Plenty of SMBs and midmarket companies have messy data. That doesn’t mean the business stinks. It just means they’ll have a harder time putting AI to work.
But maybe your records are immaculate. The software still needs to know how your business works. What does “available” mean? Which customer records belong together? Who can approve an exception? Your team knows things that aren’t written down anywhere.
Now that the AI revolution is well underway, we need to make more of that explicit. Data cleaning is part of it. So is building an ontology: a model of the things in your business, how they relate, and what can happen to them.
We tend to call the whole exercise sorting your shit out. Which, depending on how long you’ve been accumulating it, can be a huge fucking pain in the ass.
Think about your garage
You need an old connector cable. You’re pretty sure it’s in there somewhere. You rummage around for fifteen minutes, move a broken printer, open the wrong box twice, and find it. Fine. It was your mess. You remember making it.
Now send a robot in there.
“Get the cable” leaves a few things out. Which cable? For what? The one in the drawer is broken. The one that works is in a box labeled “Christmas,” for reasons that made sense when you put it there.
Maybe five years from now you buy a robot and it handles all of this perfectly. Great. I’ll happily retire the analogy. For now, it helps explain how much context you bring to a job that looks simple from the outside.
Now think about your company. Thousands of SKUs. Years of PDFs, spreadsheets, emails and records scattered across software that doesn’t all talk to each other. People know which version to use, which customer has two accounts, and who to call when the numbers look wrong. How much of that does the software know?
Giving an agent access to the files gets it into the garage. It still needs a way to work out what it’s looking at.
Think about the work behind a self-driving car. A camera image has to become something the system can act on: vehicles, pedestrians, lanes, movement. Waymo’s public datasets give a glimpse of that work, with millions of labeled objects alongside sensor data and maps.
Your business needs some of that same preparation. What are we looking at? What does it mean? What can the system do with it?
What does “available” mean?
Maybe your data isn’t messy at all. You’ve got proper systems. Somebody actually labels things. Congratulations, you’re ahead of my garage.
Your warehouse says you have a hundred shirts. Sales says you have fifty available. They can both be right. Forty are already promised to customers and another ten are damaged. One system counts what’s physically there. The other counts what you can sell.
Now ask an AI to accept an order for seventy.
The records can be perfectly clean. You still need to establish what “available” means for this decision. Can you ship fifty now and twenty later? Does the customer have to agree? Who’s allowed to make that promise?
That’s part of sorting your shit out too. Figuring out how the business works, then making enough of that explicit for the software to do the job.
Start with the nouns: the customer, the product, the order. Are both systems talking about the same shirt? Add the adjectives: available, reserved, damaged. Then the verbs: reserve the stock, approve the order, ship it. Each action has conditions. Reading an inventory report and promising a delivery are very different responsibilities.
And then there’s the stuff your team just knows. Maybe this customer always accepts a split shipment. The approval happens in a text message, so sales knows the arrangement and the software doesn’t. We need to decide with you how that approval gets recorded, when it applies, and when the system has to ask again.
This is why Palantir puts so much emphasis on its Ontology: a model of the business that connects its data, relationships, logic, actions and permissions. In our shirt example, that means connecting the stock to the orders already promised against it, defining what can still be sold, and deciding who can commit to delivering the rest later.
So what do we actually do?
We can start with an audit and diagnostic. Pick a workflow and show us how it gets done, including the workarounds. For the shirt order, we’d follow it from the customer’s request to the warehouse. Where do the stock figures come from? Where are reservations recorded? Who approves a split shipment? What happens when an order gets canceled? You get a map of the workflow, a prioritized repair list and a scope for the first build.
What we need to fix depends on what we find. Two systems might give the same product different names. A supplier might send quantities in cases while your warehouse counts individual shirts. We connect the records, reconcile the units and keep track of where each number came from. Your team helps us establish which figures to trust and who’s allowed to change them.
You’ll hear the term ETL: extract, transform, load. Say we pull a supplier’s delivery record, convert the cases you received into individual shirts and load those quantities into your inventory system. That’s ETL. The conversion needs to survive the next file too. Otherwise, we’ve cleaned up Tuesday and broken Wednesday. Data mining can help us spot recurring problems, such as the same supplier’s files producing the same error.
AI can help with the cleanup too. It can suggest matches, classify files and flag things that look wrong. When a record is ambiguous, we need a way to get an answer from someone who knows.
You don’t have to clean the whole company before anything useful happens. We start with enough of the garage to do the job you care about. For this order, that means the customers, products, stock and approvals. Then we test the awkward cases your team already knows about.
Ask for seventy shirts again. This time the system should check what’s available now and whether the remaining twenty can arrive when the customer needs them. It should follow your rules for a split shipment and get approval where it needs it. If the order goes ahead, it needs to record the promise and reserve the stock, so the next order can’t quietly sell the same fifty shirts again.
That’s AI prep. It’s a huge part of what we do.
Mulholland Data Cleaning Corporation. We sort your shit out.
Plumber photograph by ocean yamaha, CC BY 2.0. Resized and converted to WebP.
Technical background: IBM on data cleaning, ETL and data mining. Film still: Robert De Niro as Harry Tuttle in Brazil (1985), directed by Terry Gilliam. Cinematography by Roger Pratt. Frame via Film Grab.


