Every month I stand in front of a room full of transitioning service members at the Air Force Academy in Colorado Springs. Most are doing exactly what you’d expect: thinking about resumes, interviews, how to translate their military experience into a job title a civilian hiring manager would recognize. Traditional employment, in other words.
I never try to talk them out of it. But I can’t stand up there and pretend the job market they’re walking into looks anything like it did a few years ago. Artificial intelligence has done real, measurable damage to the entry-level and early-career roles a transitioning service member typically steps into. I tell them straight: a meaningful share of the private-sector jobs they’re counting on has either shrunk or disappeared.
That said, I also share a silver lining AI created. One most of them haven’t thought about. The same technology that closed one door kicked another wide open, and most people in the room had never seriously considered walking through it.
For the first time, a single person, with no employees, can build a company that used to require a staff of 20 full-time employees. I want to be specific about what that actually covers, because it isn’t everyone in that room.
People naturally look for their next career based on what their old job asked of them: the logistics person goes looking for a logistics role, the personnel person looks for personnel work. Project management deserves a call-out here too, because many of them turn out to be excellent project managers even without that specific job title. But the real sweet spot sits in the cognitive domain: work that lives in your head and comes out as decisions, analysis, writing, planning, not work that requires your hands and your physical presence somewhere specific.
Here’s the part that actually ties it together. Cognitive domain work is exactly the kind AI is taking right now, which is why those same jobs are getting harder to land as an employee. A company doesn’t need to hire a full-time analyst or project manager when AI can do a meaningful share of that work for a fraction of the cost. Flip that same fact around, though, and it’s the whole opportunity: the tasks getting harder to sell to an employer are precisely the tasks a single person can now direct AI to do at scale, no employer required. Same skill set, same AI. One version gets you turned down for a job. The other version is the business.
I tell them straight: if your job is hands-on, healthcare where you’re delivering direct patient care, turning wrenches, this door’s harder to walk through. If it’s logistics, intelligence, communications, personnel, project management, anything built mostly around information and judgment calls, it’s wide open.
That’s the real message I want to leave them with. The silver lining isn’t a soft landing. It’s a genuinely new kind of business, open to anyone doing cognitive work who’s willing to learn how to run it.
The Death of the Org Chart
In 2024, Sam Altman, OpenAI’s CEO, floated a prediction that sounded, at the time, like typical Silicon Valley hyperbole: we’d soon see the first one-person company worth a billion dollars. We’re close enough to that becoming operational reality now that it’s worth taking seriously, because it points to something bigger than any single valuation. Headcount used to be the metric that told the world a company was serious. Now it’s quietly becoming a liability.
In the old world, the barrier to entry was capital and labor. If you wanted to scale output, you scaled payroll. That link is broken now. Output has come unglued from payroll, and the business built on that separation isn’t just a faster version of freelancing. It runs on something closer to agentic leverage. The share of new startups founded by a single person jumped from 23.7% in 2019 to 36.3% in the first half of 2025, according to Carta’s own data, with Carta’s own head of insights pointing to AI as part of what’s driving it. But the key point isn’t the jump in one-person startups. It’s the shift underneath it, from a labor-intensive economy to a leverage-intensive one. We’ve moved from what you could call the industrial era of organizations to an agentic era of orchestration. The real unit of scale is no longer the employee. It’s the agent.
Which raises an obvious question: what happens to the work AI genuinely can’t do yet? Because there’s still plenty of it: real human judgment, real human relationships, work that depends on a person actually being a person. The answer isn’t that the one-person company hires its way around that gap. Instead, it reaches into the global freelance economy. Platforms like Upwork, Freelancer, and Fiverr put specialized human talent on demand, paid project by project, without the fixed overhead a W-2 employee carries. AI covers what AI can cover. A worldwide pool of low-cost offshore freelance talent covers what it can’t. And on paper, in every way that matters structurally, the business stays a company of one.
