Teamday
← Back

Boulton & Watt's Sam Gerstenzang on Selling Risk, Not SaaS

Stripe
businessstartupfuture-of-workautomationinterview

How Boulton & Watt Rebuilt Its Companies Around AI Coding

Sam Gerstenzang is co-founder of Boulton & Watt, which Stripe co-founder John Collison introduces as "the world's slowest startup incubator." It starts one company every few years, runs it until well past product-market fit, then hires a CEO. The first is Moxie, which helps nurses and doctors open their own med spas. The second serves funeral homes and the third serves independent freight brokers. Gerstenzang, who used to work at Stripe, tells Collison what changed in those companies once AI could write the software, and what didn't.

Start from the customer's P&L, not from a product. Boulton & Watt looks at where a small business earns and spends money and tries to take over all of it. For Moxie that began with legal setup, which otherwise costs about $20,000 in lawyer fees per state. Then came marketing, payments software and supplies. Med spas spend about 30% of revenue on Botox and other pharma, and Gerstenzang says Moxie is now one of the largest Botox buyers in the US. Moxie charges 13%, and the sales pitch compares a customer's P&L before and after: with the savings, he says, it will "feel like it costs negative 2%."

Custom software for each customer: One of the freight brokers works out of his customer's office in Louisiana, where three brokers compete for every job through the customer's own portal. Boulton & Watt spent two hours building software that automates that portal work for him. "Now we're sort of building this flexible custom software for each customer." Collison asks how that scales and how it avoids becoming unmaintainable. The answer is a fixed set of backend APIs with a new front end for each customer, and a software engineer reviews everything that ships to customers. Customers can't yet prompt changes themselves. Gerstenzang thinks that is where the maintenance problems would start.

One product manager per engineer: Stripe ran at about one PM to six or eight engineers. Boulton & Watt's two newer companies run at one to one, and the first company, started before AI coding, does not. Total spending on software development is about flat and the number of engineers is the same. What changed is the unit: a team of 13 building one feature became pairs who own an area and go deep. "Part of the reason the one-on-one team works so well is there's no one else to talk to." Deciding what customers want and managing the rollout are still hard. Only the building got easier.

The four-year-old company needed two refactors: The first was in the code, where one engineer knew what everything was named and the codebase had to be restructured so AI could understand it too. The second was in the people. "Log out of Slack, shut your computer, you're joining a new company." Gerstenzang says the newer companies are less collaborative in a good way: people get the data themselves, and nobody waits for an engineer.

Where AI disappointed: Moxie had built a large human team to answer front desk calls for 100 med spas, then launched a voice AI version. Gerstenzang says it was actually better than the human team. "Our users were so much less tolerant of mistakes." Owners understood why a person got something wrong and expected the machine to be perfect. He also says AI summarizes well but can't synthesize, for example to separate user segments or propose a new business. Adoption inside the existing companies took far longer than he expected because of data integration and change management.

6 Takeaways from Sam Gerstenzang on Building Companies with AI

  • Sell risk, not software - Customers pay to offload compliance, credit and delivery risk. "No nurse is going to vibe code her EMR." She cares too much about HIPAA and her license.
  • Software quality no longer sets companies apart - Gerstenzang says the variance between software teams is collapsing, so pure software businesses need scale and buying power or a risk worth paying for.
  • Systems of record are exposed from both sides - Cheap point solutions take the low end and businesses that do the actual work take the high end. Once agents operate the data, he asks, does anyone still need to log into Workday?
  • Review standards differ by audience - Internal apps get a lighter standard. Everything shipped to customers is reviewed by a software engineer.
  • Nobody was fired and legal costs didn't fall - The companies now run contracts past AI that nobody used to check, so they consume more legal services. Term sheets still go to a lawyer.
  • Go-to-market is physical - Moxie takes nurses shoe shopping at Jimmy Choo and is opening a training center for injectors in Los Angeles, taught by people with TikTok followings.

What Boulton & Watt's Model Means for AI-Era Software Businesses

Gerstenzang's argument is that software became the cheap part. What customers still pay for is someone who takes responsibility for the outcome, such as the license, the payment or the truckload. AI lets a small team cover far more of a customer's P&L, but only on a stable backend, with human review, and with changes rolled out to people slowly. For the next company he is looking at restoration after burst pipes, fire safety and pet hospice. In each he sees demand that will grow without him, and a part of the job where being wrong costs someone a license.