I use AI models every day to make a living. I also read the same reporting you do about water tables and power draw and the way these companies are racing to get data centers poured before anyone can organize against them.
Those two things are supposed to cancel each other out. I’ve been told so, more than once, usually by people I agree with about almost everything else.
I’m not going to argue you out of that. What I can do is tell you what this looks like from inside a room where it actually happened, because I was standing in one, and what happened there is not what either side of this argument told me to expect.
The department had about fifty people in it. Ten of them wanted to write code and weren’t allowed to.
Not because they couldn’t. Several were better engineers than the job ever let them prove. They were DevOps, and DevOps was the lane, and the lane was the point.
I learned that in my first week, walking through how a Fortune 500 company’s core infrastructure team spent its days. I was there to work out how to get them using models. I came in with one rule for myself, which was that I would not hand anyone a plan that ended with their coworkers being walked out of the building.
Then discovery told me something I didn’t want to put in a deck. The way that department worked was inefficient enough that if you dropped models into it honestly, you wouldn’t be augmenting those people. You’d be replacing a good number of them.
The Busywork Was the Point
The rest of the developers spent their days on API maintenance. Not building APIs. Maintaining them. And when I got far enough into discovery to look at how those APIs were actually designed, I found something that took me a while to say out loud, because once you say it you can’t unsee it. They were built in a way that guaranteed they would always need maintaining.
Here’s one I still think about. Support tickets came in as text files, dropped into a folder. A script pulled them into a spreadsheet. The script wasn’t reliable, so someone checked its output. The script broke whenever the input shifted, so someone maintained the script. Every spreadsheet got saved and kept, because the boss had decided that was what auditing required. Then an API read the spreadsheet.
A machine-readable file, converted by a machine, verified by a person, maintained by another person, so a different machine could read it again. The accountants didn’t like the arrangement either. They said yes anyway.
And it wasn’t just tickets. Purchase orders went through something similar. So did recalls. A whole catalog of things the company runs on, each one with its own variation of the same arrangement, each one with somebody’s week attached to keeping it upright.
Nothing ever got redesigned. It got extended. A new requirement showed up and somebody bolted a step onto the end of a chain that was already too long, because rebuilding it properly would have taken a quarter and nobody was going to approve a quarter. So the endpoints got treated like a redheaded stepchild, patched and extended and patched again, for years.
This is 2026.
And the whole chain sat one stray character from failure. A malformed filename, a date typed as text, a column nudged one to the right, and the API downstream was reading nothing, and somebody lost an afternoon working out which file did it.
Nobody had to conspire to make that happen. But the effect was clean. Fifty people were always busy. Busy is what upper management sees. Busy is what you point at in a quarterly review when someone asks what your department is for.
And from two levels up, manufactured busywork and actual productivity look exactly the same.
The man running that department had been at the company thirty-five years. His people knew how the rest of the company saw them. Second rate. They were smart. They knew what the company’s real problems were. They could list the top ten if you asked them, and several of them did. They just weren’t permitted anywhere near them.
They also knew they were falling behind the companies they measured themselves against. One of the developers put it to me plainly. “If we’re not using models now, this department is done. It doesn’t have to do everything but it would be a huge help.”
That’s what I heard in it, though I should be careful here. I don’t know that anyone upstairs was paying attention. I never met the COO or the CFO and I have no idea whether that department was on anyone’s list. What I had was a feeling, and the feeling came from having seen this shape before. In a company that size there is usually a day when somebody takes a serious look at a fifty person cost center and works out what the days actually consist of. When that happens, nobody gets retrained. They get a number attached to them and a date.
So the department’s real choice may never have been whether to change. It may have been whether to change on its own terms or wait to be found out. That’s my read. Nobody said it to me in those words.
I took the finding to the person who hired me, a lead developer with serious leverage and one of the department’s managers. Then the other managers joined. And within a few minutes I understood that I had walked into the middle of something.
The Only Person in the Room Who Said No
He was untouchable, and it’s worth being precise about why. Thirty-five years and genuine respect. That combination is armor in a company that size. Nobody above him was going to move on him over a proposal from an outside consultant, and everybody knew it. The only people who could touch him were his own.