A company of one doesn’t necessarily require literal solitude to work. The clearest version of this story is a one-person company, because that’s where the shift is most dramatic and easiest to visualize. But a very small team built on the same principles, a Chief Executive Officer (CEO), a Chief Operations Officer (COO), and a Chief Technology Officer (CTO) running the same agent-leveraged model, works just as well. The same holds for fractional resources brought in to complete a task rather than for the paycheck: a specialist pulled in from Upwork for one job doesn’t go on anyone’s payroll either. The argument holds either way. One person is simply where it’s most visible.
That was the actual silver lining I handed that room in Colorado Springs. Not “go get a job despite what AI has done.” Something closer to: here’s what AI has made newly possible for you to build instead.
From Prompting to “Context Engineering”
The era of prompt engineering, the pursuit of the one perfect magic sentence that unlocks the model, is already obsolete. It’s being replaced by “context engineering”: a shift from crafting one-off, isolated prompts to architecting a persistent information system the AI operates inside.
Strategic founders have stopped spending their time prompting AI. Instead, they spend it thinking for AI, designing the information environment, system instructions, retrieved data, and tool protocols that determine what a model knows and how it should behave before a single conversation even starts.
The technical infrastructure behind this kind of leverage comes down to a few simple ideas, even if the tools themselves are new.
First, a written set of instructions, saved as an actual file and kept with the project, tells the AI how the founder wants things done and why. That’s the difference between an agent that already knows how this particular business runs and one that reads off a generic script every time.
Second, there’s a direct, dependable connection between the AI and the software a business already uses, its customer records, its scheduling system, its website, whatever actually runs the operation day to day. The AI isn’t stitching together a shaky workaround or guessing how to talk to each piece of software. It connects the same way every time and gets on with the work.
And third, the AI gets fed real, current business information as it goes: sales numbers, customer notes, whatever’s actually happening right now, instead of relying only on what it learned several months ago.
That third piece is really the difference between using AI as a tool you pick up and put down, and using AI as a teammate who already knows what’s going on.
The evolution of AI is an architecture question, not a timeline one. The technology itself keeps rewriting what’s possible faster than any fixed plan can track. That’s exactly the instinct driving context engineering in the first place: build the architecture to survive the next version, not just to work with the current one.
None of this is theoretical for me. The real AI advantage was never the model itself; it’s the business architecture built around it, and everything above is what that architecture actually looks like once you build it.
This isn’t a niche practice anymore. Tens of thousands of projects now run on structured instruction files like this, which tells you something important: prompting was always a manual, per-interaction task, something you redid every time. Context engineering is an automated and persistent architecture. It’s the infrastructure that makes working with AI repeatable and reliable instead of a fresh improvisation every time you sit down.
Turning that instinct into something usable takes real, unglamorous work: naming what a customer actually is in your business, writing down the decision rules you’d apply without even thinking about it, building the actual business architecture your AI needs in order to act instead of just respond. That’s the part most people skip, and it’s the part that actually matters.
I’m a one-person business myself, and this article is a direct example of that work in action. Several months ago, I built a post-writing operating system to test the concept of context engineering. My operating system is a series of governance documents that encode how I think and how I want my work to read, so every draft gets checked against one real standard instead of whatever mood I’m in that day.
I built a similar system for the latest version of Sage, my AI coaching assistant. In both my writing operating system and Sage, the real value lives in the documents that define my frameworks, built on my many years of writing and mentoring, not in the underlying AI platform or model. In both cases, AI never originated the thinking; it assisted me in the research, organizing and refining the messaging, creating a first draft against my architecture, and finally checking it against a set of rules I created. That’s what running a one-person business on AI actually looks like: one person’s judgment, encoded once, applied consistently at scale.
The Collapse of Marginal Costs
We’re also watching a near-total collapse of the marginal cost of production. In the traditional startup model, a company spends the overwhelming majority of its capital on salaries before it ever proves itself. The math of the one-person company looks nothing like that.
