Issue 009 - The Melting Moat

For most of my career, I bought services by the hour. Agencies, consultancies, integrators, staffing firms. The proposals changed from vendor to vendor. The invoice didn’t. It primarily counted hours, and behind every hour was a person on somebody’s payroll.

That is the quiet fact underneath a large part of the US economy: much of professional work is priced in person-hours. When the price is hours and the cost is salaries, the size of the business is the size of its roster. Capacity is headcount. Growth is hiring. Margin is how full each person’s calendar is.

It also means the moat is the payroll. What a 200-person agency has that a talented person working alone does not is capacity: enough people to staff the account, survive a resignation and pass procurement hurdles.

I call what is happening the Melting Moat.

When software can do a meaningful share of the billable work, the hours stop requiring the headcount. The roster stops being the advantage. And one capable person, working with AI, can sell what used to take a team.

That is the core engine fueling my prediction of The Artificial Entrepreneur: a cohort of founders who build companies that grow without hiring the way their growth would predict.

This issue lays out the idea, the evidence I have for it so far, the condition under which I said I would abandon it, and the first time that condition was met by federal data. It came back on September 29. The thesis survived. Here is exactly what that does and does not mean.

The buyers have noticed. In a 2024 survey of more than 80 multinational advertisers by the World Federation of Advertisers and MediaSense, three in four said they intend to change how they pay their agencies within three years, and 58% want more of it tied to outputs or outcomes rather than time.¹ Only about a quarter named AI as the reason. Most named alignment: paying for results instead of effort.

That distinction matters. The pressure on hourly pricing did not start with AI. What AI changes is the cost of the hour. Suppose a task that took a junior analyst six hours now takes 20 minutes with a large language model. A firm that bills by the hour has two bad choices: bill the 20 minutes and shrink, or keep billing six hours and hope the client never notices. Clients notice.

The same survey caught a paradox. Most of those brands expect their agency fees to rise over the next three years, and most also expect to pay less as AI spreads.¹ It resolves once you stop treating the work as one thing. Buyers expect to pay more for judgment, for strategy and for the person who can read the data and make the call. They expect to pay less for everything that can be produced. The billable middle is what melts. Not the top. Not the bottom.

On September 10, before I had looked at any outcome data, I wrote down which industries I expected to be hit hardest and by how much. Then I checked.

I got some of it wrong, and the wrong part taught me the most.

I expected the line to run between kinds of professions, and I expected analytical professions such as law and accounting to lose ground. They didn’t. Employment in legal services and in accounting both grew. Of the five industries I ranked, two landed where I said they would.

What the data showed instead is that the line runs through the content of the work. Inside the business-services sector, the industries that sell contact and help with paperwork (contact centers, collections, document preparation, temporary office help) lost jobs between 2019 and 2025. The industries next door that sell physical work at similar wages (pest control, landscaping, building services) mostly gained them. The median contact-and-paperwork industry fell 14%. The median physical-work industry grew 13%.²

An ordinary downturn cuts hourly work broadly. It does not grow pest control by 22% while contact-center employment falls by more than a third. Something is selecting on what the work is, not what it costs.

Three things keep me from overselling that.

The decline in call centers is older than AI. Telemarketing employment peaked in 2016 and has fallen every year since, for reasons that include offshoring and robocall rules. What changed is the speed: about 1% a year from 2015 to 2019, about 10% a year from 2022 to 2025.² The acceleration is the finding. The decline is not.

The evidence isn’t uniform. A catch-all category of other business support services grew 18%. I don’t have an explanation for that yet, and I won’t pretend I do.

Court reporting looks like a second exception, and I think it is the most instructive thing in the chart. Turning speech into text is one of the most automatable tasks on this list, yet private court-reporting employment fell only 6%. My explanation: freedom and lives are at stake. A court reporter certifies the record of a trial. If software mangles the transcript of a murder trial, the cost of that error is a person’s liberty, or a just verdict being thrown out on a technicality. Until AI is far less fallible than a human reporter, nobody is going to hand it that job.

What protects work, in other words, may not be whether AI can do it. It may be whether a named person has to vouch for the result and answer for it when it’s wrong. Regulation alone doesn’t do it. Collection agencies are heavily regulated, and they lost 31%. The difference is that a collector attests to nothing. A court reporter, a lawyer signing a filing and an auditor signing an opinion all do.² It’s a hypothesis, not a finding, and it is a small industry. But it is a question the data can answer, which brings me to law and accounting.

2019 to 2025 runs straight through a pandemic. Both groups lived through it. Only one of them shrank.

Law and accounting look like exceptions to everything above. Both added jobs from 2019 to 2025: legal services grew 7%, and accounting, tax preparation, bookkeeping and payroll grew 12%. But an industry total only tells you how many people a firm employs. It doesn’t tell you what those people do. So I looked inside, at the occupations.⁵

In law firms, the people doing the legal work grew. Lawyers rose from about 403,000 to 448,000, up 11%. Paralegals rose from about 255,000 to 305,000, up 20%, faster than the lawyers they work alongside. The clerical layer shrank. Legal secretaries fell 10%. General office clerks fell 23%. Office and administrative support as a whole fell 12%, about 43,500 fewer people, in an industry that was growing.