As for what was actually driving the people around him, I have no idea. It’s entirely possible somebody wanted his seat. In a department that size there’s usually at least one person doing that arithmetic. But I didn’t want any of that shit. I was just trying to help, and I spent a fair amount of that first engagement being circled.
That part is not unique to this company. Walk into any billion dollar organization as a consultant and you are an unknown quantity with a badge. Before anyone listens to a word you say they are working out whose side you’re on, who brought you in, and what your presence means about their own position. You get sniffed at first. You get heard later, if at all.
What I could read was the shape of it. Status quo against people who wanted to build something. That fight was already running when I arrived, and the proposal I handed over just gave it something to happen about.
Here’s what I did not expect.
The workers wanted it. Not cautiously. Not after a six month change management program with a communications plan and a feelings survey. Forty of them were ninety to ninety-five percent in, and they got there fast, because they could see what I could see. The remedial work was the wall between them and anything that mattered.
About ten were loyal to him. I want to be fair about them too, because they weren’t fools. Some of them had built standing in that department on being the person who knew how the mess worked. Take away the busywork and you don’t just take away their tasks. You take away what made them necessary. Anyone who has worked anywhere knows those ten.
The only person who said no was the one whose signature I needed.
Nobody told me no. That isn’t how it works at that level.
What I got was an email from one of the ten.
”Thank you for taking the time to understand our issues, Mr. Castro. We will get back to you.”
Anyone who has done contract work knows that sentence. It is perfectly polite and it means the thing is dead, and it means nobody is going to be on record having killed it. There’s no argument to have, no decision to appeal, nobody to follow up with, because on paper nothing has happened yet and nothing ever will.
So I walked. No contract, no project, nothing to show for the discovery work.
I want to be fair to him here, because the easy version of this story has a villain in it and the easy version is worse. He was not stupid and he was not lazy. He read the proposal as a threat to his position and he was completely correct. If his department stopped being busy and started solving the top ten problems in the company, somebody was eventually going to ask what he had been doing with fifty people for the last decade. He understood the stakes better than anyone else in the building.
He was right about the threat. He was wrong about who it was bad for.
The Chairs That Were Empty
Two weeks later they called me back.
I recognized most of the room. Not all of it. I’d had real conversations in that first meeting, people who’d walked me through what their week actually looked like, and some of them weren’t there anymore. So I asked.
No longer here. Retired. Moved to a different department. The answers came back flat and short and nobody volunteered a fifth word. I asked a few different people and got a few different versions, and none of them added up to an explanation. I still don’t know what happened in that room after I left. I wasn’t in it.
The boss had retired. That much was clear. Whether he jumped or was pushed I can’t tell you, and the people who could weren’t offering.
The first thing I did was ask what the audit requirement actually was. Which regulation. Which retention schedule. Who had ever asked to see one of those files.
Nobody could tell me. Not the accountants who had been saving them, not the developers maintaining the script that produced them. The department was a silo. The man who knew had retired, and the reason retired with him.
What I could tell was that the people who remained had already decided. There was no pitch to make. Nobody needed convincing that the maintenance work was a wall, or that models could take it down. They’d worked that out on their own, probably well before I showed up. Something had shifted in that department, and I was not the cause of it. I was the occasion.
What Fifty People Do When You Take the Busywork Away
I keep coming back to the test cases.
They wrote them constantly. Wrote them, updated them, rewrote them when anything upstream moved. It was the largest single block of work in that department and every person doing it hated it. Not in the way people grumble about a task they’d rather not do. In the way you hate something you’ve decided is beneath what you’re capable of and have to do anyway, every week, for years.
The model writes the test cases now.
Seven of them had been assigned to documentation. Not documenting their own work. Documentation as the job, doc-only pull requests, comments in other people’s code when others forgot to. Sticking in reasons for the change request on a markdown file the moment a branch goes out. Seven developers leaning more as business analysts, in a department of fifty, at a Fortune 500 company. They didn’t want to get too comfy with this routine.
