The original thesis said agencies were hollow engines. It was right. It was also too small.
Preamble
In October 2025 I published a 97-page thesis called The Infrastructure Operator Doctrine. The argument was structural. The agency model, the thing that employs roughly 500,000 firms in the United States alone, is a hollow engine. It produces value only while it's running. It stores nothing. Every dollar of revenue requires a corresponding dollar of labor. The founder gets trapped inside the machine because the machine is the founder. And the more successful the agency appears from the outside, the more fragile it becomes on the inside, because success means more clients, more clients means more labor, more labor means more management, and none of that mass makes the engine any less hollow.
I ran my own agency before I wrote that thesis. Scaled it to mid five figures a month, lived the model from the inside, felt the break points, and left for software. For the past three years I've been building custom software and systems for agency owners. The ones who wanted out of the model. The ones who needed a platform for their clients, or wanted to turn their domain expertise into a product they could actually own instead of a service they had to personally deliver. Five years in the SaaS and AI space total.
Most people who follow my work know the SaaS and AI side. That's not the full picture. I built a SaaS and exited it for mid five figures when I was 21. Then started a B2B AI sales platform, closed a million dollars in LOI pilots, and watched it die on the runway because we bootstrapped, couldn't raise without collected revenue, and a third-party dev team missed deadlines for six months straight until the window closed. Five years in SaaS and AI. Not all wins. But all real.
Roughly thirty clients over the past three years, all transformed from labor-dependent to system-dependent. About twenty of those were agencies who hired me to build out a software product, an app, or an internal system for their business. The ones who saw the hollow engine before I gave it a name. The rest spanned verticals that had nothing to do with agencies. A wearable device platform for a nine-figure health-tech company. Manufacturing infrastructure for one of the largest peptide producers in the US. Channel systems for a publicly traded commerce company that did a million in its first twenty days. Health-tech, biotech, finance, commerce, martech. Same pattern every time. The business ran on humans doing repeatable cognitive work, and I built the system that replaced the repetition with infrastructure the founder could own.
The thesis wasn't academic for me. It came from doing this, over and over, across industries that had nothing in common except the same structural problem.
When I published it, I thought it was forward-looking. I thought the market would arrive at the conclusions I'd outlined over the next three to five years as AI tools matured and the economics shifted. I thought agency owners would slowly feel the pressure, that some would adapt and some wouldn't, and the transition would play out over a normal business cycle. Gradual. Manageable. Something you could plan for.
I was right about the direction. I was wrong about the timeline by a factor of ten.
Five months later I'm rewriting the thesis. Not because the core argument broke. Because it turned out to be too narrow. The hollow engine isn't an agency problem. It's a business structure problem. And what I described as a slow-moving structural shift has turned into a pressure test that's hitting every industry, every business model, and every assumption about how work gets done, all at once, all right now, in March 2026.
The original thesis argued that the agency model was structurally doomed and that the way out was to distill the founder's domain expertise into systems and infrastructure. Stop selling hours. Start building something you own. Something that generates recurring revenue, that the business holds as intellectual property, that gets valued like software rather than like a founder-dependent service. That argument still holds. Agencies that stayed in the labor model are still declining. The numbers I'll show you confirm it.
But what I missed, what I don't think anyone fully anticipated at this speed, is that the same structural weakness exists in any business carrying weight that isn't load-bearing. Any business where the value-producing work is performed by humans who could be replaced by a sufficiently capable system. Any business where the overhead exists because, until recently, there was no alternative.
That's not just agencies. That's most businesses.
This is the expanded version. It covers agencies, but it goes further. It covers the SaaS model, which stopped being a product category and started being a component layer. It covers the expertise economy, which is being systematized and distributed at near-zero marginal cost. It covers the emergence of skills as the atomic unit of business capability. It covers the nuances that the hype cycle glosses over, the cost math at scale, the liability gap that nobody has figured out, the security problems that haven't gone away. And it covers the uncomfortable reality that most of the economy is built on structures that exist because, until about sixty days ago, there was no alternative to them.
I'm writing this for anyone who runs a business, works in a business, or is thinking about building one. The pressure test is coming. For some of you it already started. This is what it looks like, why it's happening, and what the other side might be.
I'm also writing it for myself. Because when I published the first version, I was looking at the agency model from the outside, having run one myself and then spent years building systems for agency owners trying to escape it. Now I'm an operator. I'm building in the AI skills space. I'm deploying autonomous systems into business operations. I'm in the weeds every day. And the thing I see from the weeds is that my total addressable market expanded by an order of magnitude in five months, not because I changed what I offer but because the market changed what it needs. My original thesis wasn't only validated by the market, it was simultaneously forced to evolve with it.
That should concern everyone who hasn't started adapting.
And it should be the most exciting thing in the world for everyone who has.
The hollow engine, revisited
The original argument went like this.
An agency sells hours. A client needs a website, a marketing campaign, a brand identity, a piece of software. The agency assigns people to the project. Those people do the work. The client pays for their time. The agency takes its margin. Then the project ends and everyone starts over.
Nothing accumulates. The website lives on the client's server, not the agency's. The marketing strategy belongs to the client, not the agency. The code, the copy, the designs, the research, all of it walks out the door when the invoice gets paid. The agency retains nothing except the experience in its employees' heads, and those employees can leave tomorrow.
This is what I called the hollow engine. A machine that produces output but builds no internal mass. Revenue requires labor. Labor requires management. Management requires the founder. The founder becomes the load-bearing wall in a building that looks impressive from the outside but has no structural integrity beyond what one person can hold up.
The model works until it doesn't. And the moment it stops working, there's nothing underneath. No proprietary technology. No compounding asset. No flywheel that spins on its own. Just a founder, a team, and a Slack workspace full of context that will evaporate the day someone quits.
The thesis argued that the escape route was to distill the founder's domain expertise, the thing that makes the agency's work valuable, into systems and infrastructure. Encode what the founder knows into something that runs without the founder in the middle. Build a system that scales without linear labor growth. Something the founder can step away from. Something that generates recurring revenue, that the business owns as IP, that gets valued like software rather than like a founder who can't take a vacation.
That was the Infrastructure Operator Model. The argument wasn't about AI replacing agencies. It was about agency founders replacing themselves, by turning their expertise into a product that compounds instead of a service that exhausts.
That argument was correct. But it was too small.
The frame that was missing
I framed hollowness as an agency problem because agencies are where I started, where I spent years building, and because the structural weakness is most visible there. But the hollowness isn't unique to agencies. It's a property of any business where the value-producing work is performed by humans who could be performing that work somewhere else, using methods that could be replicated by a sufficiently capable system.
Think about what that actually covers.
An accounting firm where CPAs process returns using the same software and the same rules every year. The firm is hollow. The CPAs are the engine. If they leave, or if a system can do what they do, the firm is a shell.
A logistics company where dispatchers manually reconcile shipments against purchase orders by comparing spreadsheets. The dispatchers are the engine. Everything else is infrastructure around them.
A real estate brokerage where agents follow the same lead nurturing sequence for every inquiry, manually entering data into a CRM and sending templated follow-ups. The agents are the engine. The brokerage is the shell.
A mid-size SaaS company with a customer success team of twelve people who spend 70 percent of their time answering the same questions, escalating the same issues, and running the same onboarding playbook. Those twelve people are the engine. The org chart around them is the shell.
