I started building with OpenAI in 2020. Jasper in 2021. Most people met AI in late 2022, when ChatGPT turned it into a headline. By then I had already spent two years quietly running content, operations, and analytics on it, and, just as usefully, two years making the mistakes.
I do not say this to plant a flag. I say it because a five-year head start turns out to be a head start on the wrong lessons as much as the right ones, and the wrong lessons are the more valuable ones to share.
Here is what those five years actually taught me.
1. The moat was never the model.
In 2020 the model was the exciting part. Every few months a better one shipped, and it was tempting to believe the company with the best model would win. That is almost never how it played out. The model is a commodity that gets cheaper and better while you sleep. What compounds is everything you wrap around it: the workflow you redesign, the proprietary data you feed it, and the distribution you already own. I have watched teams with mediocre models and excellent workflows quietly beat teams with the reverse. If your AI advantage disappears the moment a competitor swaps in the same API, it was never an advantage.
2. "AI-native" is a decision you make before you hire.
The companies that got real leverage from AI did not bolt it on afterward. They asked a different question at the start: what would this team look like if agents did the operational work by default? That question changes who you hire, what you build in-house, and how fast you can move. "AI-native" is not a feature you add in a sprint. It is an operating posture you adopt before the org chart hardens, because it is much harder to retrofit.
3. Most AI projects fail for a boring reason.
They bolt a chatbot onto a broken process. The process was the problem; the chatbot just makes the broken process faster and more confident. The work that pays off is unglamorous: pick one high-volume, low-judgment workflow, map what actually happens, and redesign it around what a machine is genuinely good at. Then automate. The order matters. Automation applied to a mess produces a faster mess.
4. Give the agent real tools, and keep a human on the 5% that matters.
A model with a clever prompt is a party trick. A model wired to real tools, to your systems, your data, and actions it can actually take, is infrastructure. The unlock in the last two years was not smarter models; it was giving them hands. But hands need supervision. The teams that trust agents blindly get burned, and the teams that supervise everything get no leverage. The answer is a human gate on the small slice where a mistake is expensive, and autonomy everywhere else. Finding that line is the whole job.
5. Measure output in hours, not headcount.
The old reflex when you need more done is to hire. The AI-native reflex is to ask whether the next unit of work needs a person at all, or a well-supervised agent. I now watch a different number: not how many people we added, but how many hours a task takes end to end, and whether that number is falling. A company optimized for headcount and a company optimized for throughput look identical on the surface and behave completely differently under load.
6. Early does not mean right. It means you had time to be wrong.
The real gift of starting in 2020 was not foresight. It was reps. I built things that did not work, trusted models that were not ready, and automated processes I should have fixed first. Every one of those was cheaper to learn in 2021 than it is to learn now with real stakes. When someone tells me they have been "doing AI" since last quarter, I do not doubt their intelligence. I just know they have not yet paid the tuition that only time charges.
None of this requires believing AI will change everything. I do not trade in that kind of prediction. I trade in what I can build and measure. Five years in, the pattern is clear enough to bet on: the winners will not be whoever has the best model. They will be whoever redesigned their work around agents earliest, wrapped it in data and distribution nobody else has, and kept a steady hand on the few decisions that actually matter.
I have spent 25 years as a technical founder building the software that AI now runs on: decision engines, fintech platforms, the systems underneath a global network. For the last five, I have been building with the AI itself. These days I build the AI, and I back the founders doing the same.
That is the head start worth having. Not the tools. The mistakes.
If you are building AI-native, not AI-flavored, I would like to hear about it.