A lot of companies are adding AI to what they do. That’s one kind of challenge. This post is about a different one: companies where AI is the core of the product itself.
Wisdom Partners Co-Founder Charles Jolley has led several of these, including one AI company he built over four years and sold to Facebook. The biggest lesson from that work is simple to say and hard to absorb. An AI company doesn’t run like a normal product company, and founders who try to run it like one get into trouble.
Science is different from engineering
Most software companies are engineering businesses. You know what you want to build. You write a spec, plan the milestones, and do the work. The uncertainty is mostly about execution: how long it will take and how well you’ll do it.
An AI company is part engineering and part science. And science works differently.
Engineering starts with a known end state and finds the path to it. Science starts with a process and discovers the end state along the way. You run experiments, learn what works, and only then decide what to build. You can’t fully predict what an AI system will be able to do until you try it.
That’s the first thing founders entering this space have to learn. If you plan an AI company like a product roadmap, you’ll keep promising things you might not be able to deliver.
Climbing a hill in the fog
The best way we’ve found to describe building an AI business is hill climbing.
You build the best version you can. Then, a research team tries hundreds of small changes, keeps the ones that make the system better, and gradually improves it. Customer usage feeds new examples back to the research team, which drives the next round of improvements. Then you repeat the loop.
Picture the hills around the San Francisco Bay Area when the fog rolls in. You can tell there’s a hill in front of you. You can’t see the top. You don’t know how big it is, or whether there’s a peak at all. You just know which way is up.
That’s what building an AI company feels like. You’re moving in a clear direction without knowing exactly where you’ll end up.
The upside of this uncertainty is real. Hill climbing is hard to copy. A competitor can see exactly where you are, yet they still have to start lower on the hill and make their own climb. Google is the classic example. It was one of several search companies when it started. It had a better algorithm, but its real advantage was a pipeline that kept improving search, year after year. Even competitors willing to spend billions couldn’t catch up overnight, and by the time they’d climbed, Google was further up.
Give yourself many ways to win
Because you can’t know in advance what will work, the best strategy is to pursue several promising directions at once.
Say you’re building AI for insurance companies. Following the approach we described in Build Shoulder-to-Shoulder, you sit down with customers and find the places where they’re most constrained and where AI could plausibly create big leverage. Usually you’ll find three or four: intake for new customers, getting quotes from carriers faster, issuing demand letters, and so on.
You don’t know which one you can solve first, or best. So you explore all of them, run fast experiments, and focus on whichever experiment shows the most promise. Your product ends up shaped as much by what the technology can do as by what the customer asked for.
It’s a bit like mining. Your research team takes samples all over the property. Your engineering team builds the mine where the samples look best. Sometimes, you go looking for silver and find a huge vein of copper instead. When that happens, you follow the copper.
Organize around a deeper pipeline
A typical software company has product and engineering. An AI company has a longer pipeline:
Research runs experiments and discovers what’s possible.
Engineering and infrastructure turn those discoveries into something reliable and repeatable.
Product figures out how to package what works into something customers will use and pay for.
Each group runs on a different rhythm, and they have to stay connected. In one AI company Charles led, all-hands meetings rotated across all three. Researchers shared new discoveries, like a pattern they’d noticed in the data and an early system that could extract it. The product team listened and started imagining ways to bring it to customers. Engineering worked out how to make it work at scale.
Measure progress by what you unlocked
This changes how you set goals. In a traditional company, you can set a target and measure whether you hit it. In an AI company, many experiments will fail, and that’s expected.
The better question is: What did we unlock? What did we move from unknown to known? Which risks did we turn from mysteries into problems we can actually manage?
Once you’ve closed that gap, you can push hard and move fast. The problem is that you can’t predict how long it will take to close it. That’s why internal clarity matters so much. Your team needs to know the difference between what you’ve proven and what you’re still exploring, especially when customers start asking for delivery dates.
Ownership culture matters even more
All of this makes ownership culture more important, not less.
When the path forward depends on experiments, you can’t direct every move from the top. People need clear outcomes to pursue, the freedom to run their own experiments, tools that show them when they’re making progress, and fast feedback loops. The faster your team can learn, the faster you climb.
AI companies can be among the most defensible businesses being built today. But only if founders accept the kind of business they’re running. You’re climbing a hill in the fog. Plan for discovery, build for fast learning, and let the climb shape what you build.