Meet the Slowest Startup Incubator in the World—Pumping Out Billion-dollar Companies

Dan Shipper interviews Sam and Dan, the founders of Bolton & Watt, a startup incubator that deliberately moves slowly — launching one new company every two years, operating it themselves to millions in revenue, then handing it off. Their portfolio includes a Series C med spa platform and California’s largest funeral services provider, neither of which existed before they personally built them from scratch. The conversation reveals a sharp distinction between “AI-native” and “AI-durable” businesses, with Bolton & Watt firmly planted in the latter camp. They’ve built custom AI agents to accelerate their company discovery process, but found hard limits on what AI can replace — particularly when it comes to validating whether real humans will actually pay for something. Their experience deploying AI inside operationally complex, real-world businesses offers a grounded counterpoint to the breathless transformation narratives common in tech circles.

The interview also surfaces a practical framework for getting teams to adopt AI without creating what the founders call “AI participation trophies” — rewarding output quality rather than tool usage, seeding examples through power users, and maintaining the expectation that AI is table stakes, not extra credit.

The Bolton & Watt Model

Bolton & Watt is named after the 1775 company formed to commercialize the steam engine — a deliberate signal about how they think about technology and business. They call themselves “the slowest incubator in the world,” and the model is radically different from the typical incubator playbook.

Where a traditional incubator might fund many ideas and coach from the sidelines, Bolton & Watt operates on extreme concentration. The founders generate the idea, then personally run the company as operators until it reaches $5–10 million in revenue. Only then do they recruit a CEO to take it to the next stage, remaining involved as active board members and partners who have spent thousands of hours in the competitive landscape.

They aim to start one new company every two years. Their two constraints on speed are finding great ideas and finding great talent — they won’t compromise on either. They’ve been asking themselves whether they could compress the cycle to 18 months or even one year, and have recently brought on a partner who specializes in the concept development phase so that when one founder rolls off an operating company, they’re rolling into a validated idea rather than a cold start.

The Tag-Team Operating Model

The founders tag-team: one runs the current company while the other starts the next. They sit next to each other, maintain full context on each other’s work, and serve as thought partners and emotional support. As one of them put it: “We have all the upside from co-founders but also a much larger scope and span.”

They’ve also started to institutionalize and specialize across the different phases of company building — concept development, early operations, scaling — recognizing that the skill set required at each stage is fundamentally different, even though the startup world pretends it’s the same all the way through.

Why Slow Is the Point

One founder previously built a company over 10 years, and years 4 through 10 brought a constant existential question: “Am I doing the thing that is most interesting and most useful? Am I spending my time the right way?” The Bolton & Watt model is partly a solution to that — by specializing in years 0 through 3 across multiple companies, they stay in the phase they find most energizing and where they believe they create the most value. They’re also accumulating reps in early-stage building that very few people get, seeing what success looks like on the other side over and over again.

Current Portfolio

1. Moxy

A “business-in-a-box” platform for nurses to open their own medical spas. Moxy provides back-office support — compliance, growth tools, and everything needed to run a med spa — so nurses can focus on aesthetics: Botox, filler, lasers, and similar treatments.

• Stage: Series C company
• Revenue: Tens of millions
• Customers: 600+
• Team: ~200 globally
• Notable timing: Launched a few months before ChatGPT’s release, making it a “just before AI” company that has had to adapt rather than being AI-native from conception

2. Contemporary Funeral Home

A funeral services provider with zero physical real estate. All arrangements are made online or over the phone. In-person services are held at wedding venues — booked a year in advance, taking advantage of the fact that venues are fully booked on Saturday nights but wide open on Tuesday mornings.

• Position: Largest provider of funeral services in California
• Expansion: About to launch in multiple new states

Both businesses reflect Bolton & Watt’s taste for “weird” businesses that aren’t YC-zeitgeistic — niche verticals that combine software, services, and a real-world component, requiring the founders to build what is essentially a services company and a software company simultaneously.

Investment Thesis: AI-Native vs. AI-Durable

Bolton & Watt draws a sharp distinction between two types of companies worth building right now:

1. AI-Native: Companies that push the technological ball forward within a category
2. AI-Durable: Companies where the core mechanism of the business remains unchanged by AI, but can use AI to become more efficient

They focus exclusively on the AI-durable category. Their reasoning: the core value delivery of a med spa (a physical injection) or a funeral home (an in-person service) cannot be digitized or automated by an LLM. There’s no robotic injector coming in the next 7–10 years. AI is an accelerant for the operations around these businesses, not a replacement for the business model itself.