The model does that now too.
What I want you to notice is what did not happen next. Nobody was walked out. The seven documentation developers went and worked on the components they’d been asking to touch. The DevOps people who did nothing but reviewed pull requests started writing code. The scripts went away and the endpoints went direct, and the API design got rebuilt granular, with the model used less as a code generator than as a sounding board, a place to throw endpoint structures at the wall and see which ones held up.
The refactoring is where I saw the clearest version of what this tool actually does. That codebase was enormous. It makes BenchBoard look like a high school swimming pool right next to Lake Merritt. The thing that stops anyone from cleaning up a codebase that size isn’t the editing, it’s that you cannot see the blast radius. Change this, and what breaks three systems over? Nobody knew, so nobody touched it, so it stayed broken. The model could trace the impact. They still did enormous amounts of manual work, because it’s an enormous codebase and there’s no version of this where that goes away. But they could finally see what they were about to hit.
Most of them worked hybrid - part agentic, part manual. Write the real logic themselves, hand the remedial parts to the model and keep it in plan mode when they wanted suggestions rather than changes. Some went straight for the UI work they’d never been allowed near and iterated faster than I expected.
Here’s the part that surprised me, and it’s the reason I’ve stopped calling any of this training.
I barely did anything.
I showed them prompts I’d used to produce interfaces quickly. I flagged the gaps that agentic development opens up if you’re not watching for them, security, login flows, database design, the internal things that still matter no matter how much of the typing gets handled for you. And then I mostly stood there. They hit the ground running. They were like kids with a new toy. They didn’t need convincing and they didn’t need a seminar. They needed permission and a guide for when they hit a snag.
That was it. That was the whole intervention. Twice a week for eight weeks. Sixteen days on site, and most of what I did on those days was answer questions and point at things to watch out for.
By the end of it, ideas were surfacing that had been shelved for years. Things people had wanted to build and had written off as impossible, not because they were too hard but because there was no time and there was never going to be any time. I didn’t chase those down. I wasn’t there to. But they knew what they were, and they had known the whole time.
The other half of that department wasn’t developers. Analysts, people who worked with numbers, some accountants. I couldn’t tell you exactly who was what, and I won’t pretend I mapped it.
What I could see is that their work moved up. They had been spending their days on requirements analysis and impact analysis, the sort of thing the model handles now. So they went at the problems they’d been describing to me all along. Better processes. Self service interfaces so other departments could stop filing requests and waiting. Tighter integrations with outside vendors.
That last one is the clearest example of what changed for them. Getting two vendor APIs to talk to each other used to be a project, which meant a request, which meant a queue, which meant it didn’t happen. Now one of them would prototype the endpoint, stand up a rough admin page, and go. Not a finished product. Enough to prove the thing worked and hand it to someone who could take it the rest of the way.
What I Can’t Tell You
I’ve just told you a story where it worked. Before you decide what it proves, here’s everything I can’t account for.
Start with the part that looks best for me. My contract ended two months ago, and they’re still hiring. The bigger problems need people to steer the models, and steering takes judgment that doesn’t come out of a model.
I want to be honest about what that is. It’s secondhand and it’s two months out. I don’t have their org chart and I can’t verify their reasons. What I can tell you is that two months after an intervention that eliminated an enormous amount of work, that department was hiring rather than shrinking. That’s what I know. It isn’t a trend, and I’d be lying if I told you a budget cycle from now looks the same.
Here’s what I actually can’t see, and it’s the thing the people worried about this are right to ask about.
Nobody in that department lost a job. But I can’t tell you about the job that never got posted. Somewhere there’s a company that would have hired a fourth accountant and now runs fine with three, and that job never existed, so it never shows up in anybody’s numbers, including mine. It won’t appear on a layoff list. It won’t appear in a survey. I can’t prove it’s happening and neither can anyone else, and the fact that it’s unmeasurable is not an argument that it isn’t real.
Then there’s the obvious one. This is the department where it went well. That’s why I’m telling you about it.
