• All
• AI & operators
• Finance & rails
• Real estate

AI: Revolution or Renaissance

Kris Van Meter · July 28, 2026 · AI Strategy

This morning, before I finished my coffee, my agents briefed me on two businesses, ran my market desk, and rendered a client document. The payroll for all of it, paid in tokens, cost less than the coffee. A decade ago that work took a floor of people, and I know because I helped run the floor.

Everyone building or buying AI ends up at the same question sooner or later. Is this a revolution or a renaissance? I used to treat that as a rhetorical setup. I've come to think the two words are doing real work, because they name two different capabilities that history has always put in two different kinds of people.

The Renaissance, at its peak, was breadth in a single mind. Da Vinci held the whole library in one head: painting, anatomy, hydraulics, optics, weapons. And he shipped almost none of it. The helicopter stayed a sketch in a notebook for four hundred years, because having the idea and having the means to execute it were separated by centuries of unsolved physics and a supply chain that didn't exist.

The Industrial Revolution was the opposite capability. Ford had no pretense of range. One product, one process, and a discipline of execution so severe it cut the price of a car by two-thirds and paid the men on the line enough to buy the thing they built. Range without a factory. A factory without range. For five hundred years you got to be one or the other, and most people got to be neither.

Cheap intelligence closes the distance between those two men. It is the third leg of the stool, the one history never delivered, and a stool that finally stands is made for climbing, toward heights neither man could reach on his own.

I am not an engineer

At least not in the classically trained sense of the word, with a degree in an engineering field and a title to match. That path is real and I'm not selling a replacement for it. Mine ran the other direction, through necessity and hard constraints: a problem sitting in front of me, no budget to hire it out, nobody to hand it to. You go get what you need. That habit is the thing this technology rewards most, because there is almost nothing left you can't teach yourself if you're willing to go after it. I made a career out of knowing enough about a lot of things to check the work of people who knew one thing deeply. In construction and renovation that meant approving the bid, then walking the job to verify the work firsthand. Finance meant building the budget and reconciling the variances against accounting policy and timing. In product I wrote the design request, sat with the engineers, and graded what came back. And at the far edge of my range, when the data outgrew Excel and crashed it, I wrote my own code the hard way, character by character at first, then by stitching old snippets and scripts into new ones: cleaning pipelines, building implied yield curves for market and submarket research, teaching myself regression, k-means, isolation forests, the principles of machine learning this whole moment was built on. A colleague and I used to call ourselves jedi masters of copy-paste data transformations, and we were only half joking.

That career built a specific instrument: someone whose whole job was setting the standard, approving the plan, and verifying what specialists shipped. I just didn't know the specialists were about to become software.

In the past year I shipped a market desk that runs every morning whether I'm at the keyboard or on a job site, a multi-county parcel data pipeline built for land brokers, a document engine that renders client work to my firm's standards, and an iOS app for financial literacy and education. I typed almost none of that code. I use AI to build systems, not draft emails, and I invest time and tokens into making those systems more deterministic every month.

The unlock was a single realization, and my career had been preparing me for it without my permission. The AI agent is the engineer. I was never going to write production code myself, and I didn't have to. The job was running engineers, and I'd been trained for it by the best kind of luck: working beside people far more talented than me, patient enough to bear my dabbling and my never-ending questions. Because I never gripped any single craft too tightly, I could let go of the details and keep the abstraction, and when I built my firm's operating system, TUG.os, those people became the pattern. I designed the roles as skills, a CFO, a research analyst, an entire product and engineering team, each shaped by how the best person I ever watched do that job actually worked. Roles stay siloed, so a specialist stays sharp, and consequential decisions go in front of more than one perspective before they ship. No agent reviews its own work. It's a corporate org chart. The staff happens to be software, and unlike the generous humans who trained me, it never tires of my questions, but it still doesn't always laugh at my jokes.

It was not always like this

Everyone who sticks with this technology climbs the same ladder.

1. A better search engine.
2. A writing partner.
3. A pair programmer.
4. Agents working unattended.
5. Systems that build systems.

Nobody skips a rung. I can date mine by the scars.

Two years ago the iteration loop was miserable. Ask for an email draft, then ask for an edit, and the second version was often worse than the first. Small matters of taste were faster to fix by hand than to explain. The output was only ever as good as the input, and getting the input right took real work. What changed is that context windows grew and the models improved, so the rails could come off. What did not change is the part that still separates working systems from expensive chat: the better you understand your own architecture and plan, the easier it is to make a small update without unintended consequences. That was true of every operation I ran before AI. It is true of this one.

Two exercises taught me more about working with the machine than everything else combined. Describe a problem and make the AI interview you first, rounds of questions until it understands before it answers. Then, on problems where you can afford the detour, let go of the reins entirely and see where the horse takes you. I stay where I'm useful, on the ideas and the problem itself, and the transcript of that thinking becomes the build request. Work I used to wait months for, or never got to at all in a backlog, now comes back in hours looking like it went through a professional shop.

The agent rung is where everyone gets burned, and my burn taught me more than the wins. For three months my market desk posted a clean, confident brief every morning while reading nothing at all. One flag in one script had broken its credentials, and instead of failing, the system degraded gracefully, wrapping good formatting around empty data while a hundred and fifty-six alerts piled into a log nothing read. The corpse kept sending me status updates. The bug was one flag. The disease was the incentive I had given the system, look alive, and it optimized for exactly what I rewarded.

Latent rules rot. Enforced rules survive.

The rebuild moved everything that matters out of the model's good intentions and into structure the model can't ignore: hard failures instead of graceful degradation, a watchdog that treats silence as guilt, and checks that run whether anyone remembers them or not, because nobody will. That rebuild was the last rung of the climb: systems that build systems, and check them.