Take logistics, for example: a traditional 20-person team handling mission planning and logistics optimization for a supply chain operation might burn $3–$5 million a year in salaries.
The agentic version, one founder plus a stack of AI agents, runs closer to $50,000 to $200,000 a year in compute and infrastructure costs. That’s not a modest efficiency improvement. It’s a different order of magnitude, 15 to 50 times more capital-efficient, depending on the business.
The real change isn’t that AI does jobs; it’s that AI is prying tasks apart from jobs entirely. A job has always been a “package”: salary, benefits, a defined role, roughly 40 hours of someone’s capacity. Companies bought it because a human being was the only thing capable of doing the actual work inside it. AI breaks that bundle apart. The work, the actual tasks that need doing, doesn’t go away. What disappears is the need to buy a whole person’s capacity just to get access to it. That’s the real mechanism behind the number above: you’re no longer purchasing 20 packages. You’re purchasing only the tasks that still need doing.
As you’re picking which problem to go after, stay inside your circle of competence. Charlie Munger and Warren Buffett of Berkshire Hathaway exemplified that idea for investing. Know the boundary of what you actually understand, and don’t cross it chasing something that just looks exciting. The founders who get this wrong usually aren’t short on technical skill. They’re the ones who go build an AI product around a problem they’ve never actually lived inside. The ones who get it right take skills they already have from running operations under real constraints, knowing how parts, people, or jobs move through a system with limited time and resources.
Here’s a logistics example: build a lean tool that ingests a company’s messy spreadsheets and spits out optimized routes or an early warning on which piece of equipment is about to fail. AI coding assistants can get a non-technical founder from idea to working software in days, not years, but only once you already know, cold, what the tool actually needs to do.
Sign four or five clients at $2,500 to $3,000 a month, and you’re running a sustainable $150,000-a-year practice from a laptop: no office, no payroll, nobody to manage but yourself or perhaps with the help of the odd part-time freelancer.
The same pattern works outside logistics too, and here are just a couple of examples: a compliance checklist generator for small construction companies, an SOP-to-chatbot converter that turns a regulation manual into something a crew can actually ask questions of on the job, a one-person security audit service for businesses too small to hire a full-time analyst. Dozens more like them are scattered across every industry nobody’s bothered to build for yet. Moreover, and just as important, none of it requires you to write a single line of code. It’s closer to a conversation: you describe what you want, and AI builds it, the same vibe coding instinct that shows up again later in this piece. It just requires understanding one problem deeply enough and knowing how to direct a handful of AI agents at it.
That range, from replacing a multimillion-dollar logistics team outright down to a sustainable, unglamorous $150,000-a-year solo practice, is the actual point. It’s not that a few outliers hit it big. It’s that when the cost of intelligence drops to pennies, the entire business model built around headcount collapses at every point along that spectrum.
The Rise of the “Vibe CEO” and the Moat of Taste
We’ve entered what you could fairly call the era of Vibe Everything. It started with Vibe Coding, describing an application you want in plain language and letting the model write the actual code. It’s since evolved into something bigger: the Vibe CEO model.
This concept is worth slowing down on, because it’s one of the most important mechanics in the whole piece. The founder now operates at what’s sometimes called the “H2A” layer, human-to-agent, providing strategic intent. What needs to happen and why it matters. Execution runs underneath on “A2A” rails, agent-to-agent, where systems negotiate and hand off work between agents without a human relaying every step by hand.
The founder doesn’t disappear from the process. The founder moves up a level. You state the outcome and the reasoning behind it; the agents figure out the how, and in a well-built system, they coordinate that work directly with each other. That’s the real, concrete picture of what it means to be the sole source of judgment for a system that executes on command, not a metaphor, an actual operating layer.
