---
name: analytics-pitfalls-four-hundred-and-twelve-boxes
description: "\"Four Hundred and Twelve Boxes\": a pitfall card from Analytics Pitfalls & Principles. Use when a plan, request, dashboard, metric or AI project might be falling into it."
license: CC-BY-4.0
---

# Four Hundred and Twelve Boxes (Building the AI Ontology From the Top Down)

A pitfall card from Analytics Pitfalls & Principles, drawn from *The Second Factory*, Brian Cariveau's next novel (https://www.analyticspitfalls.com/second-factory).
https://www.analyticspitfalls.com/cards/79

## How to use this card

Use it when what someone is working on, or something they heard at work, looks like the problem below: a plan, a roadmap, a request, a dashboard, a metric, an AI project.

1. Name the card by its title and say in a sentence why it fits. Say the work looks like this pitfall; never tell someone they have it.
2. Give the principle in the card's own words. Don't rewrite it.
3. Take the steps one at a time against their actual situation: what is already done, what isn't, and which to do first. If you need a fact to do that, ask one question, not five.
4. Then push past the technical fix, if the work needs people to work differently (most does). The steps are the if-then part. Ask what has to change in how people work, what they have to learn or believe, or what they give up, for it to stick. That's the adaptive work (Ronald Heifetz's term), and it's usually the real work.
5. If the card doesn't fit, say so plainly. Don't stretch it.
6. Finish with the card's link: https://www.analyticspitfalls.com/cards/79

## The problem

Someone stands up with a deck. Every AI-native company has an ontology, a model of the business that agents can understand. The next slide is a diagram with four hundred boxes and so many lines it looks like a plate of spaghetti. Eighteen months, enterprise-wide, top down, with a consulting partner.

You've seen this dream before. Ten years ago it was a single source of truth. Then a semantic layer. Now an ontology. Same dream, new word, same ending: out of date before it's finished, drawn by people who've never answered a customer's call, and nobody uses it.

Meanwhile your agents are guessing. If a new hire would be confused by your data, the agent will be too. It'll just be confused confidently, at scale.

## The principle

Simple for humans is simple for agents. Build the map one trusted domain at a time, in words a new hire would understand.

An ontology is a map of the things your business cares about, what each one means and how they connect. A seller lists a vehicle. A part fits certain vehicles. You may already have most of it. If your data products are rated, owned and joined on the same keys, that's the backbone. All that's missing are the words.

Start with one question a real leader asks every week. Write a plain-English definition for every term it depends on, each pointing at the data it comes from. Let AI stand only on data you trust: three stars and up, on a scale of zero to four. Then test it like code, sit beside the leader while they ask, and pick the next domain.

## The steps

1. **Pick one question and one sponsor.** Not an enterprise model. One question a real leader asks every week, written down word for word.
2. **Rate the data it needs.** Zero to four stars, honestly. Three stars means it has an owner and joins reliably on your core keys, and that's the line for anything AI touches. If someone argues their data is three stars, test the join in front of them.
3. **Write the words.** A plain-English definition for every term, each pointing at its data. Read each one aloud and ask whether someone in their first week would understand it.
4. **Test it like code.** Twenty real questions with known answers. Fix the definitions, not the answers. Don't show anyone until it gets eighteen right and cites its sources.

## In practice

At ClassicMotors' weekly leadership meeting, the head of Enterprise Architecture put one word on his first slide: ONTOLOGY. His second was a diagram with four hundred and twelve boxes. Eighteen months, enterprise-wide, scoped with a consulting partner.

Jin Park, who ran the data factory floor, looked at the ceiling. "Ten years ago, everybody needed a single source of truth. A few years later, everybody needed a semantic layer. Now it's an ontology. New word. Same dream. Same plate of spaghetti."

That evening, Nora Berg asked the AI to explain an ontology three ways, the last as if she'd never heard the word and was suspicious of it. A map of the things a business cares about, what each one means and how they connect. She stopped walking in front of a neighbor's garage.

The next morning, she stood at the end of the catwalk in front of a wall she'd walked past a hundred times: the domain map Sarah Chen and Jin had built years earlier. Twelve domains, the data products under each, and a star rating, zero to four, next to every one. Eleven of the twelve domains were at three stars or better.

"Oh," Jin said, when he saw what she saw. "Oh, that's annoying. We did the hard part."

---

© 2026 Brian Cariveau, Analytics Pitfalls & Principles. Free to share and adapt under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/): keep this credit, "Brian Cariveau, Analytics Pitfalls & Principles (https://www.analyticspitfalls.com)", and say if you changed it.
The names Analytics Pitfalls & Principles, The Auction Block and The Second Factory are not part of the license. Don't use them for your own version, or to suggest it comes from us.
