If you're new here: this newsletter is about what it actually takes to make AI work at the infrastructure layer. Not trends. Not model comparisons. The unglamorous, foundational work that separates AI that changes how a company operates from AI that lives in a slide deck.
Every week, for years, a team somewhere in our operation would pack up a banker's box full of paperwork and ship it to the home office.
Not upload it. Not email it. Put it in a box, tape it shut, and put it on a truck.
I learned about this approximately six weeks into leading digital transformation at a 79-year-old maritime logistics company. I was mapping how operational decisions actually got made. Not how the org chart said they got made. The real sequence of events between something happening and someone deciding what to do about it.
The banker's box was the answer for one class of decisions. The paperwork in that box was the data those decisions ran on. Once a week.
That's the data infrastructure underneath a lot of companies currently running AI initiatives.
Most companies don't have an AI problem. They have a data problem they're hoping AI will solve.
The data problem looks like this.
Decisions happen on calls with no record. The conclusion gets emailed around afterward, if anyone remembers to do it. The next person who needs to make the same call finds a two-year-old email chain, a person who's changed roles, and a spreadsheet last touched by someone who no longer works there.
Reports live in PDFs in someone's inbox. If you need a number from six months ago, you're searching email. If you need a trend, you're asking the person who built the original report to rebuild it manually.
Institutional knowledge retires when the person does. The most experienced operators carry the entire decision logic for their domain in their heads. When they leave, years of context walks out with them.
This isn't a criticism. These systems existed before the modern data stack, and they worked well enough to build lasting companies. The banker's box got the data home. It just doesn't work when you're trying to train a model on it.
AI needs structured inputs. It needs data captured intentionally, in a format that can be read by something other than the person who created it, at the moment a decision requires it, not once a week on a truck.
When those inputs aren't there, AI doesn't solve the problem. It amplifies it. You get wrong answers you can't argue with, generated at scale, from a system that sounds authoritative. That's worse than the spreadsheet.
Vendors want to show demos. Boards want to hear about the AI initiative. Teams want to build something that feels like progress.
The companies winning on AI right now aren't the ones who moved fastest to the tools. They're the ones who did the quiet work of building the foundation first.
The Foundation-First Sequence has four steps. I've run it three times: across advertising technology, retail AI at Kroger, and maritime logistics.
Start by mapping the decisions that actually drive the operation. For each one, trace backward: what data does this decision require? Where does that data currently live? Is it captured at all, or does it exist only in someone's memory? At Pasha, this meant pulling every system that touched a cargo decision: ERP records, spreadsheets, email threads, the WhatsApp groups people actually used to communicate vessel status. The largest category was decisions happening with no record at all.
For every gap that surfaces, design the capture point. This is not a technology conversation yet. It's a data-flow conversation: what needs to be true about how information moves from event to record? We picked three decisions that happened more than 20 times a day and worked backward: what information would make each of these obvious? That became the data we needed to capture. The instinct is to capture everything. Don't. Decide what you actually need before you build anything.
Build it in production, not in demo. A proof of concept in a sandbox is not a foundation. The foundation is what runs when nobody is watching it. Structured inputs where freetext had been. A single vessel status source that three teams had been maintaining in parallel spreadsheets. Forms that replaced phone calls. No one wanted to be in those meetings. This is the work that doesn't go in the press release.
Then verify it works without the person who built it in the room. If the system requires ongoing expert interpretation to be useful, it's not a system. It's a person with extra steps.
Most AI initiatives skip to the model and come back to this sequence after something breaks. The ones that don't skip it don't break.
If you removed AI from your initiative entirely, would the underlying data problems still be there?
If yes, that's what you're actually working on. The model is downstream of that.
The banker's box is no longer on a truck. That work is done. The data that lived in it now lands in a system that can be queried, trended, and eventually fed into something that makes decisions faster than once a week.
It took longer than anyone wanted and produced fewer press releases than the AI layer will.
That's the job. The Foundation Letter is where I write about it honestly: what it takes, what it costs, and what it looks like when it works.
If that's useful to you, subscribe. It's free.
The Foundation Letter publishes weekly. One idea from inside real transformations, not from a conference stage.