All of these businesses look solid from the outside. They have offices, employees, revenue, clients. But strip away the humans who do the repetitive cognitive work and ask: what's left? What does the business own that produces value independently of the people showing up on Monday morning?
In most cases the answer is: not much.
That's the expanded hollow engine. It's not an agency concept. It's a structural concept. And it applies to any business where the work being done is primarily information processing, rule application, or pattern-based decision-making performed by humans who are, functionally, acting as middleware between systems.
Aerodynamics
Here's a better way to think about it.
Every business is a vehicle. It's trying to go somewhere, to grow, to serve more customers, to generate more revenue, to build something durable. The question is: how much of what the business carries is actually moving it forward, and how much of it is drag?
Before AI agents, you couldn't really answer that question. You needed a marketing department because that's how marketing got done. You needed an accounting team because that's how books got kept. You needed customer support reps because that's how tickets got resolved. Every function required humans, so every human was, by definition, load-bearing. You couldn't remove them without the function stopping.
Now you can ask the question for real. Not hypothetically. Not as a consulting exercise where someone draws boxes on a whiteboard. Actually ask: if I could replace this function with a system that runs autonomously, would the business still work? Would the output be the same? Would the clients notice?
If the answer is yes, that function isn't load-bearing. It's mass.
Mass isn't useless. The humans doing that work weren't doing a bad job. Most of them were doing a fine job. But fine isn't the same as necessary. And when a system can do a fine job at a fraction of the cost, without the overhead, without the management layer, without the attrition risk, without the PTO calendar, the mass becomes drag. It's weight the vehicle is carrying that slows it down relative to a competitor who figured out how to move without it.
This is what I mean by aerodynamics. The businesses that survive the next three to five years aren't the ones with the most resources or the most employees or the biggest org charts. They're the ones that figured out, early, what's actually moving them forward versus what's just along for the ride. The ones that stripped down to what's structurally necessary and let systems handle everything else.
That sounds ruthless. Maybe it is. But the market doesn't care about your org chart's feelings. It cares about output, speed, and cost. And the business that can produce equivalent output at one-tenth the cost with one-fifth the headcount will eventually win. Not because it's morally better. Because the math is the math.
The businesses that get this wrong almost always make the same mistake. They assume they already know what's structural. They protect the functions they're emotionally attached to and cut the ones that seem expendable, which is usually backwards. The thing a founder built with their own hands isn't necessarily the thing that's keeping the business alive. It's just the thing they can't imagine operating without.
The original thesis said agencies were hollow. The expanded thesis says most businesses are carrying weight they don't need, and AI just made it visible by showing everyone what the stripped-down version looks like.
The pressure test is simple. What happens when your competitor figures this out before you do?
What happened in five months
I need to talk about pace because it's the thing that makes this thesis urgent instead of academic. If these changes were happening over five years, everyone would have time to read the white papers, run the pilots, form the committees, and adapt gracefully. They're not happening over five years. They're happening in weeks.
Let me walk through what I watched happen between October 2025 and March 2026.
When I published the original thesis, Claude Code was already impressive but still felt like a developer tool. MCP existed but the ecosystem was thin. Agent frameworks were proliferating but most of them were demos, not production systems. The skills economy was barely a concept. A few people were building autonomous systems. Most people were still treating AI as a fancy autocomplete that occasionally said something wrong.
By December, that had shifted. I was on calls with agency owners who had spent a weekend with Claude Code and come out the other side looking like they'd seen a ghost. Not because the technology was scary in some abstract sci-fi way. Because it was capable. Genuinely, practically, right-now capable of doing work they were charging clients $15,000 a month to perform. One guy told me he couldn't sleep for three days. Another cancelled a hiring round the following Monday. These weren't small operators running lifestyle businesses. These were $100K-a-month agencies with teams of seven to twelve, and they were suddenly doing the arithmetic on what happens when the work that employs six of those people can be handled by a system that costs $200 a month.
By January 12th, Anthropic launched Claude Cowork. Not another chatbot. Not another assistant. AI agents built for specific business functions: data analysis, legal document drafting, meeting preparation, pipeline management, financial reporting. Agents that don't assist workers. Agents that execute workflows independently.
By January 25th, OpenClaw happened. The project I wrote about in the ClawdBot analysis, the one Anthropic hit with a trademark complaint and forced to rebrand. Peter Steinberger, one person, an Austrian developer who'd made his money selling his previous company Nutrient, built an autonomous agent framework that pulled 9,000 GitHub stars on its first day. The rebrand didn't slow it down. The project kept growing.
Then January 30th. Two things on the same day. Anthropic dropped eleven open-source plugins covering legal, sales, finance, marketing, customer support, and product management. Eleven vertical-specific toolkits, available to anyone, all at once. And OpenClaw crossed 150,000 GitHub stars. The corporate side of AI and the open-source side of AI both hit a threshold on the same Friday.
Wall Street noticed. Between January 30th and February 4th, $300 billion in market value evaporated from application software stocks. Five days. Three hundred billion dollars. Because the market understood, instantly, what it meant when Anthropic was shipping production-ready business agents and a quarter of a million developers were adopting an autonomous agent framework simultaneously. The per-seat model wasn't under pressure, it was under attack from both directions at once.
Within two weeks of launch, OpenClaw had 190,000 stars. By early March, 247,000 stars and 47,700 forks. By February 14th, Steinberger had joined OpenAI and the project was moving to a foundation.
OpenClaw wasn't a business. Steinberger was spending $10,000 to $20,000 a month out of pocket to run it. Open source, no revenue model. None of that matters as a business story. All of it matters as a market indicator. One person, working with the tools that now exist, built something that a quarter of a million developers adopted in weeks. The distance between concept and global adoption compressed to a timeline that makes traditional product development look like a process from a different century.
Then Nat Eliason gave an AI agent a thousand dollars and told it to run a business. The agent, Felix, built a website, created products, launched a marketplace called Claw Mart, opened a Stripe account, and started processing payments. It's pulled in nearly $80,000 in revenue since early February. Its annual run rate is approaching a million dollars. It runs on a $200-a-month Claude Pro Max subscription. An AI agent. Running a business. Collecting revenue. Managing operations. With minimal human oversight.
Nat went on Bankless and the episode title was "Building a Million Dollar Zero Human Company with OpenClaw."
Is Felix a real business in the way we traditionally define that term? I genuinely don't know. Maybe it's a proof of concept. Maybe it's a stunt that happened to work. But the fact that it's possible, that an AI agent can perform the actual mechanical functions of running a small business and generate real revenue doing it, that fact is the one that keeps me up at night. Not because I'm afraid of it. Because it validates everything in this thesis at a speed I didn't think we'd see for years.
Meanwhile, every week between January and March, new agent frameworks shipped. New MCP servers appeared on GitHub. New skills ecosystems launched. The "I built X with AI in Y days" posts became so relentless they stopped being remarkable. Someone shipping a complete SaaS product in a weekend became background noise. That's how fast the baseline moved. The thing that would have made headlines six months ago became a Tuesday.
I'm in group chats with people who are doing tens of millions a year. Operators who've built and sold companies, who run portfolios of products, who've been through every hype cycle the internet has produced in the last fifteen years. These are not excitable people. They don't get swept up in trends. They're the ones who usually sit back and wait for the dust to settle before they move.
