An AI entrepreneur is two different jobs, not one
A definition of "AI entrepreneur", the two distinct senses the word carries, and four tests for deciding whether a business actually qualifies.
The definition
- AI entrepreneur
- An AI entrepreneur is someone who starts and runs a business in which AI systems do a measurable share of the operating work, not just someone who sells an AI product.
The load-bearing word is operating. Not advising. Not prototyping. Work that, if it stopped on a Tuesday morning, would show up in a revenue number or a support queue by Thursday.
Metadata.io went from $0 to $15M ARR on $50M raised, with six patents in AI-driven marketing. It sold AI to marketers and ran on people. Before that, Silver Spotfire, $50K to $1.5M ARR — same pattern. I am now publicly building a zero-human company, where the agents do the work and I pay their bills. Same founder, two different jobs. That gap is the whole entry.
Two senses, labelled
The phrase gets used for two populations who have almost nothing in common operationally. Name them before arguing about them.
| Sense A — AI-sector founder | Sense B — AI-operated founder | |
|---|---|---|
| What's AI | The product | The staff |
| Sells to | Customers who want AI | Customers who want an outcome |
| Headcount | Normal for the ARR | Deliberately flat |
| Main cost line | Salaries | Salaries plus inference |
| Typical failure | Model is a feature, not a moat | Unbounded loops and silent wrong output |
| How you verify | Look at the product | Look at the bill and the trigger |
Most press coverage means Sense A. Most people typing the query in 2025 mean Sense B — they want to know what it looks like to run a company where software does the shifts. A founder can be both. The two senses impose different engineering, and conflating them is how people end up pitching a wrapper as a workforce. I wrote about the role split in more detail in Who is "the AI founder"?.
Where the word came from
Entrepreneur predates all of this by roughly three centuries. Richard Cantillon used it in his Essai sur la nature du commerce en général, published posthumously in 1755, to mean the person who buys at a known price and sells at an unknown one — the bearer of price risk. That definition still does the work. The AI entrepreneur buys inference at a published per-token price and sells an outcome at a price set before the token count is known.
Artificial intelligence was named in August 1955, in the proposal McCarthy, Minsky, Rochester and Shannon wrote for the Dartmouth summer project. Four men, one document, a name that stuck — I covered that in Nobody founded AI.
The compound is recent, and its usefulness is datable. Before 30 November 2022, "AI entrepreneur" almost always meant Sense A, because you could not buy general-purpose reasoning by the call. What made Sense B mechanically possible was tool use: OpenAI shipped function calling in the API on 13 June 2023, so a model could decide to hit an endpoint instead of only emitting prose. That is the point at which a model stops being a feature and starts being a worker with hands.
What it is not
It is not an "AI-enabled founder". That is the neighbouring term, and it is the one people substitute when they want the label without the bill. An AI-enabled founder uses ChatGPT to draft positioning, Cursor to write the first version of a feature, and Midjourney for the deck. Useful. Not this. The AI is a tool they pick up and put down, inside a company whose work is still done by humans on a payroll.
The boundary is not sophistication. It is autonomy and cost. Does something run when nobody is awake, and does it generate a bill?
One worked example
A webhook fires. An agent enriches the inbound lead and routes it. No person started that.
The loop is capped at a fixed number of steps, enforced in your own code rather than in the prompt, because a model asked to stop will sometimes decide not to. Anthropic's own agent guidance is explicit that these systems run in a loop and need a stopping condition. Unbounded loops are the most common production failure in agent systems, and the fix is always outside the model: a step counter, a wall-clock timeout, a spend ceiling on the API key.
The unit economics are arithmetic. Measure tokens in and out for one run, multiply by the published per-million price on the provider's pricing page, add the enrichment vendor's per-lookup fee. That is your cost of goods for one task. Then charge for it — Stripe meter events exist for exactly this, and I walked through a working build in Charge for one agent task.
A trigger you did not pull, a bounded loop, a known cost per run, a price above it. That is the whole qualifying structure. Start with one, bounded — that argument is here.
Four tests
Apply these to any business claiming the label. Three out of four is Sense B. Fewer than two is an AI-enabled founder with good marketing.
- The trigger test. Does the work start without a human starting it? A cron job, a webhook, a queue — not someone opening a tab.
- The bill test. Is there a line item for inference that you would notice if it doubled? No bill means no production.
- The bound test. Is there a step cap, a timeout and a spend ceiling enforced in code outside the model? If the only limit is a prompt saying "be efficient", the system is a demo.
- The revenue test. Is the agent's output something a customer pays for, directly or through a cost you no longer carry? Internal time saved does not count unless you can name the task that stopped.
The honest version of the label admits what it has not done yet. I would not give an agent unsupervised spend authority today, and anyone who tells you they have is either lying or has not read their logs. Autonomy is a dial, and the number on the dial is the definition.
Sources
- A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (AI Magazine reprint) — The August 1955 proposal in which McCarthy, Minsky, Rochester and Shannon put the phrase 'artificial intelligence' in writing.
- Function calling and other API updates — OpenAI — 13 June 2023: the date tool-calling became a standard API feature, which is what made model-run operating work practical.
- Introducing ChatGPT — OpenAI — 30 November 2022 launch date, used to mark when the term entered mass circulation.
- Building effective agents — Anthropic — Vendor engineering guidance that agents run in a loop and need explicit stopping conditions and guardrails.
- OpenAI API pricing — Published per-million-token prices, the input to any per-run unit cost calculation.
- Usage-based billing — Stripe Docs — Meter events and usage-based pricing, the billing primitive for charging per agent task.
- Richard Cantillon — Britannica — Cantillon's Essai sur la nature du commerce en général, published posthumously in 1755, as the source of 'entrepreneur' as a bearer of price risk.