Planned Isn't In Stock (When an Agent Reads Data Nobody Defined)
The Problem
A marketer builds an agent that finds owners of a particular car and emails them about parts. Everyone cheers. It's exactly the kind of builder you hoped for. Nobody notices where it gets its information. It's pointed at a shared folder, and inside is the product roadmap: every part the team plans to make, with target dates. The agent reads "tail-light lens, Q2" and concludes, perfectly reasonably, that you sell tail-light lenses. So it tells hundreds of customers they're in stock. Every person in the building knows the difference between planned and in stock. Nobody ever wrote it down, because nobody had to. An agent only knows what's been written down, and it acts on whatever it reads at full speed.
The Principle
Customer-facing agents read only from defined, trusted data. If a word matters to the business, write down what it means before an agent acts on it. Planned, in production, on order and in stock are four different things. People learn the difference in their first month. Agents learn it from a definition or not at all. So define the words that carry promises (in stock, available, delivered, eligible, active), put each definition next to the data it comes from, and give it an owner. Then draw a line. Anything that talks to a customer reads from certified data only, three stars or better. No roadmaps. No shared folders. Just the truth.
4 Steps to Take Action
- 1
List the words that carry promises
In stock, available, delivery date, refund window, eligible. Write a plain-English definition of each, pointing at the data it comes from, with an owner.
- 2
Draw the three-star line
Anything that talks to a customer reads from certified data only. No shared folders, no roadmaps, no exceptions on request.
- 3
Ask what every agent reads
For each customer-facing agent, list its sources. Anything you can't name is a promise waiting to go wrong.
- 4
Correct it honestly
When one does go wrong, tell customers the truth, short and in plain English, and offer them a place in line.
💡 Inspired By
Chapters 8 and 9 of The Second Factory. And the discovery that nothing makes a team fix its data faster than an AI politely refusing to use it.
This is what Analytics Pitfalls & Principles is about: people helping people.
In Practice
Lena Brooks, a campaign manager at ClassicMotors, built her first agent in one of the company's AI sessions. It found owners of a particular model in the auction data, wrote each of them a personal email and sent it. She showed it off in the company's AI channel. It got forty reactions. She'd pointed it at a shared folder called Parts Launch, where the parts team kept its roadmap. The agent read "1967 Shelby GT500 tail-light lens, Q2" and told nine hundred and fourteen Shelby owners the lenses were in stock. By nine that night, a hundred and forty-one had replied and twenty-three had tried to order. One wrote, in capital letters, I HAVE BEEN LOOKING FOR THESE FOR ELEVEN YEARS. "Nobody told it the difference between planned and real," Lena kept saying. "Nobody told it because nobody's told anyone," Nora said. At the board meeting the next day, the lead independent director mentioned she'd had the email. She owned a 1967 Shelby.