Invest once, reap recursive rewards

That rebuild followed an operating model I'd been converging on for a year, and a builder named @joezunenet has put it better than I've managed to:

"Agents should handle uncertainty, exploration, and exceptions. Once a behavior becomes understood and repeatable, capture it in deterministic code, API/CLI, tools, or workflows. The goal is not to eliminate agents. It is to stop paying for the same cognition over and over again. Tokens should fund discovery, not rent-seeking on already-solved problems."

That is the whole thing in four sentences. Agents explore. The moment a behavior is understood and repeatable, I harden it into code, and from then on I edit the code, or fork it into alternate paths, instead of asking a model to re-derive it every morning. When I correct a mistake, the correction gets written into rules that load at the start of every future session, so the same mistake can't be made twice. Discovery gets paid for once. Everything after that runs for cents, and the hundredth run costs effectively nothing.

That split is the accounting most AI budgets get backwards. Tokens spent at design time, on the smartest model available, are capital: they buy an asset that pays you back every run. Tokens spent re-deriving solved problems are operating expense, rent with nothing on the balance sheet to show for it. Book the build as the investment it is and the CFO's question changes, from what the subscription costs to what the asset returns per run. Every owner already budgets a renovation this way. Almost nobody budgets AI this way yet.

The compounding runs in both directions. My corrections flow down into the structure, and each model generation's gains flow up through it. I have refactored TUG.os on every major release, GPT-4, then Opus, now Fable, and the first move each time is the same. Before pointing the new intelligence at any new work, I reevaluate the whole machine and reinvest in the agents and the factory. Every behavior I captured under the last model runs cheaper and extends further under the next one. The outputs keep improving with the models because the structure was built to inherit their gains.

If you've worked near software engineers, none of this should sound revolutionary, and that's the point. Engineers have worked this way forever. You'd be shocked how much of professional coding is copy and paste: libraries, templates, patterns, solved problems reused instead of re-solved. The syntax was always the cheap part. The discipline is knowing what to capture, when it's ready, and how to organize whoever is doing the capturing. That is management work, and operators are allowed to be good at it.

Uncle Bob Martin, who started coding in the late sixties and spent a career teaching engineers how to hold the craft, describes his own setup this way:

"My current strategy is to not read any of the code written by my agents. That's the only way I can take advantage of their productivity. What I do instead is to surround the agents with extreme constraints. Unit tests, gherkin tests, QA procedures, quality metrics, mutation testing, test coverage, and a plethora of others. In the end, I have very high confidence in the code they produce because they've had to run the gauntlet of all of my constraints and tests."

Fifty years of craft, and the job he describes is building the gauntlet rather than writing the lines. Someone still has to push the frontier models themselves forward, and that work belongs to people with the training for it. Building something that runs a real business is a different job at a different bar, and it runs on constraints. Constraints are what operators have lived inside their entire careers.

Show me the incentive

So why did S&P Global find 42 percent of companies abandoning most of their AI initiatives last year, up from 17 percent the year before, while the tools improved the whole time? Munger answered that decades ago. Show me the incentive and I will show you the outcome.

Inside a funded org, AI usage optimizes for what gets rewarded, and employees are paid in attention. A forty-page generated deck photographs like the future. Token consumption becomes the new lines of code, legible to a boss and correlated with none of the numbers a business lives on. The owner is paid differently. Every hour a machine takes off my plate lands on my P&L, not on a performance review. I have nobody to look busy for, so I am free to want what everyone should want: to be fired from as much of my own job as the machine can take, so I can spend the hours on the work and people I love. The firing is only half the trade. The same machine multiplies what's left, ten times the depth on research, a hundred times the ground covered in a morning, and the judgment hours get sharper because they land on better inputs. Fire me from the tedium, and make me better at what remains. I keep a sticky note taped to my monitor for the bogged-down moments, the places I get stuck. It reads: how can AI help me do this? That's the entire strategy, on one square of paper. That is why this dividend is landing first with owner-operators and bootstrappers, and why it keeps missing the org charts, where the incentive to perform work beats the incentive to eliminate it.

Did AI write this?

Sure, in the same way my agents run the market desk. I wrote and wrote, messy and unorganized, and then put the draft through the machine I've been describing. An editor agent enforces my rules and strips the patterns AI writing is known for. A voice pass, trained on years of my own emails and writing, pulls the tone back toward mine. And my own hands make the final edit, taking some suggestions and leaving others, until the idea lands in a voice I recognize. The ideas are mine, as much as anyone can own any idea today anyway. The discipline that got them to you clean is the factory's. This piece is its own demonstration.

The final leg of the stool

Da Vinci gave us the notebook. Ford gave us the line. Between them sat four hundred years of solving the physics, the leg of the journey nobody romanticizes. Cheap intelligence solves the physics on demand, and what that produces is a third kind of operator, one who speaks both languages: breadth of ideas held the way Da Vinci held them, execution discipline enforced the way Ford enforced it, with agents doing the engineering in the middle. One person, one desk, running what used to take a floor of people. I'm not describing a future. I'm describing my morning. The career that trained me for it was ordinary: approve the bid, check the work, reconcile the numbers, evaluate what the engineers shipped. AI changed one variable in that job, the leverage. And the leverage is the hopeful part, because it reaches well past operators like me. A rising tide is supposed to lift all ships. This one can reach the driftwood, the pieces the last economy wrote off.

Which leaves the question worth taking back to your desk: where do you and your business sit? Apply intelligence and this is your renaissance, the notebook finally connected to the line. Ration it, billing hours for what a machine now does, renting software by the seat, and the revolution lands closer to home. Most businesses I know have a foot on each side. The honest way to find your footing is Munger's: ask what your AI spend is rewarded for. Two numbers settle it: revenue per token and cost per token. Your token bill is already helping you answer.