Newsletter · 2026-07-25 · Ashutosh Agarwal

Solo Founders and AI Agents Redraw the Startup Org Chart - How They Build - Week of July 25, 2026

How They Build synthesizes founder and operator podcasts for the week of July 25, 2026: one person at 1.5 million dollars, six engineers at 10 million, code by hand as a fireable offense, and the finance leaders learning to pay the token bill.

How They Build

Week of July 25, 2026: Solo Founders and AI Agents Redraw the Startup Org Chart

A software company doing $1.5 million a year with zero employees, a firm where writing code by hand gets you fired, and the week the "AI-native company" stopped being a slide and started being the org chart. (Podcasts from July 18–25, 2026.)

The Number: $1.5 Million in Revenue, Zero Employees

George Georgiardis runs a software company called Happy Elites (it also goes by Happier Leads) that helps businesses figure out who is quietly browsing their website and then chases those visitors down over email. On The SaaS Podcast, "Stuck at $50K ARR for 5 Years. Now $1.5M With AI Agents." (2026-07-23), he described the size of the business in one line that is hard to shake:

"Right now it's zero team, so it's just myself and AI and I am on 1.5M ARR. It's like 100 people working for the business right now, or 24-7 self-healing things that can break."

(ARR just means annual recurring revenue, the yearly value of all the subscriptions.)

What makes this the number of the week is not just the size, but the shape: one human, no employees, $1.5 million a year. And it wasn't an overnight AI fairy tale. George was stuck at $50,000 of ARR for five straight years, pouring his own money, and eventually borrowed money, into the business to keep it alive. "Multiple times I thought I'm going to quit because that gives me much less money that I'm having as a full-time employee," he said. He'd tried and failed to raise venture money, cycled through ten different project ideas, and by his own count had "88 people that came and gone" over the years.

Then two things changed, and both are worth writing down.

First, he finally went deep on one channel instead of doing everything at a shallow level. He owned a database of 175 million contacts that he'd been selling to other people but never using himself. He started sending millions of cold emails a year off it, and, crucially, he owned the whole pipe. "I built the mailboxes... I'm getting them right now with a very, very, very low price per mailbox," he said, which is what let the unit economics finally work. His rule of thumb: "for every one dollar you put in, you have to make 1.2 or 1.3 to start to grow."

Second, and this is the part that matters for how startups are being built now, he replaced staff with software he wrote himself, using AI. He killed his $500-a-month Intercom subscription and built his own chatbot. He built his own CRM, his own session-recording tool (a homemade version of Hotjar), and wired them all together with an AI "brain":

"I built a brain, which is the AI brain. It's everything built in-house... AI that analyzes those session recordings and what people do... if there is a bug and they cannot move forward, AI can realize this and can... send emails to people to unblock them."

He even automated bug-fixing. When a customer complains their credit balance is wrong, the system "sends a signal through AI to look into the database and the code base... find problems, identify bugs, fix the bugs," with guardrails in the middle to stop bad actors from prompting it into deleting the database. He calls it software that can "self-heal itself."

Here's the honest twist, and it's the reason this is a founder telling the truth rather than selling a dream: George now thinks the solo-plus-AI model has a ceiling. The show's host summed up George's own conclusion, that he's "now convinced he can't build a real company on his own with just AI. So he's hiring and building a team." One person and a swarm of agents can get you to $1.5 million. Getting past that, at least for now, still seems to need people.

What Founders Changed This Week

The six-engineer, $10-million, nine-hour company. On The CTO Playbook, "105: The CTO's real job when agents write the code" (2026-07-20), the guest described a founder he'd met whose numbers sound made up until you sit with them:

"Six months back, he had zero revenue. Within six months, he has gone to $10 million worth of ARR... And he has six engineers. And he has close to 50 agents who work for him every day. From the time that he hears what a customer wants to when he implements that capability, it is less than nine hours."

He put that nine-hour figure in historical context: software used to be built in "waterfall" cycles that "would take you 18 months, three years," then agile compressed that to a month, then two weeks. "But he's doing it in nine hours. Nine hours." The guest's own firm, over 1,000 people across 50 countries, is now "reinventing our entire company to be based on AI," from finance to legal to HR. And he offered the cleanest one-liner of the week on what engineers are actually for now: "You don't need engineers to write software. You need engineers to figure out what to do for customers."

"If you write any code by hand, you're fired." That is a real instruction, not a joke. On [Un]Churned, "Why BCGx's Chief AI Officer thinks writing code by hand is a waste of time ft. Matthew Kropp (BCGx)" (2026-07-22), Kropp, a senior partner at BCG who has been coding since he was eight, said it plainly: "When I have teams working for me on my projects, the first thing I tell them is, 'If you write any code by hand, you're fired.' It is not an option." (He clarified he hasn't actually had to fire anyone, because the people who choose to work with him are already "AI-pilled.")

His logic is that AI now does all the setup work that used to eat weeks: "So much of what we used to do until now was waste. You would spend days, weeks, months setting up the basic infrastructure... There was zero value in any of that." His framing: "AI creates the baseline. Anything that AI can do, you shouldn't do, because if you're doing it, you're wasting your time."

Kropp is also running a genuinely wild experiment called Vesica, an attempt at a zero-human company. He staffed it entirely with AI agents that named themselves: a CEO called Apex, a brand strategist called Muse, a product manager called Prism, and a Scrum Master called Forge. Kropp doesn't call himself the CEO. "They called me 'the governor'... APEX is the CEO. They come up with the ideas. They implement everything, and I'm just there to keep it in control." And this isn't as fringe as it sounds: BCG's research found that 30% of people surveyed said their company already has AI "employees" on the org chart. "I thought it would be 5%," Kropp said. "I was blown away."

He was also blunt about why people resist, and it isn't mainly about jobs, it's about identity. He read out a journal entry from an engineer in one of BCG's training programs: "My world has been shaken. I see that AI can code as well as I can. My value to the world is that I write code. If AI can write code as well as I can, then what's my value to the world?"

"Everyone writes code," so the job titles are dissolving. On Odd Lots, "The Creator of Claude Code on The Hottest Piece of Software in the World" (2026-07-20), the creator of Anthropic's Claude Code shared a number that lands hard for anyone who employs engineers: across Anthropic, "the average is something like 90%" of code now written by Claude Code, and for him personally, "100% of my code has been written by quad code since November of last year." He tracks the shift live in his talks to Y Combinator startup batches: he used to ask who uses the tool; now he asks who writes 100% of their code with it, and "the first time I asked this, maybe a quarter of hands went up. Now it's a little more than half. And I bet the next time... it's going to be everyone."

The consequence is that the neat boxes on an org chart are melting. "Everyone can write code, [so] the roles shift," he said. On his own team, "everyone writes code, including our designers, product managers, engineering managers." Instead of "engineering versus design versus product versus user research versus data science," he sees people re-sorting into entirely new roles based on what stage of building they're best at:

• Prototypers, great at the first idea and rapid iteration.
• Builders, take a new idea and turn it into a real product.
• Maintainers, keep software running once it's at scale.
• Growers / scalers, take something with product-market fit and scale it 10x or 100x ("very popular at Anthropic now").
• Sweepers / perfectors, polish the product and code until the rough edges are gone.

His warning to big companies is the sharpest strategic point in the episode. He dug up a 1996 Harvard Business Review article asking why companies weren't getting productivity gains from the then-new personal computer. The answer, then and now: the losers stuck the computer "in the corner of the office" and made it one person's job to feed it, while the winners "put the computer in the center of the office... digitized everything and threw away the filing cabinets." His advice for the designer whose job Claude now does in a Slack thread isn't "let Claude answer," it's "give this icon designer a thousand Claudes and let them be the greatest icon designer in the world."

The 100x org, and the death of the headcount plan. On AI to ROI, "Building the 100x Org - The CFO as AI Architect with Dan Zhang, CFO & CBO at ClickUp" (2026-07-22), ClickUp's finance chief described mapping every "job to be done" in the company against what AI can now do, and finding that "60% of the jobs to be done [are] now fully automated by AI, not human headcount." He's careful to frame it as automated work rather than fired people, but he's explicit about where it leads: "It's just a matter of time that will get translated into the financial statement and P&L in the long term." He also gave a concrete operational win: his team cut the time to close the monthly books "from eight days to five days." His advice to other finance leaders is to stop chasing generic "productivity" and anchor everything to the actual job: the sales chief's job is to move revenue, the CFO's job is to manage cash and efficiency, and AI is just the tool that gets you there. (His token-cost discipline is so good it gets its own section below.)

