What is AI arbitrage? The real business, and the funnel built on top of it

AI arbitrage is the practice of buying something from an AI provider cheaply, usually at a per-token price, and reselling the output as a service at a much higher price. The gap between the two is the arbitrage. It is a real business model, it is legal when done honestly, and it is also the frame around which a specific category of business-opportunity scheme has been built.

Where the margin actually comes from

Three gaps, and only one of them is durable.

The price gap is the obvious one. A language model API charges per token of input and output, which for most text work amounts to fractions of a cent per request. A client paying for a finished deliverable is paying for the deliverable, not for the compute. Charging a few hundred dollars for work that cost a few dollars of inference is the arbitrage in its simplest form.

The knowledge gap is the second. Someone who knows which model to use, how to prompt it, how to check its output and what to do when it is wrong is selling judgement, not tokens. That is an ordinary consultancy business with an unusual input cost.

The attention gap is the third and the shakiest. Many early AI businesses existed because the client did not know the tool existed. That gap closes on its own as the tools get famous, and it has been closing fast.

Is AI arbitrage legal?

Reselling AI output is legal in itself, subject to three things most explainers of this topic skip.

The provider's terms. Every major model provider sets out what you may do with outputs, whether you may resell them, what you must disclose and what categories of use are prohibited. Those terms change, and building a business on a clause you have not read is a risk you have taken without pricing it.

Disclosure. Whether you must tell a client that a deliverable was produced with AI depends on the jurisdiction, the sector and often the contract. Silence is not automatically a problem; an assurance that it was human-made when it was not is a different matter entirely.

Copyright and liability. Output you resell can be wrong, can defame someone, or can reproduce something it should not. You are the supplier in that relationship, which means the client's recourse is against you rather than against the model provider.

Why the US regulator is already involved

Because a version of this has been sold as a get-rich scheme, and the Federal Trade Commission brought cases.

On 25 September 2024 the FTC announced Operation AI Comply, a sweep of five enforcement actions against companies it said used AI to deceive or harm consumers. Three of the four settlements involved what the FTC described as business opportunity schemes claiming AI could help people earn more money, faster.

The named defendants are worth knowing, because the pattern repeats. Ascend Ecom, which the FTC said had operated under several names since 2021 and claimed proprietary AI software would maximise clients' success. Ecommerce Empire Builders, charged with falsely claiming to help consumers build an AI-powered ecommerce empire through training programmes and done-for-you storefronts. Empire Holdings Group, alleged to have falsely claimed consumers would earn significant sums from AI-driven online stores. FBA Machine, alleged to have promised guaranteed income from AI-powered storefronts.

A detail runs through several of those complaints: the FTC alleged failures to provide the disclosure documents required under its Business Opportunity Rule. That rule exists precisely because selling someone a business is different from selling them a product, and it is the legal hook that catches a great deal of this.

What the FTC also went after, and why it matters here

Two of the five actions were not business opportunities at all, and they define the edges of the category.

The FTC alleged that DoNotPay had offered an AI service it promoted as the world's first robot lawyer which did not meet its promises of substituting for human legal expertise. And it took action over Rytr, whose service could produce large volumes of product testimonials, which the FTC said had no reasonable legitimate use.

Those two cases are about the claims made for AI output rather than about reselling it. Together with the business-opportunity cases they sketch the enforcement line: you may resell AI work, you may not misrepresent what it is, what it can do, or what buying into your programme will earn somebody.

How to tell a business from a funnel

The structural tell is who the customer is.

In a real AI services business, the customer is a client buying a deliverable, and the operator's revenue comes from doing work. In a funnel, the customer is an aspiring operator buying a course, a template pack or a done-for-you storefront, and the revenue comes from selling the opportunity. The second model does not require the underlying business to work at all.

Four questions separate them quickly. Does the seller make money from clients or from students? Is there a specific, checkable claim about earnings, and is it accompanied by the disclosures the Business Opportunity Rule requires? Is the pricing for the outcome or for access? And can you see any named client of the seller's own operation rather than testimonials about the course?

What the honest version looks like

Narrower and less exciting than the pitch, and it does exist.

The durable businesses in this space sell a specific outcome to a specific industry, where the operator understands the domain well enough to check the output. Transcription and summarisation for a professional sector with compliance requirements. First-draft copy for a client who will edit it. Data extraction and structuring from documents a client already owns. Translation with a human reviewer who speaks both languages.

What those have in common is that the AI is an input cost, not the product, and the operator is accountable for the result. That is simply a services business with a cheap supplier, and the margin erodes at the normal rate for services businesses as competitors arrive.

Why the easy version is closing

The attention gap and the knowledge gap are both narrowing, and the price gap is narrowing differently.

Model prices per token have fallen repeatedly, which sounds like good news for an arbitrageur and is not. A falling input price lowers the barrier for the next competitor and for the client, who can increasingly do the simple version themselves. When a client can get a serviceable draft from a free consumer tool, the margin on selling drafts goes to zero.

The providers also keep absorbing the layer above. Features that were products a year ago become buttons inside the platform, which removes the business built on them. Anyone whose plan is to resell a thin wrapper is competing with the provider's own roadmap. We have looked at the same dynamic in the apps that get absorbed into platforms from the pricing side.

The honest summary

AI arbitrage is a real description of a real thing: there is a large gap between what inference costs and what finished work is worth. Reselling across that gap is a legitimate business when you are accountable for the output and honest about how it was made.

What the term mostly refers to online is something else: a course or a done-for-you package sold to people who want to run that business. The FTC has brought cases against several operators of exactly that model, and the Business Opportunity Rule is the instrument it used.

If you are considering either side of this, the useful question is which product you are actually being sold. One of them has clients. The other has students.

This article explains a business practice and is not legal or financial advice.