Revision 2. ~2,440 words. One blocking bracket remains. Do not publish until it is filled or the sentence is cut. Two permission items are listed at the foot. Both need a conversation, not an edit.
I spent three years teaching AI to leaders, nonprofits, businesses, and churches before a single one of them made a move.
It was not that they were unimpressed. For the first time in my life the feedback was so unanimous and so consistent that I took to thinking of my demonstrations as magic shows. But unlike magicians, I actually wanted to teach people the tricks.
One morning over coffee, a friend who consults with nonprofits and does not hand out compliments made me stop and pay attention just so he could say a phrase I had never heard before. You da prize. I hoarded that one like manna for months.
And they kept asking. Again and again, show me that thing again. I have been a leader for a long time and it was the first time in my life I have ever been pursued.
I want to be honest about why that mattered so much. I had been a pastor, then a nonprofit founder, then a marketing director, and I was now slowly arriving at the understanding that I was going to be something involving AI. Imposter syndrome is difficult enough without that sentence in your head. I needed the praise more than I would have admitted at the time.
But nobody seemed to want to learn the tricks.
That was the strange part, and the contrast was almost too clean. The most enthusiastic reception I have ever had for anything, and on the other side of it, nothing. No adoption. No practice. Not a single organization doing anything differently on Monday than it had done on Friday.
For a long time I read that as a failure of nerve or attention. I do not anymore, and the thing that changed my mind was a number.
In 2025, MIT's Project NANDA published a report called The GenAI Divide, which found that ninety-five percent of enterprise pilots produced no measurable impact on profit and loss.¹ The figure has been contested and the methodology is reported inconsistently, so hold the number loosely. Hold the finding. And note what the finding actually is. It is not that those organizations failed. It is that they ran pilots and the pilots produced nothing anyone could measure. The report's own explanation was not that the models were bad. It was that the systems did not retain feedback, adapt to a context, or improve, so nobody learned anything.
Set my friends against that and they come out ahead of me. They never bothered to attempt the thing I suspect they already knew they could not grasp, or could not adjust to. I was standing in their offices telling them the water was fine. They were the ones who noticed they could not swim.
I am still not sure why or how this began to change. But about a year and a half ago, it did.
The call
I will always remember my first Zoom call with Alan Hirsch. I secretly recorded it so I could share the transcript with my wife. It is my first and only act of human fandom. That is how much I loved that guy.
I read his work in 2010 and it changed the direction of my life. I spent the following decade running a community that taught his material to young adults. Brad Brisco, who has been in almost every conversation since, sat on that community's board from the beginning. So when Alan called to say he was ready to start building, which really meant he was ready for me to, the answer was going to be yes regardless of what it did for anything else.
I have said several times since, in a tone of somber gratitude somewhat out of character for a moody overthinker, that if building the place where Alan Hirsch's vast and vital work could finally come together as a whole was the only thing I ever got to do vocationally, it would be more than enough.
What I did not notice at the time is that his request had a property none of my organizational work ever had. Alan did not need to convene anybody. No board, no task force, no budget cycle, nobody else to agree. He decided on a Tuesday and the work started that week.
I did not read that as a finding. I read it as luck.
What I thought I had agreed to
I had been asked to build a web app. That is how it was described, that is how I understood it, and for months that is how I worked, even though most of what I was actually doing had nothing to do with code.
Alan has published more than a million words in his English-language books alone. They arrived as an assortment of PDFs. Then there are hundreds of articles, podcasts, videos, assessments and tools, almost all of it either offline or spread across a variety of apps and sites, some of which no longer exist.
I treated this as prep work. Something the agents could handle in the background while I built the platform back to front with the other agents. Staging. Get the material into a usable state, then start the real work.
The prep work kept taking more of my attention. More weeks, and a great many more tokens. I never really mentioned it to Alan. I am not sure why. Probably because I still thought of it as setup, and you do not give status updates on setup.
Then I mentioned it to Brad, almost as an afterthought, somewhere inside one of the two-hour Zoom conversations we have been having for two years alongside what is very likely the longest text chain in the history of text chains.
And in the middle of describing it out loud I heard what I was saying. This was not the setup. This was the thing.
What I had actually agreed to
Here is what most people have wrong, and there is no reason they would have it right.
A PDF of a book feels like the book. You can open it, read it, search it for a word. So it seems obvious that a computer has it too.
It does not. A PDF is closer to a photograph of a book. The words are marks positioned on a page for a human eye, and nothing in the file says this line is a chapter title and that one is a footnote. Copy a few pages out and you will see it immediately. Page numbers land in the middle of sentences. Running headers repeat every thirty lines. Footnote markers come loose from the claims they belonged to. Two-column layouts interleave into nonsense. A word broken across a line break stays broken.
And that is the good case, where the file has real text in it. A book scanned from a print run in the nineties has no text in it at all. It has a picture of text, and something has to read it the way a person would, and it will get some of it wrong in ways that stay invisible until you go looking.
The rest of the material sits on a scale.
