Nearly every executive I talk to today feels the same urgency around AI and the same frustration. Despite months of experimentation, their organizations still can’t seem to create repeatable, operational value from it.

There’s no shortage of conversation about the usual suspects: model performance, cost, security, repeatability. But there’s a bigger issue that doesn’t see nearly enough conversation. Most enterprises aren’t struggling because the models are weak or too expensive. They’re struggling because their operational environments were never built to support AI at scale.

One of my favorite questions to ask leaders after they’ve decided what to build is: “Could an analyst do this today using the tools you already have?” More often than not, the answer is “yes, if they were willing to scour across five different systems to piece it together.” That’s not an AI problem. That’s the same data foundation problem that’s been holding back analytics programs for years.

The good news is that building that foundation doesn’t have to be the multi-year, million dollar project it once was. We’re standing up data platforms significantly faster than a few years ago and lightyears faster than a decade ago. And you don’t need a pristine, fully governed data warehouse to start getting real value out of AI. Raw data lakes will limit what you can trust and how far you can scale, but there’s a practical middle ground that gets you there without the full blown investment upfront.

The next problem organizations run into when they start moving experiments toward production is scale. Suddenly the spreadsheet that took an afternoon to put together needs to be manually refreshed everyday with new data. The process that worked for a few hundred records needs to handle millions. And every token you push through a model costs money, so doing it inefficiently at scale gets expensive fast.

The teams that get this right aren’t necessarily doing anything more sophisticated with AI. They’re just more disciplined about what they ask it to do. Language models are exceptional at reasoning, synthesizing and generating. They are not the right tool for filtering, aggregating or wrangling millions of rows of data. That’s what databases were built for. Query optimizers, parallel processing, partitioning, these aren’t new innovations. They’re decades of engineering designed to make exactly this kind of work fast and cheap.

The practical implication is straightforward. The more work you can do in your data platform before a record ever touches a model, the faster, cheaper and more reliable your AI workflows become. Teams that figure this out early stop thinking about AI as a replacement for their data infrastructure and start treating it as the last mile of a well engineered pipeline.

One of our most successful AI implementations came from a client that already had solid foundations in Snowflake. That head start mattered more than they expected. Their data was structured and in one place. Security was already handled by a platform their organization knew and trusted. And when it came time to scale, they used Snowflake’s Cortex features to run AI analysis through the SQL engine, processing in parallel across their existing infrastructure. What would have taken weeks of custom engineering to cobble together was running in production in a fraction of the time. The foundation they’d already built didn’t just make AI easier. It made it possible to deliver something that’s making a multimillion dollar impact on their business while costing in the hundreds of dollars.

The organizations winning with AI right now aren’t the ones with the biggest budgets or the most advanced models. They’re the ones that took the time to get their operational house in order, even imperfectly. You don’t need a perfect foundation to start. You need a good enough one, and the discipline to keep improving it as you go. The pace at which teams can stand up reliable data platforms today means that excuse is getting harder to make.

So now you have your foundation and your first AI agent in production. What’s next? Deciding what to build is perhaps the most valuable skill you can develop going forward. We’ll discuss that in next month’s article.

Share: