Thirty months in. The first eighteen taught me the lessons. The last twelve made me live them.
A year ago I ended a piece with a line I believed: AI won’t replace humans, but humans who can effectively integrate AI into their workflow will replace those who do not.
It’s an easy thing to write when you’re the executive who approves the budget, blesses the pilot, and shows up for the demo. You get to narrate the revolution from a comfortable distance. What I didn’t say out loud was the obvious follow-up, the one that sat there staring at me: did I believe it enough to do it myself?
So this year I found out. I stopped describing the work I wanted and started building it with my own hands — late at night, badly at first, then less badly. The analysis that used to take a week. The models I used to wait on. The synthesis I used to delegate and review. I stopped writing the requirements and started writing the thing.
Some of it was good. Some of it was embarrassing. All of it taught me something no demo ever has.
This is Part Two. Same trenches, a year deeper. Here’s what held up, what I had wrong, and the five lessons I could only learn with my hands on the keyboard.
But first, the quiet part, said out loud — because it may be the most useful paragraph in this essay. Most of what gets repeated about AI in boardrooms today isn’t slightly wrong. It’s expensively wrong. That AI is an IT project. That the company with the best model wins. That the bottleneck is GPUs. That the payoff is fewer people. That moving faster is the same thing as getting ahead. Every one of those is comfortable, every one is wrong, and every one carries a price tag your finance team hasn’t been handed yet. The five lessons below are really five of those assumptions, taken apart by hand — mine included, because I believed most of them too.
First, the scorecard
I won’t re-litigate all eight lessons from last year. But a thesis you never grade is just a slogan, so here’s the honest tally.
The lesson that relevance has a 24-hour half-life held up — I undersold it, if anything. In a market reorganizing itself around AI, insight no longer ages in days. It ages in hours.
The lesson to listen for the quiet signal — the thing people tell you when you finally stop talking — held up, and matters more now than it did. The machine’s best trick is still listening without ego.
And the partnership thesis — human and machine, not human versus machine — held up so completely that the rest of this piece is really just five harder, more specific versions of it.
But the lessons that held up aren’t the interesting part. The interesting part is the things I had wrong, or had right for the wrong reasons. That’s where the last twelve months actually lived.
Lesson 9: You have to put your own hands on it
Builds on last year’s Lesson 2 — “Scale expertise.”
Last year I argued you scale expertise by capturing it: record the expert, train the system, distribute the knowledge. I still believe that. But I had the org chart of it wrong. I assumed my job was to point at the work, and that someone closer to the keyboard would build it.
That’s the comfortable lie of leadership in the AI era, and it cost me months. The distance between sponsoring AI and building AI is the entire lesson. You cannot delegate your way to fluency in something this new. You cannot steer what you have never driven.
So let me kill the assumption directly: AI is not a thing your organization adopts, the way it adopted cloud or a new CRM — a program you charter, staff, fund, and review at a quarterly steering committee. Treat it that way and you will get exactly what every other company running the same playbook gets: a tidy slide deck and no capability. AI is something you build into your own hands or watch erode under you. The org chart does not save you here. Your title does not save you here.
So I drove. And the education was immediate and humbling. I watched a system do in ninety seconds what used to take days — and, in the same sitting, watched it produce something confident, polished, and completely wrong. You don’t learn either of those things from a status update. You learn them at one in the morning, with your own hands, owning the output.
Here is the part that matters for everyone reading this who leads a team: that line I wrote last year was never aimed at your employees. It was aimed at you. Humans who integrate AI will replace those who don’t applies hardest to the people who assume their seniority exempts them from the keyboard. It doesn’t. It never did. The most dangerous place to stand right now is one level above the work.
Lesson 10: The bottleneck moved from the model to the data
Builds on last year’s Lesson 5 — “Architecture matters, beyond the widget.”
Last year I warned against the chat-widget mirage — the belief that the interface is the product, when the real work lives underneath. I was right. I just aimed it too low.
Because here’s what a year of building actually taught me: the model is no longer the hard part. Everyone rents the same handful of frontier models. They are extraordinary, they improve monthly, and they are racing to become the most commoditized layer in the entire stack. If your only edge is “we use a great model,” you have no edge. So does everyone. Somewhere in your company there is a slide that credits your AI advantage to the model you selected. Delete it. It is describing a rental that your competitor can sign for by Friday.
Earlier this year I made a version of this argument in public — a two-part series for Forbes titled AI Will Finally Break How Data Is Stored — Or Break Your Data Center. I framed it then as an architecture problem: unstructured data scattered across clouds and edges, power turning from a line item into a hard ceiling, and storage forced to evolve from a box you buy into the conductor of your entire data estate. I believed every word. But I’d written it as an analyst writes — from above. A year with my hands on the work taught me the same lesson from the inside, and the inside is more brutal.
The edge is the data — specifically, whether your data is ready: discovered, classified, governed, given meaning, and shaped into something a model can actually use, before anyone asks the question. The figure that should keep every executive up at night comes from Gartner: through 2026, organizations will abandon 60% of AI projects that aren’t supported by AI-ready data. Not because the models failed — because the data underneath them was never ready. I lived this in everything I built. The model was never the thing that failed. What failed, every single time, was data that was fragmented, unlabeled, trapped across systems, or stripped of the context that made it mean anything.
So here’s the line I’d tape to the wall: the expensive hardware isn’t idle because it’s slow. It’s idle because it’s starved. That capex you personally approved — the cluster that took three meetings and a board nod to fund — is sitting underused, and the reason isn’t on the spec sheet. It’s in your data. We spent two years treating compute as the constraint while the real one sat untouched inside our own systems, unprepared and unfunded. The winners of the next phase won’t have the best model — that’s a rental everyone holds. They’ll be the ones whose data was ready to feed it. The model was never the point. The data always was.
Lesson 11: “Pull” became “push” faster than I forecast
Builds on last year’s Lesson 7 — “Start with pull, evolve to push.”
Last year I described push as the aspiration — the someday-state where insight finds you before you go looking, answering the question before you’ve even formed it. I called it the horizon. I was wrong about the timeline. The horizon arrived early.
The work stopped waiting to be asked. It now shows up where I already am, unprompted, having already done the reading. That’s genuinely transformative — the quiet tax of remembering to go check the thing simply disappears.
But push broke something I didn’t see coming. When the machine stops waiting for your question, it also removes the moment where you would have caught its mistake. Pull has a checkpoint built in: you ask, you read, you judge. Push deletes the checkpoint. The answer arrives pre-formed, confident, and frictionless — which is the exact condition under which a wrong answer does the most damage. The convenience that makes push valuable is the same convenience that makes it dangerous. Which leads directly to the lesson that surprised me most.
Lesson 12: As the output got better, trust got harder — not easier
Builds on last year’s Lesson 6 — “Behavior change and trust take time.”
Last year I said trust takes time, and I meant adoption time — the human runway you need to fold a new tool into old habits. I assumed trust was a curve that only climbs: the better the output, the more you trust it, the faster you lean on it.
That curve is a trap.
And it runs straight into the most dangerous assumption your culture holds: that speed is the same as progress. Every reflex in a high-performing company rewards moving fast — ship it, send it, put it in the deck. AI takes that reflex and turns it into a liability, because it now produces confident, polished, finished-looking work that is wrong, faster than any human ever could. You have already watched this happen. The demo that wowed the steering committee. The analysis that went straight into the board pack because it looked done. The number nobody questioned because the formatting was immaculate.
Here’s what actually happens. When the machine was visibly clunky, I checked everything, because I didn’t trust it. As it got good — genuinely, impressively good — I started checking less, because it had earned it. And that is the moment of maximum exposure. The better the work looks, the less you scrutinize it, and the more expensive the error you let through. I have watched something arrive beautifully formatted, fluent, sourced-looking, and flatly wrong — and nearly waved it through because it looked right. The polish is the camouflage.
So I had it backwards. Trust isn’t the summit you climb toward. It’s the thing you have to deliberately withhold, even — especially — as the work improves, because the improving is exactly what lowers your guard. Verification stopped being a phase and became the job. The discipline that matters now isn’t learning to trust the machine. It’s learning to keep your skepticism alive after it has given you every reason to retire it.
Lesson 13: Taste became the scarce skill
Builds on last year’s Lesson 8 — “Human + Machine, not Human vs. Machine.”
Last year I framed the partnership as complementary: the machine accelerates, the human judges. Right again — but I was vague about which human skill does the judging. A year of building made it specific.
It isn’t prompting. Prompting was the skill of 2024, and it’s already fading — the models now meet you most of the way. It isn’t producing the work, either; the machine generates more of that than I do, faster. The scarce skill — the one that actually separates the people getting leverage from the people drowning in output — is taste. The judgment to know which of five plausible answers is the right one. Which insight is real versus merely fluent. Which sentence lands and which one is hollow.
This is the human capacity that gets more valuable as the machine gets better, not less. When generating ten options costs nothing, the entire game becomes choosing — and choosing well is a deeply human act built on experience, context, scars, and the thing we don’t have a clean word for: knowing. The machine made everything cheap to produce. It made nothing cheap to judge. That gap is where the value moved, and it’s where it’s going to stay.
And this is exactly where the business case quietly falls apart. The spreadsheet that justified your AI investment almost certainly promised savings from doing the same work with fewer people. That is the wrong number, and it’s pointed in the wrong direction. The real return is stranger and far larger: the same people making sharper decisions, faster — but only if they have the judgment to tell the good answer from the merely plausible one. You cannot bootcamp that. You cannot hire it in a quarter or buy it in a license. AI did not lower the value of seasoned human judgment. It quietly inflated it — at the precise moment most companies are planning to need less of it.
The next eighteen months
A year ago I compared this moment to the early days of cloud: promising, chaotic, inevitable — and I said the winners would be the ones doing the unglamorous foundational work while everyone else chased the shiny thing. I’ll stand on that. But I’d sharpen it now, having earned the right to.
The unglamorous work isn’t a technology project. It’s a personal one. Get your own hands on the keyboard. Get your data ready before you chase another model. Keep your guard up precisely when the output stops giving you a reason to. And spend your judgment on the one thing now scarce enough to be worth it: knowing what good actually looks like.
If you keep one thing from this, keep the demolition list: the model isn’t the moat. The GPU isn’t the bottleneck. Fewer people isn’t the payoff. And speed isn’t progress. Every one of those is a story that’s easier to tell a board than the truth — and every one of them is how serious companies are going to quietly lose the next five years while believing they’re winning.
I rewrote my closing line this year — not because the original was wrong, but because it was too gentle:
AI won’t replace you. But the version of you that refused to build will be replaced by the version of you that did.
I went and became the second one. Eighteen months of watching taught me the lessons. Twelve months of building is the only thing that made me believe them.
— Field Notes, Part Two.
Sources & related reading: Shawn Rosemarin, “Human + Machine: 8 AI Lessons After 18 Months in the Trenches” (2025). Shawn Rosemarin, “AI Will Finally Break How Data Is Stored — Or Break Your Data Center”, Forbes Technology Council (Feb. 10, 2026). Gartner, “Lack of AI-Ready Data Puts AI Projects at Risk” (Feb. 26, 2025).
Leave a Reply