Coming next: The Second Factory →

Back to Library
Data Foundations⚠️ Pitfall

Seven Versions of the Truth (When Nobody Can Say Which Data Is Good)

The Problem

Someone needs survey data, so you search the warehouse. Seven tables come back. A raw dump that still has test accounts in it. A monthly file somebody uploads by hand. A quarterly feed from a vendor on a different scale. A "cleaned" copy nobody has touched in four months. A v2 that duplicates the monthly with new column names. A legacy merge with broken joins. A daily feed that might be fine. You ask the team which one is right. "Depends what you need." Nobody can say which is good, which is better or which is best, and none of them joins cleanly to the customer table. So every analyst picks whichever one loads first, two reports disagree, and the meeting turns into an argument about whose number is right. Then you check another domain and find the same thing.

The Principle

Rate your data, promote one, retire the rest. Every domain needs one data product people can trust without asking, and a rating that says how far to trust it. Score every data product from zero to four stars, like a restaurant guide. Zero: it exists. One: documented, so you can understand it without tribal knowledge. Two: tested, so you know when it breaks. Three: governed, with an owner, and it joins to the master record with standard keys. Four: reference grade, checked against business outcomes by whoever owns the number, the way Finance signs off on revenue. Three stars is the line. Below it, data is fine inside its own domain but not safe to combine. At three and above, you can join survey to customer to orders and trust the answer. Four stars is what goes to the board and what you train models on. Start from the query logs, not a form. Logs tell you what people use. They don't tell you what's good; that part you earn. Then count how many domains have a product at three stars or more, and watch that number every month. The same map that helps a new analyst find the right data in an afternoon tells an AI model where its context lives.

🎯

3 Steps to Take Action

  1. 1

    Pull the logs for one domain

    Take the domain people argue about most. Pull ninety days of query logs: which tables get read, how often, by how many teams, joined to what. The most-used product is your candidate. A table nobody has read in six months goes on the retire list.

  2. 2

    Rate every product in one meeting

    Put the person who builds the data and the person who uses it in the same room and score every product from zero to four stars. Agree on what three stars means before you start: an owner, tests, documentation, and a clean join to the master record with standard keys.

  3. 3

    Promote one, retire the rest, report coverage

    Invest in the best candidate until it reaches three stars, then give the others a retirement date. Report one number every month: how many domains have a product at three stars or more. Anything below three that goes into a board report carries its rating next to it.

💡 Inspired By

Chapter 8 of The Auction Block, "The Map We Didn't Know We Needed."

This is what Analytics Pitfalls & Principles is about: people helping people.

In Practice

At ClassicMotors, a request for survey data turned up seven survey tables. Marcus, the analyst closest to them, could explain each one and still couldn't say which was right. None joined cleanly to the seller master: one used email addresses, another the vendor's own IDs. Jin didn't send around a form asking teams what they trusted. He pulled ninety days of query logs. The daily seller-feedback table had been queried more than two thousand times by four teams, almost always joined to sellers. The monthly upload had one regular user: Marcus, the morning after each upload. The v2 hadn't been touched in six months. The legacy merge was read every night by a job that fed nothing. Then they rated all seven. None scored above two stars. They promoted the seller-feedback table to three, standardized its seller key, gave it an owner and retired the other six. Over six weeks they ran the same audit across all twelve domains. Only about 10% of the estate reached three stars, but those products covered nine of the twelve domains. When Jack, the CEO, saw the map with its stars, he asked, "How did I not know our data was this simple?" Two months later his board asked how AI-ready the data was. He answered with one slide: nine of twelve domains at three stars or above.

Take it to work

Give this card to your AI

A skill is a file your AI keeps. Add this one and it knows Seven Versions of the Truth whenever your work runs into it.

Then ask it: “Does my plan fall into Seven Versions of the Truth?”

How to add it
  • Claude (a paid plan): Customize, Skills, +, Upload a skill. Pick the file you downloaded.
  • ChatGPT or Gemini: where your plan has Skills, upload the same file.
  • Microsoft 365 Copilot Cowork: Plugins, Add, Upload, then pick the Cowork file.
  • An AI at work that won’t take a skill? Open this card as text and paste it in.

All the AI skills, including ones for the moment before an AI, data or BI strategy, a dashboard or a readout goes out.

More on Data Foundations

From the book

In the book, a new hire takes six months to become useful, mostly because of table names nobody can read. And there are five definitions of an active seller.

This card is one of 59 drawn from The Auction Block by Brian Cariveau — the story of an analytics leader with 90 days to fix what the cards describe.

Heard something like this at work? See the pitfalls behind what else you’ve heard this month.Heard at work: free, 1 minute, no account