Six hundred and seventy-two people were about to get an email from me, and I was thrilled about it.
I was thrilled. That was the problem. A number had arrived. The number was big. Big numbers feel like proof, and I had been looking for proof for about eighteen months.
Here is the setup. My boss mentioned, in passing, that several people had been asking whether there was going to be an AI newsletter. Not a survey, not a formal request, just the kind of hallway comment that lodges in your chest and stays there. People wanted the thing. After a year and a half of building an AI enablement program at a furniture company, of walking into rooms where the temperature ranged from polite to actively skeptical, somebody wanted the thing. Unprompted.
So I did what any reasonable person does now. I asked Copilot to help me build a distribution list, segmented by engagement, so I could start with the people most likely to actually read it.
It came back with a Tier 1. Six hundred and seventy-two names.
I did not think “that seems high.” I thought “that seems right.” I felt a small warm flush of vindication. Six hundred and seventy-two engaged users. That is a program. That is a slide. That is a number you say out loud in a leadership meeting with your voice very flat, because you know the number is doing the work for you.
I sat with it for maybe four minutes before something in the back of my skull started clearing its throat.
Six hundred and seventy-two was fifty-four percent of every licensed user at the company.
A tier that contains a majority of the population is not a tier. It is a mailing list wearing a lanyard.
What I found when I looked underneath was one of the more educational half hours of my year. Of those 672 people, 435 sat at exactly the threshold score. Not near it. Not clustered around it in a nice believable bell. On it. Precisely on it, all 435 of them, like a chorus line. That is not a distribution. That is a default value. Somewhere in the pipeline, a null had been quietly converted into a five, and 435 people who had never opened the tool in their lives had been promoted to Highly Engaged.
I raised the cutoff from 5 to 6. The list went from 672 to 208.
Then I went hunting for actual signal, the kind where a human being had demonstrably done something with a machine. Real usage. Repeat sessions. Evidence of a person on the other end.
Ten. Of the original 672, ten.
Also: 35 of the addresses had resolved to numeric employee-ID aliases. Things like 20112570@millerknoll.com. I would have been sending a thoughtfully written newsletter about AI literacy to a series of integers. Some of those integers may not correspond to anyone currently employed here. I did not investigate further. There are doors you do not open on a Tuesday.
And then the detail I have decided to just say out loud rather than quietly delete, because if I am going to teach this I have to be willing to be the exhibit.
I was on my own recipient list.
The system had evaluated me, the person who built the program, who runs the sessions, who wrote the curriculum, and scored me as a promising prospect for my own newsletter. Tier 1. Warm lead. Very engaged with AI. Somebody should really reach out to this Harvey guy, he seems interested.
I have thought about that a lot. The model was not lying to me. It was doing exactly what I asked, with the data it had, and it presented the result in a format that looked like analysis. Columns. Tiers. A score. The furniture of rigor without any of the load-bearing walls.
The dangerous output is never the obviously wrong one. If it had told me my Tier 1 was four million people, or that my top engaged user was a Labrador retriever, I would have laughed and started over. Obvious errors are free. They cost you nothing but a second attempt.
The expensive error is the one that tells you what you were already hoping was true, dressed in the visual language of evidence, arriving on a day when you needed good news. It does not trip any alarms because it is not surprising. It is confirming. And confirmation feels, in the body, almost exactly like understanding.
My program teaches verification as a core behavior. It is not a footnote in the curriculum, it is the spine of it. I say some version of “check the work, especially when you like the work” to every room I walk into. And I very nearly shipped a 672-person email built on a default value, because a big number showed up at a moment when I wanted a big number.
I did not fail my own curriculum. I got within about four minutes of failing it, which is close enough that I have stopped telling the story as a win.
There was one more piece of luck in this, and I should name it because it was not luck, it was a colleague. Someone in comms told me that if I was going to build a newsletter, she would rather I use the sanctioned corporate tools than some personal platform, for privacy reasons. My first reaction was mild deflation, because the sanctioned tools are less fun and the fonts are worse. My second reaction, roughly ninety seconds later, was that she was completely right and I had not thought about it for even one second. Two verification failures in one project, and I only caught one of them myself.
Warm before wide. I’ve been muttering it at myself ever since, usually when I want to go big.
Send it to the people who will actually read it. Ten is a real number. Ten people who open the thing and reply with a question is a newsletter. Six hundred and seventy-two people who have been algorithmically assigned enthusiasm is a deliverability problem with a spreadsheet attached. Earn the next hundred. Then the next. And never, under any circumstances, let the newsletter become the subject of the newsletter, because the moment a program starts reporting on its own reach, it has quietly changed jobs from helping people to describing itself.
Ten readers. That is where it starts. I know their names.