Every single one of them is saying the same thing right now. AI is legitimately changing everything. Not "AI is interesting." Not "AI has potential." AI is changing everything, right now, today, in ways that affect their businesses, their teams, their cost structures, and their competitive positions. I've never seen this group unanimous on anything. They agree on this.
I've been in this space for five years. I've watched funnel builders come and go. Chatbots. No-code tools. AI copywriting. Programmatic SEO. None of them moved like this. Not a single one. And none of them had this kind of implication. Those were tools that helped businesses do what they already did, slightly faster or slightly cheaper. Incremental improvements to existing processes.
This replaces the processes themselves.
The first real replacement for human labor
Here's something that doesn't get said enough. Before AI agents, there was never a real replacement for human cognitive labor. Not once. Not in the entire history of commerce.
Every previous wave of automation replaced physical tasks. Machines on factory floors. ATMs instead of tellers for basic transactions. Self-checkout at the grocery store. Robotic arms in warehouses. All of these replaced the body. None of them replaced the brain.
And for cognitive work, the work that happens in offices, the work that involves reading, deciding, writing, communicating, analyzing, planning, there was never a replacement. There was only relocation.
Outsourcing to India or the Philippines wasn't replacing labor. It was buying the same labor from a different labor market at a lower price point. The work was still done by a human brain. A human who read the email. A human who made the decision. A human who wrote the response. A human who processed the invoice. You were arbitraging geography, not capability. You still needed a person. You just paid them less because they lived somewhere with a lower cost of living.
That's not replacement like people like to think it is. That's relocation or optimization. The labor was still human. The work was still cognitive. The person was still in the loop.
What's happening now is different in kind, not just in degree. For the first time, the cognitive work itself, the reading, deciding, writing, communicating, analyzing, planning, can be performed by a non-human system. Not a system that assists a human in doing the work. Not a system that suggests what the human should do. A system that does the work. Receives the input. Applies the judgment. Produces the output. Routes it to the next step. Without a human in the middle.
This is the line that doesn't un-cross.
When the work itself can be performed by a system, the entire economic structure built around employing humans to do that work comes under pressure. Not theoretical pressure. Not "in a few years" pressure. Actual, measurable, showing-up-in-earnings-calls pressure. SaaS companies reporting reduced seat counts. Agencies reporting client attrition to self-service models. Consulting firms watching their pipeline change shape as clients realize they can handle more in-house than they thought.
The pressure isn't coming. It arrived while people were still arguing about whether AI could write a decent email.
But replacement is the wrong word at scale
I need to be precise here because the easy version of this argument is wrong, and I refuse to make the easy argument.
The easy version says: AI replaces all workers, nobody has a job, everything changes overnight. That's not what's happening. Not in the structural sense.
At small scale, a solo founder or a five-person startup, AI agents are a near-total replacement for many business functions. The cost is negligible. $75 to $150 a month for an AI stack that handles content creation, customer support, financial analysis, design, and development work. At that scale, the math is so lopsided it barely needs arguing. A solo founder can now operate a business that would have required a team of ten, for the cost of a nice dinner out each month. The businesses that choose not to use these tools will be outcompeted by the ones that do, and the gap will widen every quarter as the tools improve.
At mid-scale, fifty to five hundred employees, the picture gets more complicated. AI doesn't replace everyone. It replaces the need for many of them. The customer success team of twelve becomes a team of four, with agents handling the routine 70 percent and humans handling the edge cases and relationship work. The marketing department of eight becomes a department of three, with agents running campaigns, generating content, and analyzing performance while humans set strategy and make judgment calls that require actual understanding of the market. The finance team gets leaner. The operations team gets leaner. Every function that involves significant pattern-based cognitive work gets leaner.
Fewer employees. Not zero employees. And the ones who remain are doing higher-value work, the work that genuinely requires human judgment, creativity, and relationship management.
That transition, from twelve to four, from eight to three, doesn't happen by buying software. It's a redesign. The system has to be built before the team can shrink, because if you shrink first and build second, you lose the institutional knowledge the system was supposed to capture. The sequencing matters as much as the technology. Get it wrong and you don't get a leaner business. You get a broken one.
At enterprise scale, a thousand employees and up, the economics shift again. And this is where the "AI is practically free" narrative starts to crack.
Running AI systems across an organization of that size requires infrastructure. Governance frameworks. Security architecture. Compliance systems. Integration work. Monitoring and audit trails. A human oversight layer for high-stakes decisions. The mid-size enterprise sees the steepest cost escalation when it moves from isolated AI pilots to integrated, multi-departmental deployment. By the time you add up the API costs, the infrastructure, the governance, the security, the integration, and the human oversight, the total cost of an AI-augmented operation at that scale can approach, and in some cases match, the cost of the human operation it replaced.
The value is still real at that scale. The AI-augmented operation is faster, more consistent, operates around the clock, and scales without linear headcount growth. But the savings story changes. It stops being "replace expensive humans with cheap systems" and becomes "replace slow, inconsistent processes with fast, consistent ones at comparable total cost." That's a different pitch. It's still a good pitch. But it's not the "everything is free now" story that gets the retweets.
So the thesis isn't "AI replaces everyone." That's lazy thinking dressed up as insight. The thesis is that AI makes every business leaner, and the degree of leanness depends on scale. At small scale, it changes everything. One person does the work of ten. At mid-scale, teams shrink but output doesn't. At enterprise scale, the math is more complicated but the direction is the same: fewer people, better systems, faster cycles.
The throughline is identical at every level. Every business is about to carry fewer humans and more systems. The ratio shifts. The weight distribution changes. And the businesses that figure out the right ratio first have an advantage that compounds over time, because unlike labor, systems don't get tired, don't quit, and don't need to be retrained every time someone leaves.
The liability gap nobody has figured out
There's a problem underneath all of this that hasn't had enough time to surface, and I think it's going to be one of the defining issues of the next two to three years.
AI can't be held responsible for its mistakes. Responsibility can fall on the people or organizations developing, supplying or deploying it, depending on the facts and applicable law.
This sounds obvious when you say it out loud. Of course the machine can't be sued. But think about what it means in practice when you're deploying these systems into real business operations where real money moves and real decisions get made.
When an employee makes a mistake, there's a chain of accountability. The employee made the decision. Their manager oversees them. The company has processes, insurance, legal frameworks built over decades to handle errors and negligence and liability. The whole structure assumes that a human made a judgment call and can be held accountable for it.
When an AI agent makes a mistake, the agent can't be held accountable. It has no legal personhood. It can't be sued, fined, disciplined, or terminated in any meaningful sense. The person or company that deployed the agent can face liability, alongside other responsible parties where the law provides. But the controller might not have known the agent would make that specific decision. Might not have been able to predict it. The agent made a judgment call based on its training data and the context it had access to, and the judgment was wrong, and now someone has to answer for it. And the law looks to responsible people and organizations, which can include both the deployer and the provider.
The EU is trying to get ahead of this. The AI Act entered into force in August 2024, with provisions applying in stages and a general application date of August 2026. A new Product Liability Directive is being implemented across member states by December 2026. These are steps in a direction. They're not solutions. They are adopted laws addressing a problem that nobody fully understands yet because the technology is moving faster than legislators can think, let alone write law.