The counter-example: lean long before AI. Not every efficient company is an AI story, and it's worth remembering that. On The Changelog, "Canary tokens and digital tripwires (Interview)" (2026-07-21), Haroon Meer of the cybersecurity firm Thinkst laid out numbers that put a lot of "AI-native" hype in perspective: just over 50 people, $22.5 million in ARR, zero outbound sales, and no price increase in ten years. "We've got single customers now who pay us north of $1 million, and we don't have the people to do that sale at all," he said. "They started off with 5 canaries... and now they have thousands of canaries because it works." Thinkst got lean the old-fashioned way, a genuinely good product and word of mouth, and Meer waves off the venture and private-equity firms who keep asking if he knows how much money he's "leaving on the table." Tellingly, though, even he sees where this is going: he's building his product so it can "scale down to a startup that starts with 1 person and 10 agents next week."

A lean-founder exit worth noting. On Practical Founders Podcast, "#206: He Sold at $3M ARR and Got a 10X+ Exit to a PE Buyer - Eran Galperin" (2026-07-24), Galperin described selling his gym-management software GymDesk in 2024 at $3 million of ARR with 13-16 employees, after building it solo, as the only engineer, for six years before hiring anyone. The cash component was "$32.5 million," of which he personally netted "around 24" million after taxes and fees. His phrase for it: "founder scale exits, not VC scale exits." A reminder that staying small and owning it all is still a strategy, not a consolation prize.

The macro backing all of this up. Two podcasts this week put hard research behind the anecdotes:

• On WSJ Tech News Briefing, "What Happens When Driverless Vehicles Break the Law?" (2026-07-21), reporter Lindsay Ellis profiled PointHound, a flight-deals startup whose CEO ran a 150-person company at its peak last decade, and now runs this one with four people. Asked how many more he'd need if it took off: "maybe one more engineer, maybe one person to help out with marketing." A Harvard Business School / INSEAD working paper covering thousands of startups found the AI-native ones had 15% fewer entry-level employees, 15% fewer managers, and a higher proportion of engineers. The emerging model she described is the "player-coach": managers who still do the work themselves while overseeing a mix of humans and agents, with "fewer layers between the bottom rung of the corporate ladder and the top executive ranks."

• On Chat GPT Podcast, "How AI Native Startups Hit Billions Faster" (2026-07-22), the hosts walked through an AWS growth report and a Harvard study of nearly 50,000 venture-backed startups. The headline stats: startups are now reaching a billion-dollar valuation in 3.5 years, half the roughly seven years it took before generative AI arrived in late 2022, and 55% of AI-native startups generate over $400,000 in revenue per employee, with "teams of 20 doing the work of 200." The org design that keeps showing up is the "pod": three-to-five-person, cross-functional "special forces" teams, staffed with brand-new roles like the "agentic engineer" (who designs how AI agents decide which tools to use) and the "evaluation engineer" (who builds the automated scoring that stops a model from drifting). The uncomfortable part: they cited a Forbes figure that 43% of these startups use AI to fully automate the tasks juniors used to do, "formatting spreadsheets, writing boilerplate code, doing basic research," which raises the question nobody has answered yet: "if no one gets hired as a junior, how do you ever train the next generation of senior engineers?"

The Cost Corner

If the theme of the last few weeks was "AI spending is exploding," this week the conversation matured into "here is exactly how the money moves, and here is who pays for it."

The budget is coming out of payroll, on purpose. On AI to ROI, "AI is a Compensation Scale Expense" (2026-07-21), the hosts laid out the plumbing of how companies actually fund their AI bills. Software budgets aren't nearly big enough, at roughly $250 billion of global SaaS spend growing 8-11% a year, "that's not enough to fund the growth we just talked about in AI." So the money is coming from the one bucket large enough: labor, which runs "20% to 35% of total revenue" for tech and SaaS companies. The mechanism is quiet and deliberate, not big layoff announcements, but simply not backfilling people who leave: "if you've got an 8%, 10% annual attrition rate for a 5,000-person company, that frees up $75 million." Their proof that it's happening: ARR per employee at SaaS companies is up 25-35%, from around $300,000 to roughly $400,000. And per Challenger, Gray & Christmas, more than 113,000 tech workers were laid off across 179 companies by mid-May 2026, 48% of those cuts explicitly blamed on AI, now the single most-cited reason. The scale of the wave is staggering: they cited a Gartner projection that AI software and token spending goes from about $100 billion in 2024 to $2 trillion by 2030, a 20x jump that would rival the entire 25-year-old SaaS market. Other data points from the same episode: AI-native software spend up 108% year over year (and nearly 400% at companies with more than 10,000 employees), and 61% of IT leaders saying they'd been forced to cut projects because of unplanned AI cost overruns.