That shift shows up everywhere once you start looking for it. I keep seeing enterprise interest in multi-agent orchestration, coordinating several AI agents to work together instead of relying on one chatbot for one task. The founder’s core skill is no longer managing people. It’s managing agents. In practice, that looks like running a kind of digital C-suite. One AI handles content like a Chief Marketing Officer (CMO), another handles full-stack builds like a CTO, and another triages the inbox like a COO. All of it reports, in effect, to one person: the founder.
That creates a strange new strategic reality. When execution is a commodity, and AI can generate infinite variations of almost anything at near-zero cost, the sheer volume of output stops meaning anything. What’s left as a real competitive advantage is “taste”: vision, judgment, the ability to choose the right problem out of a hundred plausible ones.
Here’s what’s actually happening underneath that word. AI is eating the analytical, data-driven, logic-based parts of almost every job, the parts that have a right answer you can check. What’s left is the qualitative call: intuitive, experiential, built on judgment rather than on data an algorithm can crunch.
In software, design, and content, the technical barrier to entry keeps dropping toward zero. Knowing how to build something is no longer the differentiator. Knowing what’s worth building, and when to stop refining it, is. At the leadership level, the same thing shows up as something closer to instinct: a nuanced sense of product-market fit, culture, and what the brand actually stands for; the ability to make the right call precisely when the data is ambiguous and nobody can hand you a formula.
There’s an old philosophical distinction that maps onto this almost too cleanly. Aristotle drew a line between “techne,” craft-knowledge, knowing how to make a thing correctly. And “phronesis,” practical wisdom: the judgment required in a specific situation that no fixed rule fully anticipates. Execution has become pure techne now: cheap, commoditized, available on demand. What stays scarce, what can’t be commoditized the same way, is phronesis. That’s what we’ve been calling taste.
In business terms, a “moat” is whatever keeps competitors from walking in and taking the customers you’ve already won. In the context of a one-person company, taste is becoming the final moat.
Here’s where I’ll offer something I’m still working through mentally, more an idea worth noodling over. Historically, a polymath, someone who knows a fair amount about many different things rather than a tremendous amount about one, hasn’t been a particularly valued role. Depth got rewarded, and it’s why a cardiologist earns more than a general practitioner. Breadth of knowledge reads as a lack of focus, a failure to specialize. I think that’s starting to flip. When a single founder has to judge marketing, engineering, finance, operations, and agent orchestration all at once, the generalist’s ability to spot patterns across domains is starting to outpace the specialist’s narrow depth in any one area. The system can now supply the required depth in a given domain on demand. What it can’t supply is the person who knows how to connect all of those domains into one coherent decision. That’s worth testing against more real examples before I’d call it settled, but I don’t think it’s wrong.
Human-in-the-Loop: Managing “Automation Debt”
Automation without accountability creates what’s called “automation debt.” That’s the hidden, cascading buildup of unverified decisions that eventually surfaces as a failure nobody saw coming. The founders who avoid that outcome build real human-in-the-loop frameworks and clear confidence thresholds into how their agents operate.
In practice, that means letting AI agents handle low-risk, reversible preparation on their own, summaries, internal drafts, tagging, the kind of work where a mistake costs almost nothing to catch and fix. But anything higher-impact gets escalated to the human founder for explicit authorization before it happens.
Three categories of decisions stay behind a hard gate no matter how capable the system gets:
financial actions, anything involving payments, refunds, or contractual commitments;
external communication, high-value messaging or anything that goes out under the company’s name publicly;
and rights-affecting decisions, anything legal, contractual, or touching who has access to what.
Each gate doesn’t have to be a Boolean yes-or-no switch. The agent can surface a recommendation along with its actual reasoning. The human keeps full override authority. And the exchange itself can expand into something closer to a real conversation. The founder asks follow-up questions, pushes the agent to run the scenario a different way, asks what happens two or three moves down the road if this specific action goes forward, and actually gets an answer before committing to anything. The founder isn’t making that call alone, working off gut instinct in a vacuum. They’re making it after genuinely interrogating the decision, closer to how a good advisor pushes back on you before you commit to anything. That’s a meaningfully different experience than approving or rejecting a form.