In practice, what this means is that full autonomy, the "no human in the loop" model, works for low-stakes operations. Drafting routine emails. Scheduling meetings. Generating internal reports. Content creation for social media. The stakes are low enough that a wrong output is an inconvenience, not a lawsuit.
For high-stakes operations, financial transactions, legal analysis, medical advice, hiring decisions, compliance reporting, the liability question puts a hard limit on how much autonomy you can responsibly hand over. Someone has to be in the loop. Someone has to be able to say "I reviewed this and approved it," because when things go wrong, and they will go wrong, that's what the legal framework is going to demand.
The fully autonomous business is possible for certain functions and certain risk levels. It is not possible for everything. Not yet. Maybe not ever for the highest-stakes decisions. And the people building and deploying as though liability doesn't exist, the ones running agents with no guardrails and no audit trails and no human oversight on consequential decisions, they're not building a competitive advantage. They're building a liability that compounds quietly until the first lawsuit surfaces.
When that lawsuit lands, and it will, it'll affect everyone in the space. The responsible operators and the reckless ones alike. The regulations that follow will be written in response to the worst actors, not the best ones. That's how regulation always works.
This hasn't had time to pan out yet. The technology is months old in its current form. Existing legal frameworks apply, while AI-specific rules are being phased in. We're in a gray zone where the capability far outpaces the governance, and the gap is widening, not shrinking.
I don't have a clean answer for this one. Nobody does. But I'd rather name it honestly than pretend it doesn't exist.
SaaS stopped being a product
This is the part that nobody saw coming. Not the agency owners. Not the SaaS founders. Not the VCs. Not me.
To recap: On January 12th, Anthropic launched Claude Cowork. On January 30th, they dropped eleven open-source plugins for legal, sales, finance, marketing, customer support, and product management. Five days later, $300 billion in market value had evaporated from application software stocks. By mid-February, the total damage was roughly one trillion dollars wiped from the collective value of the software sector. The press called it the SaaSpocalypse.
I need to explain why Anthropic releasing a set of plugins caused a trillion-dollar wipeout, because the surface-level explanation is misleading and most of the commentary I've read gets it wrong.
The surface-level story goes: AI is coming for SaaS, companies will build their own tools, nobody will need software anymore. That's wrong. It's not what's happening and it misreads the data badly.
Here's what's actually happening.
SaaS companies sell seats. A CRM costs $50 per user per month. An email platform costs $30 per user per month. A project management tool costs $25 per user per month. The pricing model, the entire financial model that Wall Street values these companies on, assumes that more humans doing more work means more seats, which means more revenue. Growth equals hiring, hiring equals seats, seats equal ARR. That equation has powered SaaS economics for twenty years.
AI agents don't use seats.
An agent that manages your CRM, sends your emails, coordinates your projects, and analyzes your data doesn't need four subscriptions at $130 per month per user. It needs API access. That's a different product at a different price point with different unit economics. And the difference is large enough that it broke the financial model the entire industry was built on.
Companies started reporting in Q4 2025 that their customers were reducing seats. Not churning. Not leaving the platform. Reducing the number of humans who needed access to it. Because they'd deployed agents that did the work of five people, and five people no longer needed to be sitting in front of the software. The function still happened. The CRM still got updated. The emails still went out. The reports still got generated. But the human in the seat disappeared, and the seat revenue went with them.
The cost of doing the work didn't vanish when the seat did. It shifted from the SaaS invoice to the compute bill, and that math is less settled than anyone selling the transition wants to admit.
That's why the valuations crashed. Not because SaaS is dying. Because the financial model that SaaS has been valued on for two decades has a structural crack in it. And once the market saw the crack, a trillion dollars evaporated in a month.
But SaaS isn't dying. It's something more interesting than dying.
Here's where the analysis gets genuinely interesting and where most of the takes I've read miss the point entirely.
The SaaSpocalypse narrative assumes that if seat-based SaaS declines, SaaS itself declines. That's like saying if people stop buying CDs, music dies. Music didn't die. It moved to streaming. The underlying demand for music was as strong as ever. The packaging changed. The business model changed. The consumption pattern changed. But the thing itself, the music, was never going anywhere.
SaaS is going through the same transition right now. The demand for CRM functionality, email functionality, analytics functionality, project management functionality, none of that is going anywhere. Businesses still need these capabilities. They need them more than ever, because agents running autonomously require reliable, well-built tools to connect to. The agent can't orchestrate nothing. It needs services to orchestrate.
What's changing is the packaging.
Think about what Stripe did to payments. Before Stripe, if you wanted to process payments online, you bought a payment processing product. You negotiated with a payment provider. You managed a vendor relationship. You sat through demos and signed contracts and dealt with a sales team. The product sat on your desk, figuratively speaking, as a standalone thing you interacted with through its own interface on its own terms.
Stripe turned that product into a component. An API call. Seven lines of code and you had payment processing. You didn't manage a vendor relationship. You didn't navigate a dedicated interface. You didn't sit through a demo. You wired it in and moved on. The payment capability was as powerful as ever. More powerful, actually, because the composability opened up use cases that the standalone product never could have addressed. But it stopped being a product you bought and started being infrastructure you plugged into.
That's what's happening to the entire SaaS layer right now. All of it. At the same time.
Every standalone tool, your CRM, your email platform, your project management system, your analytics suite, your bookkeeping software, your scheduling tool, your form builder, your help desk, is in the process of becoming a component. A capability that gets wired into a larger system. The tool doesn't go away. It loses its sovereignty. It stops being the thing you open every morning and interact with through a dedicated UI, and becomes the thing your agent calls when it needs that specific function performed. The UI becomes irrelevant because no human is clicking the buttons.
The average number of SaaS applications per organization dropped from 112 to 106 by April 2025. That number is still falling. Not because companies need less software. Because they need less UI. The software still runs. The functions still execute. The data still flows. The human just stopped being the one sitting in the middle making it happen.
The magnetization
I keep coming back to this word because I think it's the most accurate description of what's happening to the software ecosystem.
Every tool is getting pulled toward a center of gravity. The center is the orchestration layer, the agent system that connects the tools together and runs the business logic between them.
Before, each tool operated on its own continent. You opened your CRM in one tab, your email in another, your analytics in a third, your project management in a fourth. You were the integration layer. You were the thing that looked at the CRM data, decided what email to send, checked the analytics to see if it worked, and updated the project board with the results. You were the middleware. The human brain sitting in the center of a dozen tools, context-switching between them, moving information from one system to another, applying judgment along the way.
Now the agent is the middleware. It sits in the center and connects to each tool through MCP or API. It looks at the CRM data, decides what action to take, checks the analytics, updates the board, sends the follow-up, logs the result. Same workflow. No human in the middle. No context-switching. No 9-to-5 constraint. No "I'll get to that tomorrow." The workflow runs continuously.
And once the agent is the middleware, the individual tools start losing their independent identity. The CRM isn't "your CRM" anymore. It's the data layer the agent queries for contact information. The email platform isn't "your email tool" anymore. It's the delivery mechanism the agent uses to send messages. Each tool becomes a function call in a larger program. It becomes a component.
