Everyone is building AI wrong
with Josh Furstoss, Alchem Accelerator
Everyone is building AI wrong
Show Notes
Josh Furstoss is an executive founder and CEO advisor who launched four companies last year - and has two more in stealth. He has coached over 120 companies through zero to one at Alchem Accelerator and holds multiple exits across his career. In this conversation, he makes a sharp and uncomfortable argument: the best AI companies are not being built by the people you would expect, in the markets you would expect, using the technology you would expect. The real arbitrage is in the relationships nobody else is making.
Josh's perspective is not theoretical. It comes from touring ball-bearing factories, talking to the guys who fix the machines for $20 an hour, and building products so niche they can be described in five words. The result is a framework for zero-to-one that is as specific and unfashionable as the markets it targets.
The Three Checkbox Framework: Don't Build Without These
Before writing a single line of code, Josh requires three boxes to be checked: distribution on the cap table, defensible IP, and an industry expert involved. If you cannot check all three, the project does not start. Not pauses - does not start.
Distribution on the cap table means giving equity to people who already have the relationships you need - people already inside the small networking circles of your target market. Defensible IP means proprietary data, SOPs, chemical formulas, machine configurations - the kind of accumulated institutional knowledge that a competitor cannot replicate with another model call. And an industry expert is the person who actually understands the domain, because Josh walks into every new industry deliberately as a "stupid person" - no assumptions, no imposed solutions, just the question of what works.
The Lower Middle Market Data Moat Nobody Is Talking About
Cabinet factories. Paint mixers. Ball-bearing manufacturers. Companies doing $100M a year, often still owned by the two brothers who built them 30 or 40 years ago. They are sitting on decades of proprietary data - internal SOPs, formulas, machine configurations, operational playbooks - and almost none of it has been touched by software engineers. Not because the opportunity is not there, but because software engineers and cattle ranchers do not hang out at the same bars.
Josh's move: tour the factory. Skip the sales team. Go find the person making $15–20 an hour who actually runs the machines and fixes things when they break. That person knows where the value is. The owner is happy when you spend time there. And once you understand the actual problem, you can build something so niche it cannot be replicated without the same relationships and domain access you now have.
The Five-Word Rule and the $83K License
If you cannot describe your product in five words or less in H1 text on a website, it is a bad product. There only needs to be one core feature. Everything else is noise that dilutes the offer and expands the surface area you have to defend.
The play Josh describes: build one hyper-specific feature so niche you can describe it in five words. Sell one enterprise license for $83,000 a month. That is $1M ARR from what might be four lines of code packaged as an executable. His lawyer gave him the canonical example - a friend at MIT discovered that the internet's routing system was losing bits, wrote a small fix, and sold the licensing deal to telecoms for millions. The code was trivial. The insight and the distribution were not.
The Class Divide Keeping Founders Away from the Money
The kids at Stanford and MIT are building for space and quantum computing. Not because those problems are more important - but because that is who they know. Software engineers and cattle ranchers do not have overlapping social networks. The result is a class divide that leaves unglamorous, high-value problems completely untouched.
Josh's antidote is deceptively simple: be a normal person. Call someone. Ask to tour the facility. Do not try to sell anything. Show up as a peer who is curious about what they do. The owners of these businesses want to partner - they just have never been approached by someone with technical skills who was willing to show up and actually talk to the people doing the work.
Zero to One Specialization: Know Your Lane, Promote Yourself Up
Josh's operational model is built on radical self-awareness about what he is actually good at: product, attracting talent, raising capital, and zero to one. He is not the best daily sales grinder. He is not the best technical architect. Rather than pretending otherwise, he builds structures around his strengths - do the zero-to-one work, then promote himself upward to a more strategic role, hire a CEO, and let the people better suited to scale do the scaling.
This is what makes four companies in a year possible without burning out. It is not a productivity hack - it is a clarity hack. When you know precisely what you are good at and what you are not, every hiring decision and every equity grant becomes a statement about where you are weak and who fills that gap.
AI Will Become Infrastructure - And Most Founders Are Misreading the Moment
Josh's macro view: AI companies will be the new small business, not the new unicorn. The foundational models - OpenAI, Anthropic, Google - are the AWS, Azure, and GCP of this generation. Just as cloud had a handful of winners and a long tail of businesses built on top of them, AI will have a handful of infrastructure winners and millions of vertical applications. Most founders are trying to be AWS when they should be building the business that runs on AWS.
The implication for early-stage founders: putting AI next to your company name is already losing its signal. Investors increasingly assume it is built in - the same way nobody put "internet-enabled" in their pitch in 2010. The differentiation has to come from something else. In Josh's world, that something else is the proprietary data and relationships that make replication impossible.
The Post-Exit Identity Crisis Is Real - and Mostly Unavoidable
Josh took a month and a half off after his last exit. He calls it a long weekend. The identity is so tied to the company that when the company is gone, the question "who am I?" has no ready answer. He found camaraderie in PEF - a group of about 1,500 post-exited tech CEOs - and describes trips to Israel during an active conflict and python hunting in the Everglades as the kinds of experiences that recalibrate something that is very hard to name.
His advice is implicit rather than stated: find things that are not outcome-based. For him it is Dungeons and Dragons every Saturday. For you it is whatever gives you identity that does not depend on a cap table. The search for meaning post-exit is not a problem to be solved - it is a condition to be lived with while experimenting.
