In October, I had the opportunity of representing Bottle Rocket and WPP Enterprise Solutions at the Databricks Data + AI World Tour in Dallas. The conversations I had in the hallways were more revealing than the sessions. Several people I talked to had already tried to bring AI into their day-to-day work. Some got a pilot running, then watched it stall once it met real data. Others hit walls on governance, or found that the same customer information was scattered across three systems that didn’t agree with each other. At different levels and in different capacities, eventually those teams realized they needed to take a step back and fix what was underneath the AI work.
That’s not the story most of us see online. The highlight reel shows what AI can do, but it rarely shows the groundwork that made it possible, or the teams that had to stop, pivot or reassess, and start again.
The highlight reel makes AI look like a simple exchange: ask a question, get an answer. Chat interfaces have trained us to expect that. Behind an enterprise AI use case, though, there’s a lot more: the model needs far more data than most teams expect, and that data has to be organized in a way that people can find and trust. Otherwise the use case never gets past the first demo.
So there’s no paved road to AI value. Every company, and often every use case within that company, has to chart its own path: what the data is, where it lives, who owns it, and what the business actually needs from it. None of that is glamorous, and none of it shows up in a product demo. It’s the grunt work, and it’s the part most people skip.
The Databricks keynote featured two large customers, American Airlines and AT&T, and both are still building out their data platforms and governance. AT&T started with a single data science team, then expanded the platform across more than 60 business units. According to Databricks’ own case study, the move cut its infrastructure footprint by about 40%. Its fraud team also used the platform to replace a rules-based system with more than 100 machine learning models, which reportedly reduced fraud attacks by 80%. American Airlines’ public material focuses on governance and data cataloging, the kind of unglamorous work that makes later AI projects possible.
A session on Toyota’s data work made the same point from the other side. Before the team built any AI, it asked whether its data was ready. Its speakers said they had to reverse one major decision along the way, which is a useful reminder that even good foundations need revisiting. The lesson isn’t that every company needs the same architecture - rather, it’s that each company has to check its own readiness first.
One of the sessions I attended was about streaming, and it’s the one closest to my day job. As an enterprise architect, I’m often weighing technical trade-offs. Take Databricks’ new Zerobus Ingest, for example. It’s meant to remove a common piece of the architecture: the message bus that sits between the systems generating large volumes of data and the lakehouse where that data lands. Fewer components means less to build and maintain, which is a real gain. But the trade-offs don’t disappear. Someone still has to decide how the data is structured, who can access it, and who owns it when something breaks. Those are the decisions teams tend to underestimate, and a simpler pipeline makes them easier to skip.
If you recognize your organization in this, you’re not behind. Most teams that stalled on AI weren’t short on ambition. They were missing a clear picture of what their data needed before the AI could work. Figuring that out is usually the first useful step, and it doesn’t require a new platform or a finished strategy.
If you want to talk through where your own data stands, book a discovery call with us. We’ll look at your current architecture, the use cases you’re considering, and what would need to be in place before those use cases can run in production.