work orders for car inspectors.
How can companies attract AI talent?
Offer exciting problems, an inclusive environment, current technology and fair pay, and upskill existing employees, the panel said.
Watch: From Hype to Reality: Practical Implications of AI Transformation, brought to you by Elder Research on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