- The Three Checkbox Framework - Before building anything: (1) distribution on the cap table, (2) defensible IP, (3) industry expert involved. All three must be checked. If you can't check all three, don't start. See Frameworks.
- The $83K License Play - One hyper-specific feature, describable in five words or less, sold as one enterprise license at $83K/month = $1M ARR. The value is not the code - it is the insight, the data, and the relationship. See Frameworks.
- Zero to One Specialization - Identify exactly what you are best at in the zero-to-one phase. Build structures to do only that. Promote yourself upward as the company scales and hire people better suited to the next phase. See Frameworks.
- Manus - Josh's tool of choice for research. AI agent platform for autonomous task execution. "The output is gorgeous" - particularly valued for computer use capabilities other tools lack.
- Lower Middle Market - Businesses generating approximately $100M/year, often family-owned for decades, with accumulated proprietary data, SOPs, and operational knowledge that is vastly underutilized by software. The highest-value untapped market for vertical AI. See Glossary.
- Distribution on Cap Table - Giving equity to people who already have the relationships, access, and trust in your target market. Not hiring salespeople - recruiting co-builders who are already inside the networks you need to reach. See Glossary.
- Defensible IP - Proprietary data, SOPs, chemical formulas, machine configurations, and operational playbooks that cannot be replicated by a competitor running the same foundational model. The moat that makes a niche AI product acquisition-resistant. See Glossary.
- Five-Word Rule - If you cannot describe your product in five words or less in H1 text on a website, it is a bad product. One core feature. Everything else dilutes the offer and expands the surface area you have to defend. See Glossary.
- GPT Wrapper - A thin AI product built on top of a foundational LLM without proprietary data, unique distribution, or genuine differentiation. The product category most founders are building and fewest investors want to fund in 2025+. See Glossary.
- Class Divide (Tech/Industry) - The socioeconomic and cultural gap that keeps software engineers from discovering and solving unglamorous problems in traditional industries. Software engineers and cattle ranchers do not hang out at the same bars - and that gap is the source of the arbitrage. See Glossary.
- Zero to One - The earliest phase of company building: defining the product, establishing distribution, and generating the first revenue. Josh Furstoss's identified specialty - the phase requiring maximum creativity and minimum institutional inertia. See Glossary.
What are the three boxes you must check before building any AI product?
Distribution on the cap table - equity given to people who already have relationships inside your target market. Defensible IP - proprietary data, SOPs, formulas, or configurations a competitor cannot replicate with another model call. An industry expert involved - someone who actually knows the domain, because Josh walks into every new market as a 'stupid person' with no assumptions. All three must be checked. If even one is missing, the project does not start.
Why are lower middle market businesses the biggest AI opportunity nobody is talking about?
Because the people who could build the software do not know the people who own the businesses. Cabinet factories, paint mixers, ball-bearing manufacturers - companies doing $100M a year, often family-owned for 30 or 40 years - are sitting on decades of proprietary data, SOPs, and operational knowledge. Nobody is touching this because software engineers and cattle ranchers do not have overlapping social networks. That gap is the arbitrage. Walk in, tour the facility, talk to the person fixing the machines, and you have access no generalist AI startup can replicate.
What is the Five-Word Rule, and why does it matter?
If you cannot describe your product in five words or less in H1 text on a website, it is a bad product. One core feature. Hyper-specific means the problem you solve is so narrow that you can articulate it instantly to the exact person who has it. Everything else is surface area to defend. The smaller and more precise the problem, the more defensible the solution - and the easier it is to price it as a specialized tool rather than commodity software.
How does four lines of code become $1M ARR?
By being the right four lines of code sold to the right person at the right price. The example: a routing protocol fix that recovered bits lost in internet data transmission - a small executable sold to telecoms as a licensing deal worth millions. The code was trivial. The insight required understanding the problem deeply, and the distribution required knowing who had it and who could pay. One enterprise license at $83K/month equals a million-dollar ARR. The leverage is in the niche, not the complexity.
Why will AI become the new small business infrastructure rather than producing new unicorns?
Because the dynamic mirrors what happened with cloud. AWS, Azure, and GCP won the infrastructure layer; every business built on top of them became the actual value. OpenAI, Anthropic, and Google are the cloud providers of this AI generation. The founders trying to become a foundational model company without proprietary data or unique distribution are competing for a race that has already been run. The better bet: find the vertical, find the data moat, and build the business that runs on the infrastructure - not the infrastructure itself.
What does 'distribution on the cap table' mean in practice?
It means giving equity to people who are already inside the networks you need to reach - not hiring salespeople, but recruiting co-builders who can open doors because they are known and trusted inside the target market. For Josh's robotics company, this meant giving early equity to people who already owned the kind of factories they were building for. Those partners became design partners, early customers, and organic referral sources - all because they had skin in the game and relationships the team never could have cold-called its way into.
How does Josh approach the post-exit identity crisis?
He took a month and a half off and calls it a long weekend. The honest answer is that the identity is so tied to the company that when it is gone, you have no ready answer to 'who am I?' His practical solutions: join a peer group of post-exited founders (PEF - 1,500 exited tech CEOs), find activities that are not outcome-based (Dungeons and Dragons every Saturday), and accept that the search for meaning is an experiment rather than a problem to be solved. He is still in it. He does not have a clean answer. That honesty is the most useful thing he says about it.