A new report from dbt reveals a telling split: 72% of analytics teams are using AI to help with coding, but only 24% have applied it to managing data pipelines. The gap comes down to a simple reality — AI coding is a solo activity. A developer can fire up Claude Code on their laptop and start generating dbt models in minutes. But building an AI system that monitors pipeline failures or triages errors requires the whole team to agree on what that looks like, how it should behave, and who owns the output. That coordination takes time, approvals, and a shared vision that most orgs haven't yet built.

The good news is your existing data stack doesn't need to be rebuilt. If you're running an ELT pattern with a warehouse, dbt, and a BI tool, you can layer AI on top just like you layered dbt on your warehouse. Start with the highest-value, lowest-risk pain point. Code review is a natural entry point because there's already a human in the loop. Set up a pilot using GitHub Actions, pick an AI model your company already has a security agreement with, and write a skill file that spells out your coding standards and business context. Let it run on a small repo first. Have your most skeptical senior engineer review the output — they'll catch things the enthusiasts miss.

The real work comes after the pilot. AI systems degrade over time as models change, costs shift, and context windows evolve. You need evaluation loops — automated checks that run a known input through your AI workflow on a regular cadence and flag when results drift. This isn't new; it's the same discipline ML teams have used for years. The difference is that now every data team needs to build it. Leadership needs to understand that AI won't 10x your velocity overnight. The efficiency gains will initially be eaten up by the work required to make the system reliable. But if you treat AI as an organizational capability rather than a personal productivity hack, you'll build something that actually lasts.

Source: dbt Labs Blog