The Model Is the Cheap Part

Last week I spent an afternoon presenting to private equity partners who buy real companies for a living. A couple of them emailed me before I'd even packed up my laptop, so the talk landed. And most of them still aren't moving on AI in any real way.

I live in a bubble. Everything I read is agents, agents, agents. So it's a useful jolt to sit in a room with people allocating serious capital and realize the actual economy hasn't moved much. We're early.

Here's the part I think most people miss. The models are good enough. I'm convinced of that. The holdup is everything around them.

AI runs at the speed of your weakest link

AI runs at the speed of your worst process, your dirtiest data, and your least willing employee.

An agent can be brilliant and still stall out on a business that keeps its numbers in 3 filing cabinets and the owner's head. You can't prompt your way past a bad process. The model just inherits the mess faster.

So when adoption stalls, the AI is rarely the problem. The weak links it's standing on are:

• Process. The workflow was never written down, just lived in someone's habits.
• Data. It's scattered, inconsistent, or trapped in formats nothing can read.
• People. Somebody has to actually change how they work, and most won't unless you make them.

I had it easy, and it still wasn't easy

I know this because I built Searcher OS solo, a profitable SaaS product with 0 engineers on payroll. The code gets written by an AI coding agent while I review instead of type. Bob handles strategy, Bridget handles support, the rest keep the lights on. None of them are people.

But all of it was greenfield. No legacy process to redesign, no messy data to clean, no team to win over. I got to build on a clean slate, which is about as forgiving as it gets.

And even then, my deal agent Marcus choked in his first month. Hallucinated numbers, missed obvious red flags. The fix took weeks of wrapping him in a real workflow and pulling the parts that needed to be exact into actual declarative code.

That's me, on a clean slate, with no one to convince. Now picture doing the same thing inside a 30-year-old company with real employees and 3 decades of accumulated process.

The model is the cheap part

The model is the cheap part. The change management, the process redesign, the data cleanup, that's the bill.

Nobody quotes you that number up front. The demo looks magic. The pilot looks magic. Then you try to run it across a real team on real data and you discover the work was never about the model at all.

This is most of what I do now on the advisory side. I sit with operators who are dreaming about what AI does to their margins, and the honest answer is usually: budget for the boring work first.

I'd bet the buyers and operators who win this round are the ones who do the unglamorous work of fixing the weak links, so the agents have something solid to stand on. The fancy part is cheap. The foundation is where the money goes.