Beyond the AI hype: Turning potential into results

Artificial intelligence has become the defining buzzword of modern supply chain transformation. Yet for all the excitement, the question remains: How do you turn AI’s potential into measurable results?

At the CSCMP EDGE conference, Shanton Wilcox of NTT Data moderated a lively discussion with Vishwa Ram, vice president of data science at Penske Logistics, and Tim Long, industry principal for manufacturing and logistics at Snowflake, exploring what it means to get “beyond the hype.”

AI is a living system

For Penske Logistics, AI is not a one-time deployment, it’s an evolving ecosystem that must adapt to shifting business conditions.

“The value of AI sometimes does erode over time,” Ram explained. “There wasn’t a sudden failure of the model, but there was just a slow liquidation of results over time. That really highlights the importance of continuous monitoring, retraining, and governance because model drift creeps in and if you’re not tracking its performance against business metrics, it’s almost always too late.”

Ram described Penske’s “driver retention” model, which once significantly reduced overtime, but gradually lost effectiveness as the business changed. The experience reinforced his belief that companies must treat AI as a living system, ‘not a set-it-and-forget-it solution.’

The data foundation comes first

Snowflake’s Long echoed that sentiment, emphasizing that data, not algorithms, determines AI’s success.

“At Snowflake we believe the most important part of an AI strategy is your data strategy,” he said. “The best results from AI are those that come from giving it the right information to answer the question you’re asking.”

He recalled one customer whose data columns combined design and model numbers into a single field. It is an easy mistake for humans to interpret but one that “made it impossible for the AI to tell them apart.” The result: incorrect insights and eroded trust.

“It’s really important to give AI every advantage you can,” Long said, “by investing in a strong data foundation.”

Beyond model accuracy: Measuring business value

One of the biggest misconceptions about AI, Ram noted, is equating technical accuracy with business success.

“Your model can be 99% accurate yet generate zero business value,” he said. In one Penske warehouse model, “lab metrics looked fantastic, but we didn’t actually see defects go down until we uncovered the process gap and fixed it alongside.”

That lesson, he added, shifted the company’s mindset: “Noisy data creates noisy AI.” Instead of chasing ever more complex models, Penske invested in a modern data stack built on Snowflake to ensure clean, consistent, and trusted data sets.

Governance: From bureaucracy to enablement

While the word “governance” often sparks resistance, Long argued that it’s the backbone of scalability.

“Poor data governance,” he said, “is when everybody does their work out of Excel.”

True governance balances protection and enablement, empowering employees to use data confidently without compromising compliance or intellectual property.

Ram agreed. “All the excitement is with AI—it’s the dessert. Data governance is the broccoli. It’s not flashy, but it’s absolutely essential,” he said.

He outlined Penske’s “Manage, Monitor, Mediate” approach: cataloging data assets, monitoring quality with business rules and machine learning, and mediating errors by identifying their source.

“When people trust the data,” he said, “they’re going to use it and that’s how you go from shiny AI experiments to creating impact.”

Fit AI to the workflow, Not the other way around

Ram shared a vivid example of integrating AI seamlessly into operations: a multimodal model that estimated trailer utilization and load stability from images. The key wasn’t the model’s sophistication, it was how it fit the workflow.

“AI doesn’t succeed because it’s clever,” he said. “It succeeds because people trust it, because it fits naturally with the way people work, and it’s aligned to the outcomes the business actually cares about.”

At Penske, that meant embedding AI outputs directly into tools like Tableau dashboards, rather than introducing new systems.

“You have to figure out how to interject it into existing processes,” he said. “The goal is to make consuming AI insights as simple as using your iPhone.”

Change management and trust

Long reinforced that the hardest part of AI adoption isn’t technical, it’s human.

“Building the models is the easy part,” he said. “The hard part is change management. Somebody’s job will be different, and you need to think about how to help them move from their current methods to better ones.”

To accelerate adoption, Ram said, business users must be part of development from the start.

“The fastest way to move AI is to those who are going to actually use it,” he noted. “When you bring business teams into the process, the model becomes more credible, more relevant and ultimately, more widely adopted.”

The next phase: Agentic AI

Looking ahead, Long said the future lies in Agentic AI—autonomous digital assistants that handle repeatable decisions.

“We’re starting to see these agents come online across our customer base,” he said. “If robotics automated repeatable tasks, AI will automate repeatable decisions.”

Ram added a closing thought that summed up the panel’s message:

“Technology moves fast, but adoption moves at the speed of trust. Invest in data, governance, and alignment with business outcomes and if you build trust, results will follow.”

SC
MR