How to Fix Your Data Before You Automate
Automation does not fix bad data. It publishes it faster and with more confidence. Reconcile one number across your systems first, and the rest of the roadmap gets easier.
· 4 min read
Patrick Laughlin, Founder, Chief AI Officer
Naval Academy, class of 2011. Decorated naval aviator, 10+ years. President, 2023 Inc. 5000 company.

A chief executive asked me for one dashboard. Pipeline, profit and loss, receivables aging, commissions, inventory, royalties. One screen, current, so he could stop asking four people for four spreadsheets on a Monday.
A reasonable ask. The figures lived in four systems: the operations platform, the CRM, the accounting package and the online store. Each had its own idea of what an order was, and of the moment it counted.
We did not start building the dashboard. We started by reconciling one number.
The exercise
Pick the figure your executive team argues about most. For most companies it is bookings, revenue recognized, or open receivables.
Write down every system that holds a version of it. Not the two you think of. All of them, including the spreadsheet someone maintains privately, which is usually the one leadership actually trusts.
For each system, write three things: who enters the value, when it gets entered, and what rule decides it. That third one is where the arguments live. One system counts an order when it is entered, another when it ships, a third when the invoice posts. None of them is wrong. They are answering different questions, and nobody wrote that down.
Then take a four-week window and reconcile the systems against each other, by hand, once.
What that produced
When the sales dashboard went up at that company, we ran it beside the report the sales director had been assembling by hand. Over a four-week window the two totals agreed to within half of one percent.
I am not going to dress that up as a time saving. What it bought was permission. Nobody argues with the dashboard now, which means nobody rebuilds it by hand on Sunday night, which means the number in the meeting is the same number every week. Reporting that is trusted gets used. Reporting that is nearly right gets checked, and checking it costs more than producing it did.
An agent is only as good as the query under it
This is the part that gets skipped in the rush to put a model on top of everything.
A language model in front of four systems that disagree does not resolve the disagreement. It picks one and states it in a full sentence with no hedging, which is considerably worse than a spreadsheet that at least looks like it might be wrong.
Fix the disagreement first. Then the model is answering a question with one answer, and its confidence is earned.
Half of the answer is usually not AI
A company accruing royalties per licensor kept its rates in spreadsheets, one per person who needed them, so every rate change had to reach every copy.
The right build there is a rate table with one owner and a version history, and a scheduled job that applies it. There is no model in that. If something can be automated deterministically, automate it deterministically. Automation is repeatable and cheap and it gives the same answer twice. A model is a probability with good manners. Save it for the parts that need judgment, like reading an unstructured document or drafting a summary, and let plain code do the arithmetic.
Stop making new mess first
The fastest data cleanup I know does not start with the history. It starts with the intake.
If a request arrives as free text in an email, every downstream system has to guess at it forever, and you will pay for that guess in every workflow you build. If it arrives as a form with six fields, it does not. Changing the intake takes an afternoon and it stops the pile growing while you deal with the pile.
The same goes for the fields nobody validates. A dropdown with five options beats a text box you will be normalizing in two years.
The order of operations
Name the number. List every system that holds it. Name the owner of each. Reconcile one window by hand. Fix the intake. Then build.
Skip to the build and you get a system that produces a figure the executive team does not believe, and the day they stop believing it is the day it stops mattering how quickly it was produced.
I will not put a model in front of data that nobody owns. It would run. It would be wrong occasionally, and nobody in the building would be able to tell you which time.