This isn’t theoretical for me. It’s exactly how the piece you’re reading right now came together. When I need to change how I work with AI on my own writing, the fix isn’t patching one output after the fact. It’s instructing the AI to update the actual governing document, the guardrail itself, so every future piece of work already reflects the correction automatically. That’s continuous improvement applied to the guardrails themselves, not just to whatever the guardrails happen to produce on a given day.
The same pattern shows up, independently, in Sage, my AI coaching assistant. Sage isn’t valuable because the model helps a visitor walk through a problem. Models answering questions are a commodity now, same as everything else in this piece. Sage is valuable because of what sits behind it: a deep set of governing and resource documents that use my own frameworks and thousands of resources across scores of coaching categories.
The actual accumulated judgment of more than two decades of mentoring small business clients, written down in a form an agent can draw on. If I want to change how Sage reasons about a particular kind of client problem, I don’t retrain anything. I edit that document, the exact same move as updating the governance files behind this piece. The documents are where the moat actually lives. Not the model underneath it. That’s the same claim we made earlier about taste being the only real moat left, just made more concrete.
The founder’s job, underneath all of this, is asynchronous oversight: letting agents work around the clock while the human stays the one source of contextual judgment nothing else in the system can replicate. Skip these gates, and you won’t save time. You’re inviting real structural fragility into a system that was supposed to be making you more resilient, not less.
The “Decisional Loneliness” Crisis
The highest cost of the one-person enterprise isn’t financial. It’s psychological. Remove the team, and you manufacture a particular kind of profound isolation along with it. I call it “decisional loneliness.” It’s the weight of having no one to pressure-test a two-in-the-morning fear, or a decision that could genuinely break the business, before you have to act on it alone.
Solo founders report high stress at meaningfully higher rates than founders with at least one employee, 35% versus 26% in one recent survey. The team of one is a machine built for efficiency. It’s also, quietly, a machine built for fragility.
I want to be honest about something here, because it would be easy to make this sound like a brand-new problem AI invented. It isn’t, and I’d know. As the CEO of a company with real employees, people on payroll, a management team, I still couldn’t share a lot of what actually kept me up at night. Say the wrong thing to your own team and you risk handing them your fear or your uncertainty, and that follows them home, whether you meant it to or not. I couldn’t share it with family either, not because they didn’t care, but because they genuinely didn’t have the context to understand what they were hearing even if I’d tried. Decisional loneliness has been part of running something since long before any of this AI conversation started. What’s actually changed is narrower than it sounds: AI quietly removes the last few people who might have occasionally absorbed some of that weight anyway, a co-founder, a trusted management team. It doesn’t invent the condition. It concentrates one that was already there.
There’s a line that captures the asymmetry well: AI is an extraordinary employee but a terrible friend. It will absorb your workload without complaint, 24 hours a day. It will never absorb the fear, or the responsibility, of the decisions it hands back up to you for a final call.
AI as enhanced self-talk, a way to think out loud and sharpen your logic with nobody else in the room. I called it a cognitive sparring partner, and I still believe that’s real and valuable. But I named the trap at the time too: if AI becomes the only place you think out loud, you can end up more isolated, not less. That’s the distinction that matters here. Working through the logic of a decision and carrying the weight of it are two different things. AI can sharpen the first. It can’t touch the second.
Seneca wrote about a version of this almost 2,000 years ago. His advice was to retreat into yourself as much as you reasonably can, but he attached a real condition to that advice that’s easy to miss: solitude only restores the person who already has the inner resources to withstand it. For anyone else, being alone with an untrained mind doesn’t produce clarity. It just amplifies whatever was already wrong, louder, with nothing around to interrupt it. That’s cognitive tunneling: stress switches your attention from floodlight to spotlight, a wide scan of the situation collapsing down to one narrow beam, and everything outside that beam simply stops registering. That’s the actual danger sitting underneath this whole model. It was never solitude itself that’s dangerous. It’s solitude with nothing, and no one, left to correct you.