Accounting firms show the same split. Bookkeeping and auditing clerks fell 10%. General office clerks fell 19%. Office and administrative support as a whole fell 10%, about 31,000 fewer people. Meanwhile the knowledge work got more people: management, business and financial, and computer jobs together grew 17%, about 104,000 more. Accountants and auditors themselves barely moved, up 1%. The growth came in the roles around them: financial managers up 42%, management analysts up 39%, tax preparers up 12%, computer and technology staff up 33%.

Put the two industries side by side and the pattern is plain. The clerical work got done with fewer people. The knowledge work was resourced with more people. Together, law and accounting employ about 75,000 fewer people in office and administrative support than they did in 2019, while adding about 267,000 in everything else.

That is why the industry totals barely registered it. Staffing firms and call centers are almost entirely clerical work, so when that work melts, the whole industry shows it. In law and accounting, the clerical layer is the smaller part of the firm, and hiring everywhere else covers the loss. Same mechanism, different visibility.

It also answers the question I raised about whether attesting to something is the key to job durability. It seems unlikely as paralegals sign nothing, and they grew fastest of all. So the line isn’t who signs the work. It’s whether the work is a procedure or a judgment. Procedures are getting done with fewer people. Judgment is getting more of them.

One caution: BLS warns that these occupation estimates are not built for tracking change over time, so I compare two years and make no claim about exactly when the shift happened.

A forecast you can’t lose isn’t a forecast. So alongside the Melting Moat, I wrote down the condition under which I would drop it entirely.

The logic was simple. When the moat melts, firms need fewer people. Some of the people they no longer need go out on their own, and so do some who see how much a solopreneur can now achieve. In federal data, that shows up first in one place: nonemployer businesses, the Census Bureau’s count of businesses with no paid employees, most of them people working for themselves.³

That count has fallen only once since 1997, in 2008. So the condition was this: if the number of US nonemployer businesses did not rise from 2023 to 2024, the moat was not melting into new businesses, and I was wrong. I would say so here.

I chose that test because it was cheap to fail and impossible to fudge. One number, published by Census, on a date I don’t control. When I wrote it down, the 2024 number did not exist yet.

Census released the 2024 count on September 29.³

The condition didn’t trip. The thesis survives its first scored test.

It is the largest one-year increase in number in the Census series back to 1997. It is not the largest in percent; 2003 grew faster, at 5.7%. The two years it passed, 2021 and 2022, each added about 1.33 million during the pandemic surge I have called The Airborne Entrepreneur.

Now the part that matters more: this is the weakest kind of win a forecaster can claim. A kill switch that doesn’t fire tells you the idea isn’t dead. It doesn’t tell you the idea is right, and it says nothing yet about AI.

I chose 2024 as my starting line precisely because I expected AI to have had minimal influence on it: some early adopters, some early job losses, not enough to move a national count. If 2024 is a record, it is a record built mostly by something else. That is the point of a baseline.

The kill switch was the floor. The prediction is the bar I have committed to clearing.

On September 11, 2026, I registered this: the United States will add five million net new nonemployer businesses between the 2024 count and the 2028 count. With 2024 now published, the 2028 count must reach at least 36,985,874. Census should publish it around 2030.

That is not a trend line. From 2012 to 2023, the count grew by an average of about 700,000 a year, roughly 2.8 million over four years.³ Five million needs 1.25 million a year, four years running.

A strong starting year makes this prediction look closer to trend than it is, and it makes the prediction harder to win honestly, not easier. So the size of the number is not the whole test. Where the growth lands is.

This is where the Melting Moat earns its place. If it is real, more and more of the new businesses should be people selling professional work on their own (consultants, analysts, designers, developers, marketers), rather than more delivery drivers and social media influencers. So I am tracking the share of growth that lands in professional, scientific and technical services and in information, excluding film and music production.

In 2024, that share was about 12% of the net increase.⁴ For comparison, the couriers-and-messengers industry, where many app-based delivery drivers file, added about as many businesses as all of professional services combined: roughly 157,000 each.³

That baseline doesn’t flatter my thesis. Good. It means the test can fail. If the five million arrives and it is mostly delivery and social media, I was wrong, and I will say so in this space.

Five million counts only businesses with no employees, and that is not the only place this cohort will show up. Some of it will form as small employers that produce more per person than any companies before them. I’m watching both.

If you sell to small business: the firms that grew added professionals and cut clerks. A product priced per back-office seat is selling into a shrinking seat count. A product that makes the lawyer, the paralegal or the accountant faster is selling into a growing one.

If you advise owners: a firm whose value rests on how many people it can bill by the hour is standing on the moat that is melting. Ask what the business is worth if the same work takes half the staff, because a buyer will.

If you buy companies: revenue per employee and the share of headcount in clerical roles will tell you more than industry growth. A roll-up thesis built on adding headcount is buying the thing that is melting.

If you write about this: a growing industry total can hide the shift. Law firm employment rose 7% while office and administrative support inside those firms fell 12%. Report the composition, not just the total.

In Issue 004, I showed you a peer-reviewed study that found a different mechanism behind The Accidental Entrepreneur’s outperformance than the one I had assumed in 2009. I was wrong about why. It didn’t matter. I was right about the cohort, and the companies that served it won.

This is how I work now. Name the cohort. Write down what would show it isn’t forming. Put a number and a date on it. Score it in public, including the parts that go against me.

This is the first prediction I’ve scored in public. It won’t be the last.