This is the magnetization. The individual pieces are getting pulled into a unified field. They're orienting around a center, the orchestration layer, and their identity as standalone products is fading. The value isn't in the individual tool anymore. It's in the composition. The orchestration. The system that makes all the pieces move as one.
SaaS companies that understand this will adapt. They'll build for composability, for API-first access, for agent consumption rather than human consumption. They'll price on usage and outcomes rather than seats. They'll stop thinking of themselves as products and start thinking of themselves as infrastructure. Some will thrive in that model. Stripe is proof that becoming infrastructure can be wildly valuable. More valuable, in fact, than being a standalone product ever was.
SaaS companies that don't understand this, that keep building for human users who sit in front of screens and click buttons, that keep pricing on seats and measuring success by how many humans log in each day, they're building for a market that's getting smaller every month.
The composition is the product now. The individual piece is just a component. And the market is going to price accordingly.
The question that follows is one nobody's spending enough time on: who does the composing? The tools exist. The components exist. The orchestration layer exists. But wiring them into a system that actually runs a specific business, that handles the operational logic of a specific company in a specific vertical, that doesn't break when the edge cases show up? That's not a prompt you type into a chatbot. That's the kind of knowledge you get by doing it, across enough different businesses that you start seeing which patterns transfer and which ones are unique to the context.
The expertise transfer
This is the one that rewrites the economics of everything else in this thesis, and almost nobody is framing it correctly.
The conventional take goes like this: AI tools help experts work faster. A senior marketer uses AI to write copy in ten minutes instead of two hours. A developer uses AI to ship features twice as fast. A consultant uses AI to produce a deck in an hour instead of a week. The expert is still the expert. They just have a faster tool. Productivity story. Efficiency play. The world looks basically the same, just sped up.
That take is true. And it's the least interesting thing happening.
The more interesting thing, the thing that actually rewrites the economics, is what happens when a non-expert gets their hands on the same tools. Not a watered-down version designed for beginners. Not a simplified interface with training wheels. The same tools, with the same capabilities, running the same models, with the same access to the same information.
And I don't mean they produce garbage that kind of looks like the real thing if you squint. That was true two years ago. The output was obviously AI-generated and anyone with domain knowledge could spot the gaps in about four seconds. What's happening now is different. A non-expert using the right tools, with the right skills configured, with the right context loaded, produces work that is genuinely, defensibly competent. Not expert-level. But competent. Usable. Deployable. Commercially viable. The kind of output that, eighteen months ago, required someone with five to seven years of domain experience and a $90,000 salary.
A founder with no marketing background can now build a content pipeline, keyword strategy, email sequence, and analytics dashboard that operates at 50 to 60 percent of what a seasoned marketing director would produce. Not because the founder suddenly learned marketing. Because the domain expertise got encoded into the system. The skills, the frameworks, the decision trees, the heuristics, the pattern recognition that used to live exclusively in someone's head after years of practice, it's in the tools now. It's in the prompts. It's in the agent's training data and its skill architecture and the MCP servers it connects to.
The expertise left the expert. It lives in the system now.
The human who built the system still matters. The person who encoded the expertise, who structured the skills, who understood the domain well enough to capture the right patterns. That person is essential. And building these systems well, building them so they actually hold up in production and not just in a demo, requires having done it across enough domains to know what transfers and what's specific to each vertical. The encoding isn't a template you fill in. It's pattern recognition earned through repetition. But the person operating the system? The person running the business that uses the system? They don't need the expertise. They need access to the system that contains the expertise. Different thing.
Why good enough wins
Here's why 50 to 60 percent isn't a failure number. It's the number that changes everything.
Most small and mid-size businesses were never buying the top 40 percent of expertise anyway. They couldn't afford it. A Fortune 500 company hires McKinsey. A mid-market company hires a respected regional firm. A fifty-person manufacturing company? They hire a local agency or a mid-career marketing manager. They were already getting 50 to 60 percent of expert-level output because that's what their budget bought them.
Now they can get the same quality from a system that costs $150 a month instead of $5,000 a month for the agency retainer. Or $8,000 a month for the mid-career hire once you factor in salary, benefits, payroll tax, office space, and the management overhead of having another employee.
Same quality tier. 95 percent lower cost. Call it optimization if you want, but that's a structural break.
I know what the objection is. "But the AI can't do what a real expert does." Correct. It can't. The top tier of expertise, the genuine strategic insight, the creative problem-solving that requires novel thinking, the judgment calls that depend on understanding context that can't be captured in a training dataset, that's still human territory. I don't think anyone serious argues otherwise. Maybe that top tier is always human territory. I'm not going to speculate on "always" because six months ago I was speculating about "three to five years" and the market proved me wrong.
But that top tier is maybe 20 to 30 percent of what most experts actually do in a given week. The rest, the 70 to 80 percent, is pattern application. It's experienced decision-making following established frameworks. It's doing the thing you've done hundreds of times before, with slight variations for the current context. Writing the email you've written a thousand versions of. Running the analysis you've run for fifty other clients. Applying the methodology you learned five years ago to a situation that's similar enough to all the other situations that the methodology works fine.
That layer is getting systematized. Encoded. Distributed at near-zero marginal cost. Available to anyone with access to the right tools.
And the businesses that were paying for the full expert were mostly paying for the pattern-application layer anyway. They just had to buy the whole package because there was no way to separate the strategic insight from the routine execution. The expert came as one unit. You hired the person, and you got both the brilliant insight and the routine work in the same salary. There was no mechanism to buy just the part you needed.
Now the routine execution is available separately. From a system. For less than the cost of the expert's daily coffee.
The expert who provides genuine strategic insight still has value. Maybe more value than before, because the commodity work disappears and the premium contribution stands out more clearly. The best marketers, the best developers, the best strategists, will likely make more money, not less, because they'll be free from the commodity work that used to eat most of their week and they can focus entirely on the part that only they can do.
But the average expert? The one who was really doing pattern application with a credential attached? That person's economic position just shifted. The thing they were selling, competent execution of established patterns, is now available from a system that costs $150 a month and doesn't need a benefits package.
I don't enjoy writing that paragraph. But I'd rather describe the world as it appears to be than as I'd like it to be.
The numbers that tell the story
Danny Postma built HeadshotPro, an AI headshot generator, to over $3.5 million ARR. By himself. No team. No investors. 300,000 users, $300K in monthly revenue.
Eric Borgos built BoredHumans to roughly $8.8 million ARR. Solo. Ad-supported plus premium tools. Launched in 2020, printing money by 2025.
Nat Eliason's FelixCraft agent has generated nearly $80,000 in revenue since early February, running on a $200-a-month subscription, with a stated goal of $1 million by end of year.
These aren't outliers anymore. They're data points in a trend.
Solo-founded startups jumped from 23.7 percent of new companies in 2019 to 36.3 percent by mid-2025. There are 41.8 million solopreneurs in the United States contributing $1.3 trillion to the economy. A complete AI stack for solo business operations, covering content creation, customer support, financial analysis, strategic planning, coding, and design, costs between $75 and $150 a month. That represents a 90 to 95 percent cost reduction compared to hiring part-time employees for those same functions.
Dario Amodei, the CEO of Anthropic, was asked at the Code with Claude conference when the first billion-dollar company with a single human employee would appear. His answer was 2026. He said he was 70 to 80 percent confident.
