Your AI Doesn't Need a Data Dump — It Needs the Right Queries

Boring on Purpose · Part 2 of 3. Previous → The Best Infrastructure Is No Infrastructure. Next → The Most Expensive Line of Code Is a Hardcoded Path.

I had a weekly analytics report that was mostly automated — except for the part where it wasn’t. Every week I’d manually export a big CSV from our data warehouse, drop it in a folder, and then trigger an AI-assisted pipeline that read the file and wrote the report. The AI did the hard part. I was the flaky human step in the middle.

The obvious “fix” is to automate the export. But that just moves the fragile step around. The better question was why the AI needed a giant file at all.

Give the model a warehouse, not a dump

I connected my AI assistant directly to the data warehouse through MCP — the same open protocol I use to give my tools access to my notes. Now the assistant could query the source itself, live, instead of parsing whatever I’d exported by hand.

The tempting first design was: have the AI pull all the rows and figure it out. That doesn’t work, and it’s worth understanding why. These conversation records have big text fields; dragging hundreds of rows of them through a tool call is slow, expensive, and often just too much for the model to reason over well. A data dump isn’t context — it’s noise you’re paying to move.

So I restructured the whole report around targeted queries instead. Rather than one giant pull, the pipeline now runs about eight small, purposeful ones: aggregates for the headline metrics, filtered pulls for the handful of rated messages that need quoting, pattern-matched queries for the deflection analysis. Each query answers one question the report actually asks.

Same report. Same quality. Zero manual steps — and the AI reasons over clean, relevant slices instead of drowning in a CSV.

The principle

It’s easy to think of an AI assistant as something you feed. You have data, so you hand it all the data. But a model connected to a live source doesn’t need to be fed — it needs to ask good questions. Aggregate first, pull detail only where the narrative genuinely needs it, and let the warehouse do what warehouses are good at.

The shift is small but it changes the ergonomics completely: the report went from “reliable except for the human” to just reliable. And the design generalizes — anytime you’re about to export a dump to feed an AI, check whether you could hand it the query engine instead. Right questions beat big files almost every time.