The closet nobody opened
When we had people coming for dinner, my mom made me clean my room. Not just the floor. The closet. The drawers. The pile behind the door.
It never made sense to me. No guest opens your closet. They came over to eat, not run a home inspection.
But that was the rule. Clean the parts people see. And, somehow, the parts they don't.
I think about that a lot now. Because data used to work exactly the same way.
We cleaned the room, not the closet
For most of my career, you cleaned what got presented. The dashboard. The board deck. The number in the QBR.
Behind that? A closet. Tables nobody remembered building. Columns named "final_v2_real". A view that only worked because one person knew not to touch it.
And it was fine. Nobody opened the closet. The guests stayed in the dining room, where everything was tidy.
Then AI started opening doors
I've written before about how the data team's job is shifting - from getting the number to owning the foundation underneath it. I keep coming back to it. This is the same idea. A memory just gave me a better way to say it.
AI doesn't stay in the dining room. It opens every door.
You point it at the warehouse and it reads all of it. The clean models and the junk drawer. The documented table and the one from 2021 nobody can explain. It doesn't know which is which. It reads all of it and answers anyway.
So the messy closet isn't hidden anymore. It's the first thing the guest opens.
Tribal knowledge was the mess we tolerated
Every data team has that person. Been there long enough to know how the whole thing actually works. "Don't use that table, the join's off. Use this one." "Ignore the revenue field on Mondays, the sync runs late."
That knowledge lived in their head. It never got written down because it never had to. You just asked them.
AI can't ask them. There's nobody to lean over to. It takes whatever field it finds, wrong one included. The tribal knowledge that quietly held the whole thing together is the exact thing AI can't see.
So it can't live in a head anymore. It has to live in the structure. In the names. In the docs. Somewhere real.
The job is keeping the house in order
This is the shift. The job isn't presenting clean data. It's keeping the whole house in order.
Knowing where everything goes. What everything is. What shape it's in. Then maintaining that - the structure, the guides, the rules - so it holds up as the business changes.
It's not glamorous. Nobody applauds a well-named table. There's no slide for it. But everything else falls over without it.
And when the house is actually in order, you can tell AI "go to the fourth drawer down on the right" and it comes back with exactly what you needed. Not because it's clever. Because you set it up. You know it's there. You put it there.
What a clean house actually means
But "keep it clean" means nothing until you get specific. So here's what specific looks like.
The dashboard graveyard can't exist anymore. For years we let it grow. Hundreds of half-built, abandoned, "I'll fix it later" dashboards nobody ever deleted. It was clutter, but it was harmless. Nobody wandered into the graveyard.
AI wanders into the graveyard. Hook an agent into your Hex repo and it'll happily pull a dead dashboard from two years ago, run its old revenue logic, and report the number like it's current. Nobody told it that one was abandoned. So now the graveyard is a liability. Certified dashboards stay lean and current, and the dead ones get deleted. Which, honestly, is what we should have been doing all along.
Your dbt project stays lean. Six models that all sort of calculate revenue means the agent picks one. Maybe the right one. A tight project with one clear model means it references the right data, because there's only one right answer to find.
Legacy code gets grouped and documented, not left in a pile. And your instructions - your README, your .md files - have to tell the AI what to ignore as much as what to use. "Don't touch this folder, it's deprecated" used to live in someone's head. Now it has to be written down, because there's nobody to say it out loud.
And you need tooling that alerts you when something drifts. A model breaks, a number jumps, a sync fails. You want to hear about it from your own system, not watch the AI hand the broken version to your CEO.
Which leads to the move nobody likes: delete things. The old trick was to shove the mess in a closet nobody opened. That stopped working. AI can't report on a dashboard you deleted, or legacy code that's actually gone. Hiding it where no one looks isn't a solution anymore. Throwing it out is.
We knew this already
None of this is new advice.
We always knew the dashboard graveyard was bad. We knew dbt should be lean. We knew the docs were stale and the legacy folder was a mess. It went on every "we should really clean this up" list, every quarter, and it never made the cut. Because it never had to. The mess was survivable. Someone always knew which dashboard to trust.
That's what changed. Not the advice. The stakes.
Tech debt used to be an issue. The kind of thing you'd get to eventually, after the roadmap stuff. Now it's a problem. A real one. The mess you could live with when humans were the only ones reading the data is the exact mess that makes AI unreliable.
And the order matters. Clean house, then trust. Not the other way around. You don't get to skip to the part where you lean on the AI, because the leaning only works once the house holds up. The trust is earned by the cleaning. The cleaning is the work we've been putting off for years.
Then you teach people to point
The last part is the one that gets skipped. A clean house doesn't help if everyone's still asking vague questions and grabbing the first answer.
"Go to the fourth drawer down on the right" only works if people know the drawers exist. And know to be specific about which one. That's a skill now - knowing what to ask, knowing what a trustworthy answer looks like, and knowing when to just open the drawer yourself.
Our job isn't only to keep the house clean. It's to teach the people walking through it how to find what they came for.
The closet is the job now
The old job was a clean room with a closet you hoped nobody opened.
The new job is a house where every drawer is clean, because every drawer is open.
Same thing my mom wanted, honestly. Tidy the parts people see. And the parts they don't.
Your house can look spotless. But AI opens the closet, and now the closet has to be clean too.