The expertise didn't disappear. It got encoded, packaged, and distributed to people who never had it before. And those people are using it to build businesses that compete with, and in some cases outperform, the ones staffed by credentialed experts. Not because the solo founder is smarter or more experienced. Because the system is capable enough and the overhead is low enough that leanness beats expertise at most scales below enterprise.
The most qualified person in the room doesn't always win anymore. The most efficiently configured system does.
Skills are everything
If the expertise is leaving the expert and entering the system, the next question is: what does the system actually look like? What's the mechanism that captures expertise and makes it executable by someone who doesn't have it?
The answer is skills.
Not skills in the human capital sense. Not "five years of marketing experience" or "proficient in Excel." Skills in the AI architecture sense. A skill is a packaged capability. A set of instructions, context, tools, and decision frameworks that turns a general-purpose AI into a specialist for a specific task. It's what makes the difference between "I have Claude" and "I have Claude that knows how to run a B2B content marketing operation using my company's voice, my industry's terminology, my audience's pain points, my competitor analysis, and my conversion framework."
That's not the same thing. "I have Claude" is a generalist with broad knowledge and no specific context. "I have Claude with the right skills" is a practitioner. Not a world-class practitioner. The 50 to 60 percent practitioner that, as I argued in the previous section, is good enough for the vast majority of the market.
The skill is the container. It's the thing that holds the domain knowledge, the decision frameworks, the quality standards, the edge-case handling, the operational context that an expert would normally carry in their head after years of practice. Package all of that into a skill, attach it to an AI agent, hand the agent to a non-expert, and the non-expert can operate at a level that used to require years of training.
The skill is what makes the expertise transfer possible. Without skills, AI is a generalist. With skills, it's a capable operator in a specific domain. The skill is the bridge between "AI is impressive but not useful for my specific situation" and "AI is running this function of my business better than the junior hire I was going to make."
Why skills matter more than models
There's a misconception in the market right now that the model is what matters. People compare Claude to GPT to Gemini to whatever shipped last Tuesday. They benchmark reasoning abilities and coding performance and context windows. They debate which model is "best" like the model is the finished product.
The model isn't the product. The model is the engine. And an engine without a vehicle doesn't go anywhere.
The vehicle is the skill. The context. The tools. The orchestration. The thing that turns raw model capability into actual business output. The thing that points the engine at a specific problem and gives it everything it needs to produce a specific result.
A good model with well-built skills will outperform a great model with no skills. Every time. Because the skill is what converts general intelligence into specific capability. The skill is what knows that in the roofing industry, leads respond best to a particular kind of follow-up within forty-eight hours. The skill is what knows that in B2B SaaS, onboarding emails should address the top three objections before the trial expires. The skill is what knows that in legal compliance, certain disclosures have to appear in a specific format in a specific location on a specific type of document.
None of that knowledge lives in the base model. It lives in the skill. And the person who creates that skill, the person who had the domain expertise and took the time to encode it into a transferable format, that person has captured something genuinely valuable. Something that gets more valuable as more people need it. Something that gets more powerful as the models it runs on improve.
That last point is worth sitting with. A skill built for Claude 3.5 runs better on Claude 4. The creator doesn't have to rebuild. The underlying engine got more powerful and the skill inherited that power automatically. That's compounding. That's an asset that appreciates over time without additional labor.
Labor doesn't compound. You sell an hour, the hour's gone. You sell another hour, you start from scratch. No accumulation. No appreciation. Just time exchanged for money until the energy runs out.
Skills compound. That's the same insight from the original thesis, the one about agencies producing value only in motion, applied to a new layer. The agency was hollow because nothing accumulated. The skill model isn't hollow. It builds on itself. Each skill that gets created, refined, and deployed makes the overall ecosystem more capable. Content skills connect to SEO skills connect to analytics skills connect to email skills, and the composition becomes more powerful than any individual skill.
When I saw this, when the pattern became clear, I started building SkillStack. A marketplace for AI skills. Not because I wanted to build a marketplace. Because the thesis demanded it. If skills are the atomic unit of the new economy, if they're the mechanism that makes the expertise transfer real, if they're the thing that turns hollow engines into compounding systems, then the infrastructure for creating, distributing, and composing skills is the most important thing to build right now. That's the action I took when the direction became clear.
The decision to build it came directly from the analysis in this document. Expertise is leaving experts and entering systems. Skills are how it travels. The platform that facilitates that transfer is sitting at the center of the biggest shift I've seen in my career.
The skill economy is the expertise economy, restructured
Here's where the threads on skills come together.
In the old economy, expertise was locked inside people. You couldn't buy it without buying the person's time. You couldn't scale it without hiring more people with the same expertise. You couldn't transfer it without years of mentorship and practice. The expert was simultaneously the bottleneck and the asset. You needed them, and they knew it, and the entire compensation structure for knowledge work reflected that leverage.
In the skill economy, expertise gets encoded into a transferable, deployable unit. The expert creates the skill. The skill gets deployed by non-experts. The expertise scales without the expert's ongoing involvement. The bottleneck breaks.
The expert still has a role. An important one. Someone has to create the skill in the first place. Someone has to know the domain well enough to encode the right decision frameworks, the right quality standards, the right edge-case handling, the right operational context. That person needs to be genuinely good at the thing. You can't encode expertise you don't have. The skill is only as good as the domain knowledge that went into building it.
But the expert's relationship to their own expertise changes. Instead of selling hours of execution, they sell the encoded version of what makes their execution valuable. Instead of trading time for money on a linear basis, they create an asset that generates value independently. Instead of being the hollow engine, doing the work, collecting the fee, starting over, they're building the machine that replaces the hollow engine.
The agency model said: "Pay me to do this for you, over and over, forever."
The skill model says: "Here's the capability to do this yourself."
One is labor. The other is product. And product scales in a way that labor never could, never has, and never will.
The junction point
Most people still use ChatGPT as a search engine.
I want that sentence to land, so let me say it again. The most powerful general-purpose cognitive tool ever created, and the average user types a question into it and reads the response like they're using a fancier version of Google. They don't build agents. They don't create skills. They don't wire autonomous systems into their business operations. They don't compose tools into orchestrated workflows. They ask it a question, read the answer, and close the tab.
That's like using the internet to check the weather in 1996.
The obvious comparison is the internet in 1996. Everyone makes it. For once it actually holds — not because the technology is similar but because the evaluation error is identical. People looked at the early web, saw a slow and ugly toy, and were right about the present state but wrong about the trajectory by a factor that cost them the next two decades. The difference now is speed. The internet took ten years to go from toy to infrastructure. AI agents went from demos to production systems in months. The window between "early adopters have an advantage" and "everyone figured it out" is compressing in a way that doesn't leave room for the wait-and-see playbook.
We're at the same junction point. Right now.
The people who see AI as a chatbot, a search engine, a thing that writes mediocre essays and occasionally hallucinates, they're making the same evaluation error. They're looking at the present state and seeing a toy. They're judging it against existing solutions, their current team, their current tools, their current processes, and concluding that it's not worth the disruption.
Meanwhile, the people who figured out what the tools can actually do are building the next generation of businesses with a fraction of the overhead and a fraction of the headcount. They're operating at a level of efficiency that the people still debating whether AI is "ready" can't compete with. And they're getting further ahead every week, because the tools keep improving and their skill in using them keeps compounding.