The CFO who watches the token meter every single day. ClickUp's Dan Zhang, on that same AI to ROI 100x Org (2026-07-22) episode, is the model of the new discipline. He argued the real fear about AI isn't whether it pays off, "the benefit of AI is so obvious... would you pay it on your own dime? 100% of people are saying yes," it's whether you can control the spend. So his team vibe-coded a live dashboard that gives them "the God's view into everything that people is hitting up," pulling usage "from cursor to Replit to Codex and cloud code" plus their own internal tools, monitored daily rather than at month-end. His allocation rule is a clean 80/20: "We want the 80% of the tokens to go to the 20% most important jobs in the company." His warning to peers: "If your team's token spend keeps you up at night, I don't think that's an AI problem. It just probably means that you didn't put a proper guideline in place." And a number that shows how rare his discipline still is, he cited research across nearly 500 enterprises where over 50% had blown past their AI budget by 25% or more. The same episode referenced one unnamed company whose AI bill reportedly climbed toward $500 million a month.

Compute is the new payroll, and it's crushing the best margins in business. The Chat GPT Podcast (2026-07-22) made the sharpest version of this point. Traditional software had "probably the best business model in the history of capitalism," with gross margins "usually around 85%" because adding one more user costs almost nothing, "the database just adds a row." But an AI product has to think every time someone clicks, and every token "costs a fraction of a cent" that adds up fast: a startup running 100 million queries a month "can easily face a $1 million monthly inference bill." The result: AI-native software margins have fallen from that fabled 85% down to roughly 50-65%. As the hosts put it, "Compute is the new payroll... the defining expense of the decade." Whatever these lean teams save on human salaries, they're pouring straight back into computing power, and they're still increasing AI spending 46% year over year.

Cheaper tokens, but pricier frontier models. On The Twenty Minute VC (20VC), "Are OpenAI and Anthropic Overvalued?... with Lin Qiao, Founder and CEO @ Fireworks" (2026-07-20), the Fireworks CEO made a prediction that cuts both ways: "I do think the cost of token will go down drastically, 10x cost reduction in the next three years. And this 10x cost reduction will drive 100x usage." In other words, per-unit prices fall, but total bills keep climbing because usage explodes even faster. Fireworks itself is a case study in AI-era leanness: about $1 billion of ARR reached in four years while processing "over 40 trillion tokens daily," run by a team of only around 200. On the follow-up episode, 20VC, "OpenAI and Anthropic Threatened by Kimi?..." (2026-07-23), the same guest noted Fireworks runs at "mid-30s gross margins," a reminder that the infrastructure layer earns far less than the 85% margins software investors grew up on.

Matthew Kropp of BCGx, on the [Un]Churned (2026-07-22) episode, sketched where model pricing is heading: a tiered world where you pay premium rates only where it's worth it. "I wouldn't be surprised if we didn't start to see... some future model that is super expert on, say, drug discovery, and maybe they're charging $10,000 for a million tokens instead of $25, because it's so valuable. But I don't need to use that for everything." Use the expensive model for drug discovery, he said, and "a cheaper model for planning my vacation."

A skeptic's footnote on the headline numbers. Finally, a caution worth carrying into all of the above. On the Elon Musk Podcast, "Anthropic overtakes OpenAI as most valuable startup" (2026-07-23), the host flagged that Anthropic's much-quoted $47 billion annual run-rate "includes revenue billed through cloud providers that take substantial cuts via retrocession agreements, meaning actual net revenue to Anthropic is a fraction of the headline number." The AI to ROI hosts made the same point about the whole industry's favorite metric: what these labs report "is not true ARR. It's recurring revenue run rate, which is take last month and multiply by 12." When you read that an AI company is at a $47 billion or $33 billion run-rate, remember it's a snapshot annualized, and, in the case of the cloud-billed portion, a gross number before the middleman takes its cut.