A web page is reachable. Plain text is reachable. A well-made ebook is mostly reachable and keeps its shape. A digital PDF is partly reachable and its structure is gone. A scan is unreachable until something reads it. An hour of a man working out an idea in front of an audience is a waveform, which is to say it is not words at all until somebody writes it down. An article on a site that closed in 2013 is unreachable in the strongest sense, which is that it is not there.
None of that tracks how good the material is. If anything it runs the other way. The blog post summarizing the book is fully reachable. The book is not.
So the work was not staging. The work was that the text had to travel, and then it had to be translated.
Travel in the ordinary sense. Out of flash drives and dead blogs and email attachments and a publisher's export, into one place prepared for it.
And translated, which is the part I had not thought about at all. Into formats with names like markdown and HTML and JSON, which are not really formats for reading. They are formats for structure. They are how you say, in a way a machine can act on, that this paragraph sits inside this section inside this chapter of this book, that this claim carries this footnote, that these words are somebody else's and this sentence is his.
Without that you do not have a body of work. You have a heap of text that a machine will confidently quote out of context in a man's name.
The part I should have known
Looking back it is strange that I underestimated this so badly, because I had already done it by hand.
In 2020 the pandemic cancelled the summer internship I was running. It was an in-person program built around Alan's material, and there was no in-person, so it had to become something people could work through online.
Which meant somebody had to get the words out of the books.
So night after night, for weeks, I typed chapters of Alan Hirsch's books into a course platform by hand. I bought one of those little hands-free reading stands to hold the pages open while I typed. I still have it.
That was the only way to do it. The books existed. They were published, edited, printed, bought and widely read. And the only way to get them in front of twenty young adults in a form they could learn from was for a man to retype them, one paragraph at a time, at a desk, at night.
Five years later I was solving the same problem with agents and still filing it under prep work.
I do not have a good explanation for that either. The closest I can get is that when something has always been done by hand, you file it under labour rather than under problem, and you do not see that it is a problem until something makes it cheap enough to look at.
What it turned out to be
What I built does all of it. A book goes in one end as a file and comes out reachable, structured, checked, with every claim traceable to the page it came from, and with a person looking at it wherever a mistake would matter.
The first one took about six months, and most of that was not building. It was finding out what the work even was.
The second one took weeks.
[PROOF, BLOCKING. One checkable thing, or the sentence above comes out. A link a reader can open, or a number with a date. This is the only thing standing between this draft and publication.]
Back to the magic shows
I was right. I want to say that plainly, because the modest version of this story is available and it is not true.
I was also years too early, and from the outside those two things look identical.
For three years I was a man dancing alone. Dancing well, I think. Certainly with conviction. And completely unfollowable, because there is no such thing as following one person doing something nobody has seen before. There is only watching him, which is what everyone did, enthusiastically, for three years.
So the enthusiasm was never an early stage of something that failed to mature. It was the whole of what was available to them, and it was complete on the first afternoon. I kept reading it as movement toward a decision because I wanted it to be that. It was not moving toward anything. It had already arrived at its finished form, and its finished form was an audience.
That is why the organizations never moved, and the reason is not that organizations are slow, though they are. It is that there was nothing yet to join. A group cannot be second. Somebody has to be first, and first is always going to be one person who can decide alone, on a Tuesday, without convening anybody.
Alan was the first follower.
I did not understand that for a long time because I was busy thinking of him as the first client. He was not. A client buys a thing that already exists. What Alan did was step out and stand next to a man dancing by himself, which is the move that turns a lone nut into a leader. That work is done by the follower, not the dancer. I had been trying to solve it from the wrong side of the problem for three years.
And then he told other people who I was. That is the second half of it and it is the half I could never have done myself. Everything since has come through somebody saying my name in a room I was not in.
So my friends were not wrong to wait. They were right that they could not grasp it, because what they were being offered was a performance by one person and no way in. Their capability never changed. What changed is that somebody joined, and once one person had, there was a thing to point at, and a group can look at a thing.
For three years I thought I needed a better trick. I needed one person to step out.
That is most of what I know now, and I would rather have found it faster.
Before this publishes
Two things in this piece involve other people and neither is fixable by writing.
The secretly recorded call. Alan should hear that this is going public from me, not read it. It is the best detail in the piece and I am not cutting it, but he gets told first.
The unnamed friend. The coffee shop line is real and the friend is real. He is unnamed here because a private person's offhand compliment does not belong in print without being asked. If he says yes, the name goes back in and the scene gets slightly better. If he does not answer, it stays as it is and loses almost nothing.
Source
1. MIT Media Lab, Project NANDA. The GenAI Divide: State of AI in Business 2025, July 2025. Preliminary report, widely covered from August 2025. For the ninety-five percent figure, which the report states as pilots stuck with no measurable P&L impact against roughly five percent extracting significant value, and for the learning gap as the stated cause. Reported methodology varies between accounts: some report 52 executive interviews and surveys of 153 leaders, others 150 interviews and 350 employees, both alongside analysis of 300 public deployments. That inconsistency is the reason for the hedge in the text. [Add the primary link once the report URL is confirmed.]
How this was made. I wrote it. AI helped me find things I had written elsewhere and check dates. Every number in it is checkable, and if you want to check one, ask me.