Why Data People Make the Best AI Engineers
Getting AI agents to do meaningful work inside a real company is mostly a data problem. Here's why 15 years of pipelines, models and governance turned out to be the perfect training ground.
If you asked a younger version of me what I wanted to do, a career in data would not have made the list. And AI, at least the kind we work with today, wasn't even a thing. Like a lot of people in this field, I fell into it sideways. What kept me here was simple: technology is one of the best tools we have for helping people solve problems.
For more than 15 years that meant data architecture, analytics products and cloud platforms. Lately it means AI agents. The surprise is how little of the job actually changed.
Agents are only as good as the data they can reach
A demo agent is easy to build. An agent that gives a CFO the right margin number, for the right business unit, as of the right date, with the right permissions, is a very different animal. That second agent depends on:
- A governed semantic layer. "Revenue" has to mean the same thing in the agent as it does in the board deck.
- Lineage. When the agent cites a number, someone has to be able to trace it back to the source system and the warehouse.
- Access control. Row-level security doesn't go away because the interface is now a chat box.
- History. Agents that reason about trends need slowly changing dimensions, snapshots and point-in-time correctness.
None of this is new to anyone who has built a data warehouse. It's the same work with a new consumer.
Tools are just well-designed data products
When you give an LLM tools, through function calling or the Model Context Protocol, you're publishing an API for a very literal, very fast and occasionally overconfident analyst. The best tool designs I've seen look a lot like good data products:
- Narrow, well-named endpoints instead of one giant "run SQL" escape hatch.
- Clear contracts: inputs, outputs and what "null" means.
- Sensible defaults and guardrails, so the cheapest path is also the safest one.
Back in my consulting days I wrote about five red-flag Tableau practices, like 50-page custom SQL scripts and Excel files on SharePoint. Every one of those anti-patterns has an agentic equivalent, and the fix is the same: move the logic upstream into something trusted and governed.
Evaluation is just testing with a better vocabulary
Data teams have always lived with the question "is this number right?" We built reconciliation checks, row-count assertions and dashboards that watch other dashboards. Evals for AI agents are the same instinct: define what good looks like, build a golden dataset, measure every change against it.
The teams that will win with AI aren't the ones with the flashiest models. They're the ones who already know how to make data trustworthy.
What I'd tell a data engineer today
Your skills transfer. Learn how LLMs call tools, get hands-on with MCP, and build one small agent against data you already understand deeply. You'll find the hard parts are the ones you've been solving for years.
More soon. I'll be writing here about agentic engineering, AI-ready data and the messy, rewarding middle where enterprise data meets AI.