Tax the tokens and you miss the money

On September 1, Fortune reported that three House members had introduced a bill to tax AI tokens and spend the proceeds on job creation. Representatives Sara Jacobs, Greg Casar, and Valerie Foushee wrote H.R. 10044 to take either 2% of the value of the tokens an AI model processes or 3% of revenue from AI services, whichever collects more, with the rate climbing automatically as unemployment climbs. The money would go to housing construction, infrastructure, child care, and elder care. Jacobs said AI could produce “the biggest wealth transfer in history from the bottom to the top.”

She is right about the direction. The bill is pointed at the wrong end of the pipe.

Run the arithmetic on one support email

Token prices are published. A frontier model runs about $1.25 per million tokens of input and $10 per million tokens of output. Now picture a real support exchange: the customer’s message, the order record, four policy documents pulled from a knowledge base, a couple of tool calls, and a written reply. Call it 30,000 tokens in and 2,000 out. That is about five cents. Round it up to a dime to cover retries, a second model checking the first, and everything else a careful setup does.

A dime.

The person who used to answer that email cost somewhere between $25 and $35 an hour once you include payroll tax, benefits, a seat, and a supervisor. In that hour they might have closed five or six tickets.

So the bill’s 2% take on the tokens for those six emails comes to roughly a tenth of a cent. The thirty dollars an hour did not evaporate. It went somewhere.

It went one layer down, to the company that stopped paying it

Last September, Marc Benioff said on a podcast that Salesforce had taken its customer support organization from 9,000 people to 5,000, with AI agents absorbing the difference. Salesforce called it a rebalance and moved many of those people into other roles, which is more than most companies do. Set the labels aside and look at the cash. Four thousand salaries stopped being spent on support. The compute that took over the work cost a small fraction of that. The gap stayed on Salesforce’s income statement.

That is the general shape, not a Salesforce quirk. Model providers sell inference into a brutally competitive market with thin margins and falling prices. The company deploying the model is the one that keeps the wage.

This is not a small pool. Forrester analysts Kate Leggett and Laura Ramos projected in May that 49% of current customer service jobs will be gone by 2030. The Bureau of Labor Statistics has customer service representative employment falling about 5% between 2024 and 2034, and that projection is already looking gentle.

A tax on tokens reads the smallest number in the whole transaction. It is like taxing the electricity a factory uses instead of what the factory sells.

The other proposals aim high too

Bernie Sanders’ American AI Sovereign Wealth Fund Act would take a one-time 50% stake in the stock of large AI companies and pay every American roughly $1,000 a year from the returns. It is far bigger than the token tax, and it has a real advantage: paying everyone means nobody has to prove they were the one replaced, which is a problem we have written about before. About 69% of Americans back the idea.

But it still bills the builders. A regional bank that closes a 200 seat contact center and keeps the payroll owes nothing under either bill, because it does not sell AI. It only bought some.

We are the layer the money passes through

We sell customer service automation. When a business connects Celeste to its email and live chat, work that a person used to do stops being done by a person, and we book revenue for that. Ours is exactly the transaction a well aimed rule should be able to see.

So we wrote our own before anyone made us. The Dividend Standard commits the greater of 5% of our qualifying revenue or 70% of our adjusted profit, every year, to the workers our software displaces, granted as an ownership stake rather than a one-time check. The revenue floor is doing the important work there, because a young company can report no profit for years while cheerfully removing jobs the whole time.

We cannot make anyone whole by ourselves. We see a slice of the savings, not the savings. That is the point of publishing the formula instead of quietly writing a check: it is meant to be copied, and it is meant to be argued with.

If Congress wants a meter that finds the money, put it where the payroll used to be. That is a line on a real company’s books, and it is a much larger number than a dime.