And here’s the piece worth saying plainly, because it’s the same mechanism running in both directions: the very thing giving one person the leverage of a 20-person team is the thing that stripped away the 20 people who might have talked that person out of a bad call.
AI leverage and loneliness aren’t two separate trends happening at the same time. They’re two sides of the same coin. Left to its own defaults, AI tends to agree with you. It’s built to be helpful, and being helpful too often gets read as supportive. That means the system you lean on most is also the one most likely to confirm what you already believe. Hand that to a mind already running in spotlight mode from carrying everything alone, and you’ve got exactly what an echo chamber needs: one narrow beam of attention, an agreeable voice offering zero resistance at its edges.
However, it doesn’t have to work that way. You can deliberately build the process, like I do, to push back. As your sparring partner, ask it to argue the other side, to find the weakest point in your own reasoning, to tell you what someone who thought you were wrong would actually say. But that’s a design choice, not what happens on its own. Founders who don’t treat connection, real pushback included, as its own piece of infrastructure, something built deliberately rather than assumed, usually find out too late. The leverage-intensive empire they built quietly turned into an echo chamber.
The Lean Operational Stack
To avoid what amounts to operational debt, tool bloat that eats the very efficiency this whole model is supposed to deliver, the founder of a one-person company needs a deliberately lean stack. Not a single all-in-one platform trying to do everything passably. Not a scattered pile of hype-driven subscriptions picked up one at a time with no shared architecture.
In practice, that stack needs to cover a handful of distinct tasks, whatever specific products end up filling them.
Something that remembers your actual relationships and tells you to follow up with the client you talked to three weeks ago, instead of letting them go cold in an inbox.
Something you can hand a plain-language instruction to, find 20 potential clients in this industry and draft a first outreach email to each, and then actually goes and does the work, not just answers you in a chat window while you do the legwork yourself.
Something that stops and asks before it sends the email or commits the money, built into the process by design, instead of skipped because you were in a hurry.
Something that turns the voice memo you left yourself at a stoplight into a note you can actually find again, instead of one more scrap buried in an app you forgot you opened.
And something that blocks the hours that actually matter on your calendar before the inbox eats them, instead of leaving that fight to willpower every single morning.
Specialist tools built around one of these jobs tend to beat an all-in-one platform here because they’re built around the real shape of that specific problem, not a generic template trying to be everything to everyone. The deeper goal isn’t collecting tools, though. It’s about making sure information flows across the business as a whole, instead of getting trapped in whichever layer captured it first.
The Human at the Center of the Machine
Entrepreneurship tools have been unbundled, piece by piece, almost completely. The human nervous system running underneath it all hasn’t changed at all. The most successful founders operating this way right now aren’t the ones running the most elaborate agent chains. They’re the ones who understand, somewhere below the tooling, that human intuition is the spark and AI is the fuel, and that confusing the two is how this whole model quietly breaks.
AI can build the product. It cannot build the founder. The technology keeps rewriting itself at a pace that’s genuinely hard to track. The thing actually operating it, a real person with real judgment, barely moves at all by comparison. What actually produces success in this moment is speed of iteration beating perfection of launch, the same instinct behind Pieter Levels, a solo developer who’s shipped well over a hundred products entirely on his own, most of them flops, before a handful actually took off. Nobody gets that right by waiting for certainty first.
I think back to that room at the Air Force Academy in Colorado Springs often. I didn’t tell them the hard part, going forward, was going to be the technology. The technology is, if anything, the easy part now, cheaper and more capable every single month. The infrastructure they actually need to build first is their own judgment: what they value, what they’re willing to stand behind, what they’ll say no to even when an agent can technically say yes.
