There is an email thread in my archive dated December 8, 2021. One of my managers is sending two prototype prints, engineering drawings, "about as simple as they come," to a technology partner, to run through an OCR engine and start experimenting. The partner writes back the same afternoon asking for two more: a medium print and a complex one, so we can compare how the extraction holds up across all three grades of difficulty. Everyone signs off cheerfully and gets on with the week.
Nobody in that thread had any idea what was coming. ChatGPT was three hundred and fifty-seven days away. "AI" was not yet a word your dentist used. We were not futurists and we were not researchers; we were a manufacturing company trying to get data off our own drawings, because in a job shop the blueprint is where the money lives and a person has to read every one.
I keep that thread the way other people keep a first dollar, and I am writing it up here because the timing turned out to contain a lesson that matters more than the experiment did.
Start with why we were doing it at all. Every industrial business is, underneath the machines, a reading operation. Work arrives as documents: drawings, specs, purchase orders, packing lists, certs. The drawing is the dense one, dimensions, tolerances, material callouts, revision notes, and the reading of it gates everything downstream, quoting first of all. We did not need a vision of the future to want machines reading blueprints. We needed to quote faster than the shop across town. The problem was fully visible from the operating side years before the tools were ready, which is the first half of the lesson.
Now the second half, which is what the tools of that era were actually like, because the contrast is the point. Document AI existed in 2021 and 2022; we used it, through good partners who were honest about its shape. To read a document type well, you trained a custom model, and the training was hungry: you could start with fifty labeled examples and hope, but the hard cases wanted thousands, sometimes millions. The known frontier, stated plainly in my correspondence from that period, was variance: forms that differ visually while carrying the same information. Invoices, purchase orders, bills of lading, the exact paper an industrial business swims in, were the canonical hard case. The playbook was to find your highest-volume constant formats, train narrowly on those, and patch the model's repeated mistakes with hand-written fixes downstream. On formats that held still, accuracy in the mid-nineties was achievable. On formats that moved, you labeled more data and lowered your expectations. It half-worked, and half-working on your highest-volume paper was genuinely worth having. That is where things stood in the fall of 2022.
The last email in that chapter of the archive is from a document-extraction partner, dated November 19, 2022, walking through the options for improving accuracy: bigger labeled datasets, pretrained models plus programmatic fixes, narrowing to constant formats.
Eleven days later, the bell rang.
I will not retell what happened after November 30, 2022; you were there. What I can tell you is what it looked like from inside a company that had already spent a year on the problem. The entire cost structure of our project inverted. The thing we had been buying with labeled examples, the machine's ability to cope with documents it had not seen before, arrived as a general capability. Variance, the named villain of every technical conversation we had in 2021 and 2022, stopped being the frontier and became a Tuesday. Problems we had carefully scoped down to make tractable un-scoped themselves. We did not have to change what we wanted. The want was suddenly affordable.
Here is the lesson, and note carefully what it is not. It is not that we saw the wave coming. We did not; nobody in that email thread predicted transformers would eat the document stack within a year. The lesson is that we did not need to. What we had, and what any operator can have, is an inventory of our reading: a precise, priced understanding of where documents gated our money, drawn from running the business, not from reading the news. When capability jumped, we did not spend a year discovering our use cases. We already had them, ranked, with the pain quantified. The prediction was worthless and the preparation was everything.
That is the general form of the advantage operators hold in all of this, and it is the reason these field notes exist. The frontier is built in San Francisco, but the problems are inventoried in plants and back offices, and the problems arrive years before the headlines do. You cannot control when the bell rings. You can control whether, when it rings, you are holding a list.
The work in that 2021 thread never stopped, by the way. It just kept compounding: the blueprints led to document extraction, the extraction led to whole companies, and the drawings we once mailed to an OCR engine to see if it could find the part number are now read by software that derives the machining from the physics. Say the ambition plainly, since five years have earned it: we want our machines to read the blueprint and build the part. Five years on one problem, started before the problem was fashionable, and the problem turned out to be the size of the whole industry.
Keep your list. The next bell is already scheduled; nobody has told us the date.