The window is open. The people who figured out the internet early, who saw the trajectory before it was obvious, who built while everyone else was debating, those people built the companies that defined the next two decades. Amazon. Google. The infrastructure layer that the rest of the economy was eventually built on. The same kind of window exists right now. And like 1996, it won't stay open. Once this becomes obvious to everyone, and it will, the advantage of being early disappears. The arbitrage closes. The gap between the people who get it and the people who don't shrinks until it's too small to matter.
Right now, the gap is as wide as it's ever going to be.
The businesses I'm watching close that gap the fastest all did the same thing. They didn't try to become AI experts overnight. They found someone who had already built and installed these systems across enough contexts to know where it works clean and where it breaks, and they moved on that operator's experience instead of trying to develop their own from zero.
What's still unresolved
I need to talk about the parts of this that I can't resolve. Because the temptation with a thesis like this is to present it as clean and inevitable, and it's neither. There are real problems, real limitations, and real open questions that anyone building in this space needs to face honestly.
Security is still a disaster
I wrote about this extensively in the analysis I published alongside the ClawdBot article, and nothing has improved. If anything, it's gotten worse because the adoption rate has accelerated while the security infrastructure hasn't kept pace.
The core problems remain. Prompt injection has no solution. Not a bad solution. No solution. Every document, email, web page, and PDF an agent processes is a potential attack vector for manipulating the agent's behavior. And nobody, not Anthropic, not OpenAI, not any security research lab anywhere on earth, has demonstrated a reliable defense against it.
Community-contributed skills and MCP servers run with the user's permissions and in most setups have no sandboxing. Credentials get stored in plaintext in config files that malware is already targeting. Major infostealer families, RedLine, Lumma, Vidar, have adapted specifically to sweep agent configurations. OpenClaw gateways have been found broadcasting their presence on local networks, leaking filesystem paths, usernames, and hostnames to anyone listening.
247,000 stars. Thousands of exposed gateways. Millions of non-technical users running autonomous systems they don't fully understand, configured with skills they downloaded from anonymous GitHub accounts, connected to their email, their bank accounts, their business operations.
The gap between what's being deployed and what's being secured is the widest I've seen in any technology cycle. And the industry is scaling on enthusiasm, with security as an afterthought that nobody wants to slow down for because the gold rush is too loud to hear the warnings.
I wrote a full analysis of the security situation in January. I won't reproduce it here. But I'll say that everything in this thesis, every argument about expertise transfer and SaaS-as-component and lean business models, assumes that the underlying systems are trustworthy. When they're not, the efficiency gains come with a hidden risk premium that most operators aren't pricing in.
The top tier is still human
I've argued that AI transfers 50 to 60 percent of expert-level capability to non-experts. That leaves 40 to 50 percent that it doesn't transfer. And that 40 to 50 percent is the part that matters most when the stakes are highest.
Genuine creative insight. The ability to see a problem that nobody else has framed. The ability to connect two unrelated ideas in a way that produces something genuinely new. The kind of strategic thinking that doesn't follow a framework because the framework doesn't exist yet, because the thinker is creating the framework for the first time.
AI is very good at pattern application. It's trained on patterns. It excels at recognizing situations it's seen before, or situations similar enough to what it's seen before, and applying the appropriate response. That's most of what most experts do most of the time, which is why the 50 to 60 percent transfer works.
But it doesn't generate genuinely novel thinking. It doesn't have intuition in the way that a great designer or a great writer or a great strategist has intuition. It doesn't know when to break the rules because the rules don't apply to this specific situation. It doesn't feel that something is off before it can articulate why. It doesn't have taste.
This matters because it means the expertise economy doesn't disappear. It bifurcates. The bottom 60 percent of expert work, the pattern application, gets systematized and distributed to anyone with access to the right tools. The top 40 percent, the genuine novel contribution, becomes more valuable because the commodity layer beneath it is no longer scarce. When everyone has access to competent execution, the thing that differentiates you is the thinking that produces the strategy the execution follows.
The best people in every field will likely come out of this better positioned than before. The average people, the ones whose primary value was competent execution of known patterns, face a harder adjustment.
I'm not going to sugarcoat that. It's a real displacement. It's happening. And the right response isn't to pretend it isn't happening. It's to move up the value chain, toward the work that requires the judgment and creativity that systems can't replicate. For the people who can make that move, this is the best time to be alive. For the people who can't, or won't, the math is about to get uncomfortable.
At scale, the cost math gets complicated
The "AI is practically free" narrative works at small scale. A solo founder paying $20 or $200 a month to replace tasks that used to take hours? Real savings. Genuine value. But that number doesn't stay small when you multiply it across an organization running agents on dozens of functions, processing millions of tokens a day, operating around the clock. The pitch changes from "cheaper" to "better at comparable cost." That's a different conversation. And the people selling the cheaper version are setting up expectations that will crack when the invoices show up.
I touched on this in the SaaS section. The seat revenue disappeared, but the cost of doing the work didn't. When a company replaces twelve people with agents, the salary line drops. The compute line appears. At small scale you barely notice it. At enterprise scale, with agents running continuously across sales, support, ops, finance, legal, processing millions of tokens per day, the compute bill starts to look like a payroll line of its own. Different shape. Similar weight.
Look at the numbers right now. Anthropic's $200/month Max plan reportedly costs somewhere around $5,000/month to supply. The economics are inverted at the provider level. That's the same playbook Uber and DoorDash ran, subsidize the unit economics to capture market share, figure out profitability later. Someone absorbs that cost differential eventually. Either prices rise, usage gets throttled, or scale efficiencies materialize that haven't materialized yet. The doctrine's "near-zero marginal cost" framing depends on providers continuing to eat the loss. That's a bet, not a fact. And the people building their entire operational model on current pricing are building on someone else's subsidy.
Underneath all the software is a physical layer that doesn't compress the way software does. Nvidia controls somewhere between 80 and 90 percent of the AI chip market. H100s cost $25,000 to $40,000 each. Data center power consumption is projected to double or triple by 2028, depending on whose estimates you trust. Microsoft, Google, and Amazon are signing nuclear power deals. Actual nuclear power deals. The infrastructure that makes "AI is practically free" possible at the user level requires billions of dollars in hardware, energy, and cooling that somebody is paying for. The marginal cost of one more API call is small. The fixed cost of the infrastructure that serves it is enormous. Those two facts coexist uncomfortably.
Both things are true. The marginal cost of packaging and distributing a skill is genuinely near-zero. The skill itself, the encoded expertise, the workflow, the logic, costs nothing to replicate once it's built. That part of the thesis holds. But the marginal cost of running that skill at scale, the compute, the tokens, the infrastructure, that's not near-zero. It's real and it grows with usage. At small and mid scale the math works cleanly. $150 a month replacing $8,000 a month in labor is a real number that real businesses are seeing right now. At enterprise scale the value proposition shifts from savings to capability. You're not paying less. You're getting something you couldn't get at any price before. Speed, consistency, scale, 24/7 operation. That's worth paying for. But it's a different argument than "practically free."
The businesses that budget for real compute costs, that model the actual unit economics instead of the subsidized ones, those are the ones that will still be running when the subsidy math corrects. The cost isn't a reason not to build. It's a reason to build with numbers that survive contact with an invoice.
