America's Highest-Margin Small Businesses Are Its Most AI-Exposed

A word on scope

This analysis joins two public datasets at the industry level. Margins come from the IRS Statistics of Income — what sole proprietors actually report on their returns — ranked across sixteen sectors. AI exposure comes from the ILO's generative-AI occupational index and OpenAI's "GPTs are GPTs" study.

Both measures have limits worth stating up front. The margins describe small Schedule C filers, not two-million-dollar targets, so they are best read as a ranking rather than the exact economics of any one deal. And "AI exposure" measures task overlap — how much of an occupation's work a model can plausibly perform — not jobs eliminated. The directional conclusion is sturdy; the individual decimals are signposts.

Which small businesses report the highest margins?

Cognitive work, by a wide margin. Ranked by net margin on sole-proprietor returns, the top of the list is dominated by businesses that sell thinking rather than things, and the bottom by the businesses people most often dream of owning.

Professional and technical services, at 39.7%, are the most profitable category in the country; restaurants, at 1.9%, are dead last. This is the familiar pattern that the unglamorous business quietly out-earns the exciting one — but a second dataset turns the ranking on its head.

Does the highest-margin business also pay the owner the most?

Not always, and the exception is instructive. Margin is a percentage, not take-home dollars. Finance and insurance sits sixth on margin at 20% but first on actual dollars per owner, about $34,572 a year, because the underlying deals are larger. Professional services clears roughly $25,000 per owner; the restaurant, about $2,200.

On both the percentage and the dollar figure, the boring business wins. But neither figure prices in what a language model does to the work behind the fee.

Which of these businesses is most exposed to AI?

The highest-margin ones. On the ILO index and OpenAI's task study, accounting and bookkeeping work scores among the most exposed occupations on earth, near the top of the list alongside data entry and clerical work. By one widely cited measure, essentially 100% of an accountant's and auditor's tasks are exposed to AI. Both studies surface the same uncomfortable pattern: the higher the income and the more education a job requires, the more exposed it tends to be, not less. Roughly one in five U.S. workers could see half their tasks affected.

The least-exposed work is physical, hands-on, and licensed — the trades, caregiving, repair. No version of the current technology drives to a house, crawls under a sink, or holds a professional license. Read the two rankings together and the tension is exact: the property that makes professional services the highest-margin business in America — pure cognitive work, no trucks, no inventory, a brain and a billing rate — is precisely what a language model commoditizes. A 40% margin built that way is less a moat than a target.

What does AI do to a 40%-margin practice?

It compresses the fee by collapsing the labor behind it. A bookkeeping and tax practice earns its margin selling hours of cognitive work: categorizing transactions, reconciling accounts, drafting returns, answering whether an expense is deductible. Each of those is now a task a low-cost tool performs in seconds and does competently. The work does not vanish; the billable hours do. When the labor behind a fee collapses, the fee follows, because the client across the street has the same tool and a cheaper quote.

An HVAC company sells something a model fundamentally cannot deliver — showing up, being licensed, and being liable when the repair holds. Its 15% margin was never lush, but it is structurally protected in a way the accountant's 40% is not.

Where does durable margin actually sit?

Stacking margin against AI durability sorts every small business into one of four quadrants.

The sweet spot — durable, protected margin — is the rarest and most valuable quadrant: high margin the model cannot reach because a human body, a license, or a physical location is the bottleneck. The bullseye is where searchers and private equity are currently paying up, which places the smart-money businesses inside the blast radius.

How do two businesses diverge over five years?

Run a bookkeeping practice and an HVAC company forward from the same 3x multiple today. The practice's commodity work — the returns and reconciliations that justified the fee — gets cheaper every quarter as clients gain the same tools. Even holding the clients, the margin compresses, plausibly from 40% toward 25% over a few years, because the work can no longer be billed at the old rate. The buyer paid 3x for earnings that are quietly shrinking. The HVAC company's furnaces still break, and nothing in the model replaces the truck and the technician, so its margin holds. The business that looked worse on day one is worth more on day 1,800.

Exposure is not extinction. The research is careful on this: most of the work is transformed, not deleted, and good practices adapt by moving into advisory work a model cannot own. What erodes is the margin. The practice survives; it simply stops being a 40%-margin business, and whoever paid a premium multiple for that 40% holds the difference.

What survives both the seller and the software?

For a generation, the high-margin acquisition was to buy cognitive capacity and resell it at a markup: a practice, an agency, a consultancy — low capital, high margin, scalable. When the cognitive labor itself becomes cheap, the durable question shifts from "what is the margin" to "can a model do this." Three checks follow before paying a premium for a high-margin business: where the margin comes from (cognitive labor billed by the hour is exposed; a physical service, location, or license is protected); whether a competitor could buy the same capability cheaply off the shelf (if so, the moat is already gone and only the price lags); and what survives both the seller and the software. On the current evidence, margin built on selling thinking is the fragile kind, and margin anchored to a body, a license, or a place is the durable kind.

Sources & method

• IRS Statistics of Income — sole-proprietor net margins by sector; describes small filers and is used as a ranking, not as the exact economics of any one deal.
• ILO generative-AI occupational-exposure index — task-overlap measure of how much of an occupation's work generative AI can perform.
• OpenAI, "GPTs are GPTs" — occupational exposure to large language models.

Margins are what sole proprietors report to the IRS. AI "exposure" measures task overlap, not jobs eliminated.