Companies Winning with AI Built the Data Layer First

Every enterprise wants to be AI-driven. Fewer are willing to do the unglamorous work in the data layer. The organizations pulling ahead first create a strong data foundation and build intelligence on top of something they actually trust.

Trinity Industries is one of North America's largest railcar manufacturers and lessors, managing a leased fleet of over 141,000 railcars valued at around $8.5 billion. Moving 900+ commodities, the company operates at the intersection of heavy industry and financial services. Trinity runs its unified data and AI platform on Databricks, having migrated 95% of its enterprise data to a single lakehouse architecture.

Stephen Ecker is the Chief Data Officer at Trinity Industries, where he has spent 13 years and founded the company's analytics function. He built the team from a group of interns into a strategic capability that has driven over $100 million in measurable business impact.

Throughout our conversation, Stephen returned to a single conviction: the data layer is the strategy. Not the model, not the agent, not the dashboard. The foundation.

The cost of fragmentation

Aly McGue: Enterprise leaders often weigh the cost of fully transforming their infrastructure against the cost of not modernizing. How did you approach this, and why was data fragmentation ultimately so costly?

Stephen Ecker: It wasn't just an IT problem. It was a strategic ceiling for us. We had workloads bouncing between Azure and AWS, back to on-prem. Every model we deployed had its own serving setup. Nothing was standardized. We had an on-premises SQL warehouse where you'd run a query overnight on car location data, come back the next morning, realize you'd made a mistake, and have to run it again the next night. That's two days to answer one question.

But the bigger cost was analytics sprawl. We started with dashboards because nobody had access to any data, and they were wildly popular. But over time, a three-sheet dashboard would become a 40-sheet dashboard, each with its own transformations baked in. We calculated that we had almost 600 distinct measures across the business. A lot of those started from the same data source but had their own filters, their own lens. And then there was the knowledge silo. An analyst would spend two days on a piece of work, and six months later, someone else would start the same analysis from scratch. At one point, I felt like my biggest value was just having been here 13 years and knowing who had already done what. 

The "which number is right" debate

Aly: Without a single data layer, organizations often face the 'which number is right?' dilemma, where data from different departments doesn't match. How did this lack of a 'single source of truth' impact your leadership's trust in the data they were seeing?

Stephen: It was constant. Someone would show up with a number, and then it took an expert to dig into the code and say, ‘No, that number has these filters applied because that's what a specific person wanted three years ago.’ Even when we tried putting caveats and technical writing inside the dashboards, it didn't work. People don't read footnotes. They just grab a number and run with it.

We were logging 11,000 hours a month in these dashboards. And we kept trying to consolidate them, but we were never really consolidating anything because the demand for more dashboard scope never stopped. So during the migration, we made a hard call. We went to Medallion architecture, moved all transformations back upstream, and started scrapping legacy dashboards. You shouldn't have 600 measures, even in a multi-billion-dollar business. We needed the core measures and then an avenue for people to do their own analysis on top of that.

Unlocking AI through consolidation

Aly: How has consolidating your platform unlocked both better analytics and advanced AI models in a way that wasn't possible before?

Stephen: The gen AI angle is a big one. Unstructured data, things like emails, suddenly became really important. The other thing consolidation gave us is access to models without the overhead. We don't have to debate setting up a separate API to OpenAI or go through legal and architectural reviews every time we want to try something. We have all the protections provided by Databricks, and we can access the models we need under a single secure umbrella. That flexibility to experiment without a procurement process every time is huge for us.

We also now have agents interacting with upwards of a billion dollars in our manufacturing supply chain procurement. They're reaching out to vendors via email, synthesizing where inventory sits within the purchase order process, following up automatically. We saw an immediate 15% increase in on-time material delivery. When you think about every $10 million of working capital improvement being roughly $1 million to the bottom line, that adds up quickly.

Real-time intelligence at scale

Aly: Where have you seen real-time insights make the biggest strategic impact on your operations, and what was the architectural challenge in delivering that reliability and intelligence?