I’ve never watched the other version happen from the inside. But I can describe it without effort, and so can you. A manager takes the same efficiency, doesn’t retrain anybody, banks the freed hours against a headcount target, and hands three names to HR in the fall. Nothing about the tool prevents that. The tool is indifferent. What determined the outcome in that department wasn’t the software. It was who held budget authority over that team and what they were being measured on that quarter.
That’s the variable. Not the model.
So take this for exactly what it is. One department, one engagement, sixteen days, one outcome. It’s evidence that this can go the workers’ way, because it did, and I was standing there. It is not evidence that it usually does.
Who Actually Drives This
There’s a version of this argument where the labs are the story. They aren’t.
They built something good. I’m not taking that away from them. But the people who decide what it becomes are the ones using it and paying for it, and I don’t think that gets said enough in either direction. The anti-AI side talks about these companies as though they’re weather. The companies themselves sometimes talk about their users the same way, as a number that goes up. Both are wrong about who has the leverage.
There are a lot of models now. That’s the part people keep missing. Switching costs are real but they aren’t prohibitive, and a lab that stops being useful to the people actually building things finds out quickly.
The internet went the same way, and it’s worth remembering how many gates it actually took off the hinges.
News ran through a handful of outlets, and if they didn’t run your story, your story didn’t exist. Then anyone with a camera and an account could reach millions directly, and the institutions that had held that power spent the next decade describing the change as a crisis. Selling anything at scale meant shelf space, and shelf space meant a buyer at a big box chain deciding your product was worth a few feet. Now somebody sews quilts in their garage and sells them to the entire country without ever meeting a buyer or renting a square foot. Talk radio decided who got a microphone and who didn’t, and the list was short and the criteria were whatever the station said they were. Now it’s podcasts, and the list is everyone.
Same pattern every time. Somebody stood between the work and the people who wanted it. The internet removed them. And in every case the removal got described as degradation by the institution doing the standing.
What I think the models add to that is narrower and more useful than the way it usually gets sold. They take the tedium. The parts of running a business that nobody chose and nobody enjoys and everybody still has to do. The reconciliation, the reformatting, the maintenance on a system somebody bolted together nine years ago. Most people who start a business did it because they wanted to make something or serve somebody, and then spent years discovering how little of the week that actually gets. Handing the tedium to a model gives the week back.
That’s what I watched happen in that department. Nobody there wanted to write test cases. They wanted to solve the company’s top ten problems and talk to the people who had them.
What I see now is the same thing happening to enterprise software. That department I was in had been running on tooling that only a Fortune 500 could afford to build and maintain. I build things at that level by myself. Not a version of that level. That level. A solo operator, or a shop of three people, can now ship what used to require a department and a budget cycle and four layers of approval, and without most of the politics that decided whether ten people in that room were allowed to write code.
That’s the thing I’d want someone on the fence to weigh. Not that the technology is harmless. That the leverage it’s redistributing has been running the other direction for forty years, and it’s running downhill now, and the people arguing loudest to shut it down are not the ones who’d be protected by that.
I own its faults. I’ve written about them at length, including the ones that cost me weeks of my life. That’s most of what this newsletter is.
I’m a regular guy with aspirations who looks at this as a tool.
Not as a side to be on. We have plenty of those already, and the loudest ones are the fights where most of us have the least purchase. I can vote, and I will, and November is close. But I don’t get to personally fix what happens in Washington and, unless you’re also running for congress, neither do you.
This is different. This one is still being decided, in rooms like the one I’ve just described to you, by people whose names you’ll never know, making calls about their own departments. That’s where it gets settled. Not on the internet.
Every technology I’ve watched arrive in nineteen years has landed the same way. Powerful, misunderstood, feared for a while, and then absorbed. The internet was going to ruin us too. What made the difference wasn’t the technology. It was whether the people paying attention stayed in the room.
It needs to be understood. It needs agency, and agency only comes from humans. And it needs to stop being the thing that sorts us into camps, because the camps are how the decisions get made without us.
So I’d rather be in the room. Anti-AI, pro-AI, undecided, it doesn’t matter to me. I want the people who give a damn in there, and I know I’m one of them.