The last piece argued that a firm's productivity gain goes to whoever has the pricing power to keep it — capital, workers, or customers — and that an hourly-billing firm converts the gain into a revenue loss because it sells time rather than output.1
Drop one level. A salary is hourly billing.
Take two senior developers at $200,000. One produces what a good senior developer has always produced. The other supervises agents and produces the output of three, at the same quality, for the same $200,000. The second developer's economics are fundamentally different — for the employer.
For the developer, nothing changed. Same pay, three times the output, and the surplus went to the firm.
That is the same division problem, one level down. At the firm level, the dividend is retained by pricing power. At the individual level, pricing power means one of three things: scarcity, an outside option, or ownership.
A worker with none of them has just cut their own price, and the labor share number from the second quarter — 52.9%, the lowest since 1947 — is what that looks like in aggregate.2
What the company was selling
The historical advantage of working for a company was leverage. It gave you capital, colleagues, distribution, administrative support, research capacity, and specialists. No individual could assemble that alone, so individuals rented access to it in exchange for a salary and a share of the risk.
AI is unbundling that leverage from the corporation and handing pieces of it to the individual. Software development, design, research, customer support, marketing, financial modeling, first-draft legal work, documentation, and operations are all things a small team can now perform or amplify with agents.
The constraint moves from how many people you can hire to how well you can specify, orchestrate, verify, and sell.
The minimum viable company gets smaller. A business that took $1–2 million and 10–20 people is becoming a business that takes two or three people and $100–300K. Some software businesses already cost less than that. And a founder can run several experiments at once, kill the weak ones cheaply, and concentrate on whatever gets traction — which used to require a portfolio, not a person.
Note
I have started companies without outside capital since 1999. I also have a job. The difference now is that the capacity I can bring to the first has stopped being bounded by the hours left over from the second.
Building gets easier as building gets worthless
If you can produce a competent SaaS product with six agents, so can everyone else with six agents. Code, content, analysis, and product design are commodities in the making. Durable value migrates to what the agents don't give you: distribution, proprietary data, customer relationships, reputation, domain depth, regulatory access, network effects, and capital.
That is the firm-level moat list from the last piece, and it applies without modification to a person.
It also relocates the risk. The standard story is that AI reduces entrepreneurial execution risk, and it does — several stages of idea, raise, hire, manage, build, market, iterate collapse into one person and an afternoon. But total risk doesn't fall. It moves from building to selling. When anyone can build, every channel gets noisier, and acquisition gets more expensive, and the founder's dividend goes to whoever owns the channel: the app store, the ad network, the marketplace.
A founder whose thesis is "I can build software incredibly fast now" has a thesis about the part that stopped mattering. A founder who understands a market deeply has a thesis about the part that still does.
Which skills compound
There is a claim doing a lot of work in this argument: that using AI intensively makes you better at using AI, so that experience compounds rather than everyone getting the same lift on the same day.
Half of it is wrong. Technique learned on a weaker model depreciates when a stronger one ships. Prompt rituals, workarounds for failure modes that no longer exist, decomposition that the model now handles on its own — all of that is dead weight within a release or two.
The other half holds. Judgment about where to point the tool compounds:
- what to delegate
- what to verify
- where the models still fail
- how to break a problem so the pieces are checkable
Two people with identical Bloomberg terminals do not produce identical returns. Two people with identical IDEs do not ship identical software. Access to machinery has never equalized output, and there is no reason to think this machinery is the exception.
Important
Someone with three years of orchestrating agents is not competing with someone who subscribed yesterday. But the advantage lives in judgment, not in technique, and anyone who mistakes the one for the other will find their edge evaporating on the next model release.
The middle
The research on who AI helps most looks, at first, like it contradicts all of this. In customer support, generative AI raised productivity 14% on average and 34% for the least experienced agents.3 On professional writing tasks it narrowed the gap between weak and strong writers.4 Among consultants, below-average performers improved more than above-average ones.5
Compression, not amplification.
Both findings are right, and they describe different kinds of work. On well-specified tasks with a known correct output, AI compresses the distribution upward: the novice gets the expert's answer. On open-ended work with no fixed answer, AI amplifies: the person with judgment gets more of what they already had.
That is exactly why the middle is exposed. The most expensive people do open-ended work and get multiplied. The cheapest standardized work gets automated outright. The reasonably competent $150–250K professional sits in the band whose work is specified enough to be automated by someone senior with agents, but not senior enough to be doing the orchestrating. Their historical value was producing information, code, analysis, or coordination that a more capable person can now generate directly.
Companies don't fire that band. They stop backfilling it. Junior hiring slows, functions consolidate, and the remaining people are asked for more. A department discovers its best twenty people plus agents can do what fifty did, and the other thirty leave by attrition over two years. Worker-level adoption is already at 41% for work-related generative AI, and 78% of the labor force works at a firm that has adopted it.6
This isn't a forecast about a technology that hasn't arrived.
The answer
So: start a business or stay employed?
The question is wrong as posed, because it assumes a binary decision on day one. AI has narrowed the difference between the two paths — lower execution risk on one side, higher variance on the other — without making entrepreneurship safer than employment outright. Customer acquisition, product-market fit, and competition are as brutal as they were.
What changed is that the salary no longer has to be the price of not building.
The defining behavior of the next decade is going to be capable people keeping the job and building something substantial beside it. Not a hobby. A business that would have required quitting and hiring five people in 2019, run in the evenings by one person and a stack of agents in 2026. The salary funds the runway. The benefits cover the variance. The asset accumulates evidence.
And then the question reverses. It stops being "should I quit and gamble on this?" and becomes "why am I still selling my time at a fixed price when this asset has demonstrated traction?" That is the moment a worker acquires pricing power — not by asking for it, but by owning something the employer doesn't.
Warning
Three things get people hurt on this path, and none of them are in the inspirational version. Read the IP assignment clause in your employment agreement before you write a line of code; many assign anything you build on company equipment, on company time, or in the company's field. Keep the business on separate hardware, separate accounts, and a separate legal entity from the first day, not the day it starts making money. And don't build in your employer's domain, because the strongest founder advantage — knowing a market deeply — is also the one most likely to be contractually theirs.
The principle underneath all of it: AI is lowering the cost of owning productive capacity while raising the expected output of anyone who sells their labor. Ownership gets more attractive. Undifferentiated employment gets more vulnerable. For someone who already knows how to build products and run a business, the constraint that always capped a small founder — available human bandwidth — is precisely the constraint the agents are attacking.
What that does to the firm, and to the interview, is the arc I wrote last year.7 This is the economics of why it's happening.