Med spas are actually conveyors of the latest medical technology — they were early adopters of GLP-1s, for instance — but what happens inside the walls of a med spa is not deeply impacted by AI. The opportunities are at the edges: reaching the right customers, serving them effectively, communicating at all hours at an affordable price.

The founders noted that every idea they’ve had in the AI-native category has had multiple formidable competitors doing something almost identical. In the AI-durable space, the competition looks different — there aren’t 10 YC companies launching in each vertical. As one founder put it: “All the smart people are listening to Dan and Every, and we’d rather compete against the less smart people.”

The Intersection of AI and Secular Trends

Rather than chasing AI as the primary trend, Bolton & Watt looks for the intersection of AI with some other secular change — rising death rates, growing demand for aesthetic treatments, shifts in how services are delivered. Their next company (which they declined to reveal, citing “literal day zero” levels of embarrassment) follows this pattern.

AI in the Company Discovery Process

The company discovery process is where Bolton & Watt has seen the greatest AI-driven transformation — consistent with their observation that the more greenfield the work, the better AI performs.

Phase 1: Vertical Research

The first step in finding a new company to build is identifying verticals worth exploring. This used to be a week of Googling and calling friends to establish basic facts. Now it’s a series of mega-prompts that generate a list of categories, assess them against Bolton & Watt’s particular point of view on what makes a good business, and narrow to a few areas worth talking to real people about.

They ran an experiment where both AI curation and human judgment independently evaluated verticals. Out of three categories selected by the group, one came from the human point of view and two came from AI-identified opportunities. The founder who did the human research described the experience as feeling like John Henry competing against the tunnel-boring machine — “and fortunately came out the other end.”

For basic market and industry reports, AI now beats the low and medium quality research that was previously available. It’s particularly good at parsing public company filings — thick PDFs that contain valuable signal about a category. But the founders draw a clear line: AI is excellent at surfacing the consensus opinion about a space and uncovering facts, but it cannot generate the differentiated point of view that comes from years of earned experience. They’ve spent two years after building Moxy looking for another category where a Moxy-style business might work, talked to roughly 60 people starting “business in a box” companies, and developed a relatively unique understanding of what makes that model succeed. AI can consume massive amounts of information and fit it to their point of view, but it cannot produce the point of view itself.

Phase 2: The “Matthew Bolton” Agent

The AI work eventually forced a migration from Google Docs to Notion, which one founder had been resisting for years. Inside Notion, they built an agent identity called “Matthew Bolton” (named after the historical Bolton of Bolton & Watt) to assist with customer discovery. They demonstrated the system using their active exploration of PNC (property and casualty) insurance as an example.

The system is organized around several linked documents:

• Point of View (POV): What they think the opportunity is, the problem they’re solving, who they’re serving
• Hypothesis Tracker: The core things that must prove true about the category, what they’re looking to discover through customer calls
• Transcripts and Notes: Records from every discovery call

Matthew Bolton operates in two modes:

Pre-call preparation: The agent reviews the target persona, current hypotheses, and validation focus areas, then suggests specific areas to dive into during the conversation. The founders review and adjust, but the prep is significantly more efficient.

Post-call synthesis: Call transcripts go directly into a Notion table. Matthew Bolton then re-reads the latest POV, the hypothesis tracker, and the most recent 10 calls. For each hypothesis, it regenerates a point of view on what evidence supports it, what evidence contradicts it, the strength of the hypothesis, and where to spend more time. It pulls relevant quotes from the transcripts.

The system serves a deeper purpose beyond efficiency. Starting a new company requires “a manic energy and a little bit of a suspension of disbelief — there’s just no reason any new company should succeed.” Matthew Bolton keeps the founders rigorous and fact-based, countering that natural optimism bias. They can “be sick offense to ourselves and ask it to remain fact-based” — balancing entrepreneurial conviction with intellectual honesty.

Where AI Research Fails: The Sycophancy Problem

Bolton & Watt tried using AI as a proxy for actual customers — synthetic customer calls where the AI plays the prospective buyer. It failed completely.

The problem: no matter what they did to combat sycophancy, the AI invariably expressed a 10 out of 10 desire to buy the product. Every idea that passed even a basic sniff test got enthusiastic AI endorsement. They tried a synthetic customer validation, then flew to meet a real prospective customer and “fell flat on our face completely.”

The AI doesn’t know the nuances of the psychology of someone who has worked in an industry for 15 years and is deciding what to buy. It doesn’t understand purchasing friction. It either knows the founders want it to say yes, or it simply can’t model the complex human decision-making process.

