Intelligence Is Now Cheap.The Signature Is Not.

Almost everyone I talk to outside tech still thinks of AI as a chatbot that makes things up. They are right about the tool, and wrong about what it means for their business. That gap is where the money is right now, and this note is about how to find your piece of it.

Arbitrage, in plain words

Arbitrage is buying something where it is cheap and selling it where it is dear. Every business is some version of that gap. A restaurant buys ingredients and sells dinner. An agency buys hours and sells a campaign. The margin lives in the difference between what the inputs cost you and what the output is worth to someone else.

What changed this year is that intelligence itself moved to the cheap side, and businesses noticed. Ramp, the corporate card company, tracks where its customers' money goes. Its labor-substitution research shows that 58.5 percent of the companies that paid freelancers through marketplaces in 2022 had stopped entirely by early 2026. The share of company spend going to those marketplaces fell from 0.66 percent in late 2021 to 0.14 percent by late 2025. Over the same period, spend on AI model providers went from zero to 2.85 percent. The companies most exposed substituted at roughly one dollar of freelance spend for three cents of AI spend.

A first draft of a contract, a design or an analysis now costs a few dollars. That part has repriced, and it repriced fast.

What has not repriced

The person who stands behind the work.

The freelancers who lost that spend were selling drafts through a marketplace, no name attached. That is exactly the product a language model now produces for pennies. But read the objections people raise about AI and notice what they have in common. It hallucinates. It has no judgment. Somebody has to answer for the output.

Those objections are true. They are also the arbitrage, because nobody will pay a chatbot for a marketing plan and almost everybody will pay a person who used the chatbot, checked it, and put their name on it. The tool got cheap. Accountability did not, and today almost nobody outside tech is selling it in that form.

People said heavier-than-air machines could not fly, right up until one did. The internet in 1998 rewarded the early, not the smart.

Two things I watched this year

In April I watched two people with no engineering background, one of them a benefits specialist, demo a working enrollment platform they had built themselves on Replit, a tool that writes the code when you describe the app. It was not better than the incumbents in that market. It was at parity, from two people who a year earlier could not have built a login page. I have shipped software for nineteen years, and I did not have a rebuttal.

The second is a marketing consultant I now work with. He uses nothing more than Claude in the browser. No plugins, no code, no developer. With it he works out which ads bring a small business its customers, builds the reporting, and puts his name on it. On Wednesday he told me he had signed two more clients, was chasing a third, and needed more capacity. He is selling the signature, not the tool. His clients could open the same browser tab he does. They pay him because he will answer for the number.

It is moving faster than it looks

Two years ago AI video was a demo and an agent that drives a computer did not exist outside a lab. OpenAI announced Sora in February 2024 and did not release it until December. Anthropic's computer-use capability shipped as a beta in October 2024, and OpenAI's Operator followed in January 2025. Today I run a pipeline that turns a script into a finished video with an AI voice, and an agent that phones job candidates and interviews them. Neither took a team. Both would have been a funded startup's whole product the year I graduated.

One caveat

This is not legal or medical advice, and experts are not obsolete. For anything that can hurt you, AI gives you a strong first draft and a professional signs off on it. That is not a limitation of the argument. It is the argument. The signature is the expensive part.

How to find your arbitrage

Two things, both free.

First, do Claude Code 101 on Anthropic Courses. It is a short course from the company that makes the tool, and it is enough to understand how to put the tool to work rather than talk to it.

Second, take one task you paid a specialist for last month. Give the tool everything you gave the specialist: the brief, the background, the constraints, the examples. Then compare the two answers side by side. Where they match, you found your arbitrage, and you now know what a first draft in that category is worth. Where they do not match, you found what the specialist is for, and that is worth knowing too.

Run that comparison on three tasks and you will have a better map of your own cost structure than most consultants could sell you.

What do you pay your accountant, your lawyer or your agency for that is not the work itself?