Where the road leads
The transition looks different depending on what you're running and at what scale. But the direction is the same for everyone.
If you're an agency owner, the move is the one I outlined in the original thesis, and it's more urgent now than when I wrote it. Stop selling hours. Take the thing you know, the domain expertise your clients are actually paying for, and encode it into a system. Build the infrastructure that delivers the result without you personally doing the work. The agency that's still running on labor in 2027 won't exist. The agency that turned its expertise into a productized system, that owns IP, that generates recurring revenue, that has a software multiple instead of a founder-dependent valuation, that's the one that survives. The tools to make that transition exist right now. They didn't a year ago.
If you're running an operations-heavy business, manufacturing, logistics, distribution, the opportunity is in the processing layer. Every business like yours has a layer of humans reconciling data, routing information, managing exceptions, updating systems. That's the drag. The load-bearing work is the thing your business exists to do, making the product, moving the goods, serving the customer. The processing layer around it is what gets systematized first. Not the core. The scaffolding.
If you're mid-size, fifty to five hundred people, the play is surgical. You don't need to transform the whole company at once. You need to identify the three to five functions where human labor is acting as middleware, where the input comes in, gets processed through a set of rules, and the output goes out, and systematize those first. Customer support triage. Lead qualification. Internal reporting. Invoice processing. Onboarding workflows. Start there. Prove the model. Then expand.
If you're enterprise, a thousand-plus, this isn't a cost play for you yet. It's a capability play. The value isn't in replacing headcount. It's in building systems that operate at a speed and consistency that human teams can't match, while keeping humans on the decisions that carry real liability. Your transition is slower, more governed, and more expensive than everyone else's. But the direction is the same. Fewer humans processing information. More systems doing it. The ratio shifts. The question is whether you set the ratio or your competitors set it for you.
If you're a solo founder or just starting out, you're in the best position of anyone reading this. You have no legacy systems. No org chart to reorganize. No institutional knowledge trapped in someone else's head. You can build from zero with the architecture that everyone else is trying to retrofit. The AI stack is your team. Skills are your expertise. The lean version of the business is the only version you've ever known. Your ceiling just went up by an order of magnitude while your costs stayed flat.
The direction is the same for all of you. Strip the drag. Keep the load-bearing structure. Build systems where you used to build teams. The specifics are different for every business, every vertical, every scale. But the structural logic is identical.
What I watched go wrong
I've watched roughly thirty businesses attempt some version of this transition over the past three years. Different industries. Different scales. Different starting points. The pattern of what went wrong was almost identical every time.
They automated in the wrong order. They pointed AI at the most visible problem instead of the most impactful one. They cut headcount before the systems were ready to absorb the work. They treated the transition like a technology purchase instead of a structural redesign. They optimized the things that were easy to measure and left the things that were hard to measure untouched, which meant they streamlined the drag and left the load-bearing functions under-resourced.
One agency I worked with went straight for the content team. Eight people, biggest line item on the payroll, obvious target. They deployed an AI content pipeline, cut to three writers in six weeks, and watched output triple. By the numbers, it worked. The dashboard looked great. What the dashboard didn't show was that the original writers weren't just producing content. They were the ones who knew which accounts needed what tone, which topics had been tried and bombed, which client stakeholders had preferences that lived nowhere except in Slack threads from eighteen months ago. Three months later, two of the agency's biggest accounts flagged that the work felt generic. One of them left. Not because the AI content was bad. Because it was the same content anyone could produce. The thing that made the agency worth $12,000 a month was the human judgment about what to write and for whom. That was the load-bearing function. The writing itself was the drag. They automated the wrong layer.
The sequencing is the whole game. Which function gets systematized first matters. Which processes need to stay human matters. Where the handoff points live between agents and people matters. The right answers depend on what you're running and where. The roofing company and the biotech manufacturer and the SaaS platform all have drag, but it's in different places. The system that works for one will break the other.
The businesses that got through it treated the transition like surgery, not like shopping. Precise. Sequenced. Specific to the business on the table.
Both things are true
I keep coming back to the same tension I landed on when I first started writing about this, the same tension that sat at the center of the ClawdBot analysis. And I still can't resolve it. I don't think anyone can right now.
The opportunity is real. The structural shift is real. Every business is carrying weight it doesn't need, and for the first time the tools exist to strip that weight off and build something lean. Something aerodynamic. Something that generates the same output or better output with a fraction of the overhead. Solo founders are building businesses that compete with teams of twenty. SaaS is becoming infrastructure. Expertise is becoming transferable. Skills are becoming the atomic unit of capability. The entire economy is reorganizing around a simpler architecture: systems that run, agents that operate them, skills that encode the expertise, and humans who set direction and handle the parts that require genuine judgment and creativity.
And.
The liability framework doesn't exist. The security infrastructure is a disaster. The cost math gets complicated at scale. The top tier of expertise remains human and may always remain human. Most people haven't even started to understand what's happening. Most businesses haven't adapted at all. The regulatory environment is a patchwork of first drafts. And the gap between what's being deployed and what's been thought through, what's been secured, what's been stress-tested, is as wide as I've ever seen in any technology adoption cycle.
Both of these things are true at the same time. And holding both of them is uncomfortable and necessary.
The people who see only the opportunity will build fast and fragile. They'll move quickly, cut costs, deploy systems, and generate impressive-looking results right up until the first liability event, the first security breach, the first client who realizes the agent made a consequential mistake with no one accountable. Speed without discipline creates fragility that compounds.
The people who see only the risk will build cautiously and get outcompeted. They'll wait for perfect clarity, perfect regulations, perfect security solutions. They'll form committees and run pilots and produce internal white papers about "AI readiness" while their competitors ship, learn, and iterate. Caution without action creates irrelevance that also compounds.
The people who hold both realities, who build aggressively within the constraints of what's actually responsible, who deploy systems while accounting for liability and security and the genuine limits of what AI can do, who move fast on the things where the risk is low and move carefully on the things where the risk is real, those are the people who build the things that last. They always have been.
They've usually done this before. Not theoretically. Across real businesses, in real operations, with real money on the line. They don't experiment on your business. They install what's already been tested.
I've spent years watching one paradigm exhaust itself and five months watching the next one materialize. The hollow engine concept didn't break. It expanded to cover any business carrying weight it doesn't need, in any industry, at any scale.
Now there's an alternative to carrying it.
If you read this far, you see the direction. Strip the drag. Keep the load-bearing structure. Build systems where you used to build teams. The direction was never the hard part.
The hard part is the sequencing. Which function gets systematized first. Which ones stay human. Where the handoffs go between agents and people. What order things have to happen in so the system absorbs the work before the team shrinks. I laid out the direction for every business type in this piece. I also told you what I watched go wrong when thirty businesses tried to execute on that direction without getting the sequence right.
The pressure test isn't whether you understand the thesis. You do. The pressure test is whether you can look at your own operation, identify what's load-bearing and what's mass, and get the order right on the first pass. Because the businesses that got the order wrong didn't get a second chance to do it slowly. Their competitors had already moved.
Most businesses are still staffed for a world that stopped existing sometime in the last ninety days. That's your window.
But a window you can see isn't the same as a window you can get through.