AI Turned Being Messy From a Tax Into a Countdown

Every sports movie has the same scene.

Rocky runs the steps. Daniel waxes the cars. The Cool Runnings crew learns to work together.

The music kicks in, calendar pages fly, and three minutes later our hero is someone new. (I will die on the hill of Rocky IV being the best one. Rocky drags a sled through Siberian snow while Drago runs on a treadmill wired to computers, and somehow the entire Cold War is in there.)

Forty years of these movies, and nobody has ever filmed the other version: the hero writing a check so someone else can do the training montage for them. Because it wouldn't work, and everyone in the theater would know it. Some things transfer when you write a check. Some things you have to go through yourself.

We've never had to apply that to business. For forty years, every technological wave could be bought. AI cannot, and I can show you the data that makes it clear.

The dinosaurs were doing great

Picture a business. Any business. It may be yours, or someone else's.

There's a few hundred people, the org is 25 years old, profitable for almost all of them.

The operational backbone is a spreadsheet called MASTER_v14_FINAL.xlsx.

Three people understand it and one retires in the spring. Data arrives by email and gets rekeyed by hand into three systems to varying degrees of quality. Two if Patrick's out on vacation.

You know what I'm supposed to say next. Dinosaur, meteor, adapt or die.

Except the company grew 8% last year and the owner sleeps great.

I run a firm that builds AI and software for exactly these companies, so I have every commercial incentive to tell you the meteor is coming. But I've read thirty years of "thought leaders" saying that, and the spreadsheet companies kept winning, kept compounding, quietly outlived the consultants who wrote their obituaries.

The predictions failed because they misdiagnosed the duct tape.

Inefficiencies just cost something: five or ten margin points, paid out in rekeyed data entry and the two weeks it takes to answer a question the data could answer in two minutes if it lived somewhere besides Patrick's head.

But every competitor was bleeding the same points on their own version of inefficiency. "Use technology better or keep paying the tax everyone else is also paying" is the honest version of every digital transformation keynote from the last twenty years, and it explains why nobody drowned.

That was a tax. Universal, priced in, survivable forever.

That era is ending, and this time somebody measured it.

The tax is becoming a countdown

Two business school professors, Rem Koning at Harvard and Hyunjin Kim at INSEAD, just published a study of how companies built around AI are organized. Every Y Combinator startup since 2020, then 45,000 US venture-backed companies, AI-native firms compared against regular firms in the same industry, founded in the same year.

The headline: AI-native companies run 25% smaller with the same valuations. The number that should stop a non-tech company cold is buried deeper. In businesses that historically grew by hiring people to do knowledge work, the AI-native firms operate at roughly a third of the headcount of their peers. An AI-native exam prep company, for example, runs on 7 people. The company selling the same outcome through human tutors has 912.

These studies are on technical startups. But for the spreadsheet-run company's whole lifespan, the obvious assumption was that the tax of inefficiency was universal. Few could exploit it because almost everybody paid it. That assumption is quickly fading as a competitor can now rebuild your operation without your headcount, and the tax becomes a countdown in this AI era.

A tax is comfortable because it's stable. But the stability is fading as a gap is growing from both ends: AI-native companies compound automation and efficiency while your manual way falls a little further behind every quarter.

Sized honestly: nothing happens next quarter, probably not much next year. But you won't see it coming in their marketing. You'll see it in their pricing, two years after it's too late to respond.

The check-writing playbook

Every previous wave was a purchasing decision. ERP, cloud, the list goes on: write the check, hire the integrator, stay out of the way. That playbook worked, and it's why the spreadsheet company survived every obituary.

But this era is where the study gets uncomfortable.

The researchers split "using AI" in two: companies whose job postings name the tools their people use (ChatGPT, Claude, Cursor), and companies that rebuilt what they deliver around AI. The tools predicted nothing. Firms naming AI tools look structurally identical to firms that never mention AI. The entire difference came from companies that rebuilt the work itself.

Buying the tools is the one move the old playbook knows. It's also the move that measured as zero impact.

You've met the people who have run the "shiny new tool" playbook. They feel credible and sharp about the technology. They may have taken over a 400-person company but have never once sat with the AP team. They bought a few hundred Copilot seats, announced it at the all-hands, and eighteen months later told their board the technology isn't ready. It was ready. They aimed the zero play at a company still running old operational playbooks.

Now the owner I actually think about. Regional logistics, couldn't tell you what an API is. But he spent nine years as a dispatcher before he owned the place, and he can walk his quote-to-cash from memory. His team knows which dispatch calls take judgment and which follow a pattern they could recite in their sleep, because it used to be their sleep. When AI showed up, they pointed at the right problems and paid someone to wire it up.

The difference between them was never technical aptitude. It's whether you know your own operation at a deep level. The knowledge AI needs isn't sitting at Accenture. It's which processes actually exist versus which live in the SOP nobody follows. Which numbers describe reality and which are theater. Which report is produced weekly because an exec asked in 2019 and nobody thought to question the request.

Most teams' operations run on undocumented facts, and AI only reads the documented ones. Someone has to close that gap, and it can only be someone who lives there.

You can write a check for the gym, the trainer, the equipment. You can hire Mickey. The reps are still yours.

You can and should delegate building the solution. What you can't delegate is knowing your own company. Sweat equity is in the mapping, not the plumbing. That split is what makes a project work: when we built one operating layer for a cross-border medical transport company — intake, live quoting, mission coordination, billing — the plumbing was ours, and knowing which parts of the operation actually needed to change was theirs.

Ten honest minutes

The opportunities are huge but knowing where to start is intimidating. But you don't need an AI readiness assessment. Start with these:

• Can you describe your quote-to-cash, end to end, without asking anyone?
• Which of your reports exist because someone uses them, and which exist because someone asked once and the request never died?
• If your most tenured ops person quit tomorrow, what breaks? Could you actually list it?

If those answers come easily, you're going to be fine with AI transformation. If they're hard, you've found where your planning should start — and that mapping is the first half of an AI Jumpstart.

Delegate the wiring, keep the advantage, ignore the doomers. The efficiencies that lead to better margins and happier teams (not stuck moving data around systems and spreadsheets) existed long before ChatGPT. AI didn't create new opportunities, it just attached a solution that is more attainable for every size of business.

Cue the music

The spreadsheet company I described at the top didn't change while you read this, but its position did. Researchers just measured its future competitor: a third of the headcount, flatter, more senior, worth just as much, if not more.

The owner who wins in this era doesn't need to understand AI. They need to know their own company well enough to point at what should change and the guardrails around that. The building — what we do — becomes easier. The systems, the automation, the software that replaces Patrick's spreadsheet without losing what Patrick knows.

My honest bet is that the least technical leaders in the mid-market will quietly win the next decade, while the ones who think AI is a purchase stand around wondering why nothing happened after they bought the Copilot seats and called it a transformation.

Nobody wins a title fight alone, and the montage has always been available to anyone willing to run it. The difference is that skipping it used to cost a tax. Now it starts a countdown.