Each enterprise needs to be AI-driven. Fewer are keen to do the unglamorous work within the information layer. The organizations pulling forward first create a robust information basis and construct intelligence on prime of one thing they really belief.
Trinity Industries is one among North America’s largest railcar producers and lessors, managing a leased fleet of over 141,000 railcars valued at round $8.5 billion. Transferring 900+ commodities, the corporate operates on the intersection of heavy trade and monetary companies. Trinity runs its unified information and AI platform on Databricks, having migrated 95% of its enterprise information to a single lakehouse structure.
Stephen Ecker is the Chief Information Officer at Trinity Industries, the place he has spent 13 years and based the corporate’s analytics operate. He constructed the group from a bunch of interns right into a strategic functionality that has pushed over $100 million in measurable enterprise affect.
All through our dialog, Stephen returned to a single conviction: the information layer is the technique. Not the mannequin, not the agent, not the dashboard. The inspiration.
The price of fragmentation
Aly McGue: Enterprise leaders typically weigh the price of totally remodeling their infrastructure towards the price of not modernizing. How did you method this, and why was information fragmentation in the end so expensive?
Stephen Ecker: It wasn’t simply an IT downside. It was a strategic ceiling for us. We had workloads bouncing between Azure and AWS, again to on-prem. Each mannequin we deployed had its personal serving setup. Nothing was standardized. We had an on-premises SQL warehouse the place you’d run a question in a single day on automotive location information, come again the following morning, understand you’d made a mistake, and should run it once more the following evening. That is two days to reply one query.
However the larger price was analytics sprawl. We began with dashboards as a result of no person had entry to any information, and so they have been wildly in style. However over time, a three-sheet dashboard would turn into a 40-sheet dashboard, every with its personal transformations baked in. We calculated that we had virtually 600 distinct measures throughout the enterprise. Quite a lot of these began from the identical information supply however had their very own filters, their very own lens. After which there was the data silo. An analyst would spend two days on a bit of labor, and 6 months later, another person would begin the identical evaluation from scratch. At one level, I felt like my largest worth was simply having been right here 13 years and figuring out who had already completed what.
The “which quantity is true” debate
Aly: With out a single information layer, organizations typically face the ‘which quantity is true?’ dilemma, the place information from totally different departments would not match. How did this lack of a ‘single supply of fact’ affect your management’s belief within the information they have been seeing?
Stephen: It was fixed. Somebody would present up with a quantity, after which it took an skilled to dig into the code and say, ‘No, that quantity has these filters utilized as a result of that is what a particular individual needed three years in the past.’ Even once we tried placing caveats and technical writing contained in the dashboards, it did not work. Individuals do not learn footnotes. They only seize a quantity and run with it.
We have been logging 11,000 hours a month in these dashboards. And we stored making an attempt to consolidate them, however we have been by no means actually consolidating something as a result of the demand for extra dashboard scope by no means stopped. So in the course of the migration, we made a tough name. We went to Medallion structure, moved all transformations again upstream, and began scrapping legacy dashboards. You should not have 600 measures, even in a multi-billion-dollar enterprise. We wanted the core measures after which an avenue for individuals to do their very own evaluation on prime of that.
Unlocking AI via consolidation
Aly: How has consolidating your platform unlocked each higher analytics and superior AI fashions in a approach that wasn’t doable earlier than?
Stephen: The gen AI angle is an enormous one. Unstructured information, issues like emails, abruptly turned actually necessary. The opposite factor consolidation gave us is entry to fashions with out the overhead. We do not have to debate organising a separate API to OpenAI or undergo authorized and architectural evaluations each time we need to attempt one thing. We’ve got all of the protections offered by Databricks, and we will entry the fashions we’d like underneath a single safe umbrella. That flexibility to experiment and not using a procurement course of each time is big for us.
We additionally now have brokers interacting with upwards of a billion {dollars} in our manufacturing provide chain procurement. They’re reaching out to distributors through electronic mail, synthesizing the place stock sits throughout the buy order course of, following up robotically. We noticed an instantaneous 15% enhance in on-time materials supply. When you consider each $10 million of working capital enchancment being roughly $1 million to the underside line, that provides up shortly.
Actual-time intelligence at scale
Aly: The place have you ever seen real-time insights make the largest strategic affect in your operations, and what was the architectural problem in delivering that reliability and intelligence?