Their conclusion: AI can reduce the friction of preparing for and analyzing customer calls, but it cannot reduce the number of human conversations required to reach confidence in an idea. “It can help us talk to people effectively, but it cannot actually reduce the number of people we need to talk to.”

AI for High-Level Judgment

When asked for opinions on high-level strategic questions, the founders don’t trust AI’s output. It’s useful for finding quotes and evidence to support or challenge a hypothesis, but not for making the judgment call. The ideal outcome is bringing three key quotes from three different people that directly speak to each critical hypothesis — AI gets them there much more efficiently, but the evaluation remains human.

That said, they do find value in asking AI to make the best counter-argument to their current thinking. It sharpens their reasoning even when they’re not looking for it to tell them they’re wrong — it teases out “in what ways could you be wrong,” which is a useful exercise. Though as one founder admitted, sometimes the counter-argument prompts a response of “Oh, fuck off.”

Operationalizing AI in Established Companies

The “No AI Participation Trophies” Philosophy

Bolton & Watt’s approach to AI adoption inside their operating companies rejects the idea of having a formal “AI initiative.” Their reasoning: you don’t want to lead with the hammer. They recall the NoSQL era when everyone was putting everything in NoSQL whether it belonged there or not, or the period when everyone built a Slackbot whether one was needed or not.

Instead, their framework:

• No credit for using AI. Nobody gets points for generating copy that was clearly AI-written and bad to read.
• Credit for output quality. The expectation is that people deliver the best possible product, and it’s assumed that using AI is necessary to achieve that standard.
• Seed through power users. Find the people on the team who are naturally pushing boundaries, have them showcase their work in public forums, and use that to raise the floor for everyone.

The concern this addresses: without active management, people will default to doing things the way they already know, reasoning that their established process produces the best quality. The counter is comparing work quality across the team and making visible what’s possible — not mandating tool adoption but making excellence the standard.

The Greenfield-Legacy Divide

A consistent pattern across Bolton & Watt’s portfolio and their founder network:

• Seed-stage founders: “Our engineering is 10x faster”
• Series D founders: “We’re like 10% faster. What is everyone talking about?”

This maps directly to the greenfield vs. legacy distinction:

Where AI excels:

• New apps, landing pages, isolated tools — anything built from scratch
• Research and development phase work
• “Throwaway” apps that solve specific problems without full engineering deployment cycles

Where AI provides incremental improvement:

• Mature codebases (Series C level, ~40-person product engineering teams) with complex integrations
• Deployment, maintenance, and integration work
• Production systems that need to fit with existing headers, footers, design systems, and infrastructure

Practical AI Wins

Throwaway apps by non-engineers: A designer at the funeral home built an app where anyone can type in a hospital name or town name and it resolves whether the company services that area. Rather than spending engineering time on safe deployment and integration, it exists as a separate app accessible via link. This pattern — enabling non-engineers to build functional internal tools — has been one of their most successful AI use cases.

SamGPT for recruiting: The talent team built “SamGPT,” trained on one founder’s blog posts, to draft personalized recruiting outreach on LinkedIn that mimics his voice. The founder forgot it existed until the team reminded him — a sign it had been working quietly and effectively in the background.

Landing page generation: AI is excellent at generating landing pages and pushing design thinking, though integrating results back into Webflow with consistent headers, footers, and design systems remains a significant amount of work.

AI and Marketing Channels

The funeral home has seen increased traffic from ChatGPT, but the founders view it as just another channel — similar to paid search, with a cat-and-mouse game around organic results and an inevitable paid version. They don’t believe people will buy a funeral via chat, just as many people still prefer to search flights themselves rather than use a travel agent. It hasn’t fundamentally changed their marketing approach; it’s simply a new channel to stay ahead of.

The Reality Check on Advanced Models

Despite the release of advanced models (referenced as Claude Opus 4.5 in the conversation), the early word from Bolton & Watt’s engineering teams is “this is better, but not a step function different experience.” They’re doing work to retool and capture more benefit, but for a 40-person product engineering team working on a mature codebase, the night-and-day transformation reported on X has not materialized. The fundamental work of deployment, maintenance, and integration remains the constraint.

---

The interview closes with the founders declining to reveal their third company — they’ve only known what it is for about five days and are nowhere near launch. They promise to return once it’s public, and the hosts note their parallel paths building weird, different businesses from neighboring offices in Boerum Hill, Brooklyn — a relationship spanning 13 years in the New York tech scene.

Meta

Added: 2026-03-08