Headcount Is The Moat

A founder caught me after class at UT Dallas the other day. We’d met at an event before and he wanted to walk me through what he’s building, which is a system of record for gyms and wellness centers. CRM, scheduling, membership billing, reporting, all of it in one place and all of it AI-powered. The pain is real. Independent gyms run on four or five pieces of software that mostly don’t talk to each other, and the owner spends Sunday nights reconciling them by hand. He wants to prove it out in fitness and then move into adjacent service businesses.

People will pay for this, and it works as a business. He owned a gym, he lived the problem, and now he’s building the thing he wished he’d had. I spent most of our conversation arguing that he’s taking the wrong side of the trade.

There are two positions available here. Sell software to the people who run gyms, or run the gym. For thirty years the first one was clearly better, and it was better for a specific reason: building software was expensive and slow, and that difficulty was what let a software company earn more than the businesses buying from it. Owning the operation meant payroll, a lease, a license, and somebody physically present at the front desk, and none of that got cheaper as you grew.

Building software got cheap. The advantage that came from it being hard went away at the same time. The payroll and the lease and the Saturday morning are all still there, and they’re now the part of the business nobody can replicate in a week. Which means the businesses worth owning are the ones the old math told you to avoid. The software is the tool. The service is the product.

Who Holds The Bag?

I’ve written before that the test for an application-layer AI company is whether the model provider’s own services arm makes the company more or less valuable. If your product lives entirely between a model and somebody else’s system of record, the answer is less, and what you’re really doing is renting a position until the platform decides your segment is worth the trouble.

I still think that’s the right test, but I’m going to add a slight twist to it. The question I was asking is whether the platform could do the job. The question that actually predicts anything is whether the platform would ever want the job.

Those come apart in a specific and useful way. Frontier labs and the hyperscalers behind them are optimizing for revenue that scales without headcount, gross margins north of seventy points, and liability they can define in a terms of service document. Every product decision they make runs through that filter. It’s why the services arms keep drifting toward implementation work for large enterprises and away from anything that looks like running an operation. They’ll take work a smart generalist can do from a laptop. They won’t take work that requires a license, a payroll, a physical location, or a municipal client with opinions about your staffing ratios.

So the boundary isn’t technical difficulty. The boundary is who’s willing to hold the bag. Somebody has to be there at 5:30am when the badge reader is down. No one at a frontier lab wants either of those jobs at any price, and that’s a structural preference rather than a gap in the roadmap.

Which means the durable position is the one that comes with obligations attached, and in this case that’s the gym itself rather than the software that helps somebody else run it.

Cheaper Is Not The Same As Richer

Getting your costs down and keeping the difference are two separate events. In a fragmented service market where price gets quoted per job, they usually don’t both happen. If four HVAC companies in the same metro each reduces costs 20% out of their back office, the customer has most of it back within two bid cycles. Cost advantage in a spot market disappears quickly when there’s a lot of competition.

So the question worth asking about any of these businesses is what stops the market from taking the savings back. The best answer I’ve found is a price that’s already fixed by contract.

I’ve been spending time on a company here in DFW that operates municipal recreation facilities under long-term concession agreements. Cities own the land and the buildings, the company runs the programming and the staffing, and the terms run for years at a time. It’s employee-heavy in the way that scares venture investors, because somebody has to physically be at every site during operating hours. They also reinvest in the facilities rather than stripping them, which is a large part of why they keep winning renewals.

Now imagine that company takes thirty percent out of its scheduling, reporting, and administrative load in year three of a ten year term. Every dollar of that stays with them, because the price schedule was set at signing and there is no competitor in a position to bid it away. There’s no bidding until the term ends. The savings become margin by default rather than by force of will.

The second-order effect is better than the first. Municipal contracts don’t get awarded on lowest price, they get scored on programming quality, community access, and how much the operator puts back into the facility. So the margin doesn’t have to sit there looking embarrassing on a P&L nobody sees. It can go into better programs and more reinvestment, which is exactly what wins the next RFP and the one at the facility across town. The savings compound into competitive position instead of leaking into price.

Stack that on the first argument and you get something I actually want to own. No frontier lab is going to respond to a municipal RFP, staff a facility on a Saturday morning, or carry the liability of youth programs on public land. And once the contract is signed, nobody can price against you for a decade.

Build The Thing, Don’t Buy It

The part of this that will annoy the software founders is what a company like that should do next, which is not buy an AI-powered system of record for recreation facilities.

Build versus buy lost for thirty years, and it lost fairly. Custom internal software was a liability that accrued interest. The person who built it left, the integrations rotted, nobody could read the code, and two years later you bought the packaged product anyway at a worse price than if you’d started there. Everyone who has run a small business learned that lesson the hard way and most of them learned it more than once.

What changed is the maintenance calculus. When rebuilding a workflow takes a week instead of a quarter, you stop maintaining things and start replacing them. Custom software became disposable, and disposable software doesn’t accumulate the debt that made the old trade bad. That’s the mechanism under all of this.

The practical version is unglamorous. Keep the scheduling software you already have. Export the data. Point a good generalist and a model at the questions that actually run the business, which are utilization by hour and by site, program mix against local demand, staffing against attendance, and which contracts are drifting toward a bad renewal. You skip multi-tenancy, the settings page, the permissions model, SOC 2, and the migration path from whatever the customer used before, which is a large share of where a vertical SaaS company’s engineering budget goes. You’re not making software safe for strangers. You’re making it work for one operator who knows precisely what he needs it to do.

Where is AI Going?

I haven’t written a check into an AI tooling company in over a year, and the reason is that the ground won’t stay still. Whatever a company builds on top of today’s model capabilities is a wasting asset, and I don’t know how to underwrite a product whose central advantage might show up as a line item in somebody’s API release notes by spring. That’s a genuine problem if you’re the vendor. You have a roadmap, annual contracts, customers who bought a promise about next year, and a competitor who started six months behind you on better frontier models and is therefore ahead.

The operator has none of that problem, because he was going to throw the thing away anyway. Rapid obsolescence is a tax on everyone selling software and close to free for everyone using it, since the in-house build has no customers to disappoint, no versions to support, and no migration to manage. The instability that makes me cautious about the vendor is the same instability that makes the internal version work. I’d been treating the pace of change as a reason to stay away from AI, and I think it’s really an argument about who should be doing the building.

Where That Leaves My Own Check

The truth of this thesis is that a lot of the value lands in a shape venture doesn’t fit. Consolidating a fragmented market takes acquisition debt, an operating bench, and a seven year hold, and lower middle market private equity is built for exactly that. I’m not going to out-compete a sponsor on a roll-up and I’m not going to pretend otherwise.

The contract-protected operator is a different animal, though, and it’s the reason I’m still intrigued. Those companies are small, they don’t need much capital, they grow by winning bids rather than by buying competitors, and they’re invisible to sponsors until they’re large enough to be called a platform. That’s a window where a check my size and some help thinking through the in-house build is worth more than it would be almost anywhere else. It requires being genuinely close to the operator and comfortable with a business that has people in it, which rules out most of the people I’d otherwise be competing with for the deal.

Which brings me back to the founder from class. He owned a gym and sold it, and the software is what he wants to do next, on the theory that it’s the better business. What I told him is that building software has gotten a lot easier, which sounds like good news right up until you notice it got easier for everyone, including the gym owner who might otherwise have been his customer. He already held the thing I’ve spent this memo calling durable and traded it for the thing I think is exposed, with his own proceeds, which is a real vote and not a hypothetical one. He’s also run a gym and I never have, so it’s possible the grass really is greener over here… I guess that’s something he’ll find out.