Stephen: Our ETA prediction mannequin. That is our most technical problem. Railcars in North America are tracked by AEI tag readers, mainly reflectors on the aspect of the automotive that ping posts roughly each 10 miles. So you realize a automotive is in Dallas, however not the place in Dallas. GPS offers you extra precision, however it’s messy. Round 20% of trade information is misreported. GPS drifts.
We needed to construct a real-time cleansing algorithm and a traversal-smoothing course of that snaps GPS readings to the proper observe by analyzing current journey historical past. All that streaming information is unified right into a single structure, remodeled, after which fed to an AI mannequin that updates ETAs inside seconds. Our mannequin is now 50% extra correct than the trade’s personal ETAs, and we do not even management the locomotives.
The analyst bottleneck disappears.
Aly: One of many largest hurdles for management is the lag time between asking a query and getting a data-backed reply. How has Databricks Genie’s pure language interface helped your group bypass the standard ‘analyst queue’?
Stephen: The primary adopters of Genie weren’t the executives, truly. It was my very own analyst group. They have been doing repeat operational work, fielding stakeholder questions and spending a day or two on evaluation. As soon as they began utilizing Genie rooms, they might get a clearer, extra concise reply in half-hour. That was the sign for us.
From there, it unfold. Our CFO is now asking questions on monetary planning information in Genie rooms. Our CEO, who was a CTO at Caterpillar, is all in. We constructed a buyer 360 software that pulls information from 9 domains and synthesizes buyer summaries. Salespeople who by no means touched a dashboard are utilizing it as a result of it is simply that simple to go deep. We’re as much as over a thousand questions a month, and we’re re-architecting our total BI layer round this method.
From requesting information to conversing with it
Aly: How does offering a conversational analytics expertise to non-technical enterprise customers shift your organizational tradition from “requesting information” to “conversing with information”?
Stephen: Curiosity. That is the trustworthy reply for what’s nonetheless exhausting. Everybody likes the low-hanging fruit. They will get a solution, pull a dataset and skip the dashboard navigation. However we wish them to go deeper, understand they’re now simply as succesful as analysts, and begin asking the more durable questions.
I bear in mind a board-level measure we created years in the past evaluating upkeep prices throughout totally different outlets in our lease fleet. It took us weeks. One of many first issues I did with a Genie room was ask it to do the identical evaluation. It arrived on the similar reply in 5 minutes, utilizing the identical methodology, and was even good sufficient to flag low pattern sizes as anomalous. That is a posh evaluation we could not have dreamed of eight years in the past. Now it takes three prompts. It is like, wow, that is actually spectacular.
We have been good sufficient to begin early on the adoption aspect, too. We introduced in Microsoft Copilot within the first couple of months, not as a result of we thought it will make everybody extra environment friendly in a single day, however as a result of we needed to get individuals prompting. We needed to get them considering of an LLM as an individual, not a search engine. In order that two years later, we’re not nonetheless instructing individuals the way to ask a query. That early funding in immediate literacy is paying off now.
Recommendation for leaders beginning this work
Aly: If you happen to had one piece of recommendation for a C-level chief making an attempt to future-proof their group for AI, what would it not be?
Stephen: Do not construct AI on a damaged basis. The info layer is the technique.
You possibly can spin up POCs fairly shortly with the newest fashions. However the winner of all that is going to be whoever has the strongest foundations, whoever truly invested within the information layer. The temptation is to chase the thrilling AI use case. You will have to withstand that. Do the legwork. Our migration was painful. It took near a yr, after which one other six to eight months after that to shore every thing up. However AI is barely nearly as good as the information it runs on. If you wish to floor it in your individual information, automate actual workflows, and scale with confidence, it begins with the inspiration. It does not imply you may’t get some fast wins alongside the best way. However when you actually need to speed up the enterprise, it is within the basis.
Closing Ideas
What stands out most from this dialog is how instantly Stephen connects each AI win again to the identical choice: repair the information layer first. The ETA mannequin, the procurement brokers, the shift to conversational analytics — none of it will have been doable with out Trinity’s dedication to a painful, year-long migration that the majority organizations attempt to skip.
Corporations that can lead in enterprise AI aren’t those with the flashiest prototypes. They’re those keen to do the structural work after which construct intelligence on one thing they really management. For this 90-year-old firm, transferring bodily items throughout a continent, that readability is price being attentive to.
To study extra about the way to create an actionable roadmap for advancing your AI capabilities, obtain the Databricks AI Maturity Mannequin.