Stephen: Our ETA prediction model. That's our most technical challenge. Railcars in North America are tracked by AEI tag readers, basically reflectors on the side of the car that ping posts roughly every 10 miles. So you know a car is in Dallas, but not where in Dallas. GPS gives you more precision, but it's messy. Around 20% of industry data is misreported. GPS drifts. 

We had to build a real-time cleaning algorithm and a traversal-smoothing process that snaps GPS readings to the correct track by analyzing recent travel history. All that streaming data is unified into a single architecture, transformed, and then fed to an AI model that updates ETAs within seconds. Our model is now 50% more accurate than the industry's own ETAs, and we don't even control the locomotives.

The analyst bottleneck disappears.

Aly: One of the biggest hurdles for leadership is the lag time between asking a question and getting a data-backed answer. How has Databricks Genie’s natural language interface helped your team bypass the traditional 'analyst queue'?

Stephen: The first adopters of Genie weren't the executives, actually. It was my own analyst team. They were doing repeat operational work, fielding stakeholder questions and spending a day or two on analysis. Once they started using Genie rooms, they could get a clearer, more concise answer in 30 minutes. That was the signal for us.

From there, it spread. Our CFO is now asking questions about financial planning data in Genie rooms. Our CEO, who was a CTO at Caterpillar, is all in. We built a customer 360 application that pulls data from 9 domains and synthesizes customer summaries. Salespeople who never touched a dashboard are using it because it's just that easy to go deep. We're up to over a thousand questions a month, and we're re-architecting our entire BI layer around this approach.

From requesting data to conversing with it

Aly: How does providing a conversational analytics experience to non-technical business users shift your organizational culture from "requesting data" to "conversing with data"?

Stephen: Curiosity. That's the honest answer for what's still hard. Everyone likes the low-hanging fruit. They can get an answer, pull a dataset and skip the dashboard navigation. But we want them to go deeper, realize they're now just as capable as analysts, and start asking the harder questions.

I remember a board-level measure we created years ago comparing maintenance costs across different shops in our lease fleet. It took us weeks. One of the first things I did with a Genie room was ask it to do the same analysis. It arrived at the same answer in five minutes, using the same methodology, and was even smart enough to flag low sample sizes as anomalous. That's a complex analysis we couldn't have dreamed of eight years ago. Now it takes three prompts. It's like, wow, that's really impressive.

We were smart enough to start early on the adoption side, too. We brought in Microsoft Copilot in the first couple of months, not because we thought it would make everyone more efficient overnight, but because we had to get people prompting. We had to get them thinking of an LLM as a person, not a search engine. So that two years later, we're not still teaching people how to ask a question. That early investment in prompt literacy is paying off now.

Advice for leaders starting this work

Aly: If you had one piece of advice for a C-level leader trying to future-proof their organization for AI, what would it be?

Stephen: Don't build AI on a broken foundation. The data layer is the strategy.

You can spin up POCs pretty quickly with the latest models. But the winner of all this is going to be whoever has the strongest foundations, whoever actually invested in the data layer. The temptation is to chase the exciting AI use case. You have to resist that. Do the legwork. Our migration was painful. It took close to a year, and then another six to eight months after that to shore everything up. But AI is only as good as the data it runs on. If you want to ground it in your own data, automate real workflows, and scale with confidence, it starts with the foundation. It doesn't mean you can't get some quick wins along the way. But if you truly want to accelerate the business, it's in the foundation.

Closing Thoughts

What stands out most from this conversation is how directly Stephen connects every AI win back to the same decision: fix the data layer first. The ETA model, the procurement agents, the shift to conversational analytics — none of it would have been possible without Trinity's commitment to a painful, year-long migration that most organizations try to skip.

Companies that will lead in enterprise AI are not the ones with the flashiest prototypes. They are the ones willing to do the structural work and then build intelligence on something they actually control. For this 90-year-old company, moving physical goods across a continent, that clarity is worth paying attention to.

To learn more about how to create an actionable roadmap for advancing your AI capabilities, download the Databricks AI Maturity Model.