At a Glance
- Two years ago, a popular prediction was that AI would make the data warehouse irrelevant. The numbers now say the opposite.
- Every agent question is a query, every answer needs fresh data, and every definition needs to be agreed on. AI multiplies data work.
- Leaders who budget “AI” and “data” separately are underfunding the part that decides whether AI works.
The prediction that aged badly
You’ve probably heard some version of it: “Once LLMs can read everything, why maintain a data warehouse? Just point the model at the sources.”
It was a tempting idea. It was also wrong.
Look at Snowflake’s latest quarter. Product revenue grew 37%, the third consecutive quarter of acceleration, driven by both its core data platform and a meaningful step-up in AI revenue. AI isn’t replacing the platform. It’s pulling more work onto it.
Why AI makes data platforms busier
This is less mysterious than it sounds.
A dashboard gets opened a few times a day. An agent in Slack gets asked the same question fifty times, rephrased, drilled into and followed up. Each of those is a query. Put AI in front of your data and consumption goes up, not down.
Agents also raise the bar on data freshness. A report can live with last night’s batch load. An agent approving a discount or flagging a churn risk can’t. That’s why data platforms are quietly upgrading the pipes that move data, not just adding chatbots. Databricks’ newly GA automatic change data feed, for example, tracks row-level changes without anyone switching it on table by table.
And agents break on ambiguity. Three definitions of “active customer” is an annoyance in a dashboard. For an agent, it’s a wrong answer delivered with confidence.
Where this leaves leaders
We see the same pattern in almost every client conversation. The AI initiative has a sponsor, a budget line and a steering committee. The data platform sits in a separate budget, often treated as a cost to optimize down.
That split is backwards. The model is the part you can swap next quarter. The data foundation is the part that decides whether any model gives you a trustworthy answer.
Atgeir’s take:-
Budget AI and data together. If an AI use case doesn’t carry its share of data engineering (pipelines, freshness, definitions), its business case is incomplete. Most pilots that stall are data projects wearing an AI badge.
Plan for consumption, not just tokens. Most teams now track LLM spend closely. Far fewer have modeled what agents will do to warehouse compute. Put both on the same FinOps view before the bill forces the conversation.
Fund the unglamorous work first. Change data capture, metric definitions and data contracts don’t make good demos, but they set the ceiling on what your agents can do. Two weeks ago we argued the value of the whole stack now sits in how your business is encoded in data. This is the bill for that.
The data platform isn’t the thing AI replaces. It’s what AI runs on. Build it like you mean it.