Last time, I asked a simple question. If AI agents don’t pay withholding, what do you tax instead?
A lot of smart people have landed on the same answer. Tax the tokens. Tokens are how AI companies measure and bill the work their models do, so every time an agent answers a question or drafts a contract, the meter runs. The idea is to put a small tax on that meter and use the money to help workers.
It sounds simple. I don’t think it works. Tokens measure how much a computer runs. They can’t tell you whether anyone lost a job, and they aren’t proportional to how much it costs to care for that human or community harmed by the job loss.
The token tax isn’t a fringe idea anymore. A group of House lawmakers introduced a bill in August that would tax AI companies on token sales or AI revenue, with rates that rise as unemployment rises. DuckDuckGo’s CEO says his company would pay a 10% surcharge, roughly what employers pay in payroll taxes. Bill Gates argues the current system rewards companies for replacing people with software. Even Anthropic, the company behind Claude, has floated a 3% levy on model usage.
But when billionaires, AI companies, and Congress all agree on a token tax, it’s worth asking who it actually helps.
Your Agent Meets Their Agent
Picture this. You use an agent to book a hotel block for a conference. The hotel has its own agent. The two of them go back and forth on dates, rates, and cancellation terms. Thousands of tokens get burned on both sides, and under a token tax, both sides pay.
Nobody lost a job in that exchange. The front desk is still staffed, and you still have yours.
Now picture a company across town that laid off forty people in claims processing. It runs a cheap, efficient model that does the same work in a fraction of the tokens. Its token tax bill is small, and the state still lost forty paychecks. Those forty people may now need help from the state, like Medicaid or job training, and nothing on that token bill reflects the cost.
Why The Meter Is Broken
Worse, AI keeps getting cheaper, and a tax on usage brings in the most money when AI is expensive and the least when it’s cheap and replacing the most workers. A state budget needs the opposite.
It’s also easy to game. Even an Oxford paper that supports token taxes admits providers could misreport usage or hide tokens. The companies being taxed would be the ones doing the counting.
It may push work offshore, too. Critics say a token tax mostly makes foreign AI models more attractive.
And states can’t agree on how to tax agents today. If I live in North Carolina and book a hotel room using an agent that was built in California, and that process is routed through servers in Kentucky and Montana, who benefits from that transaction? Depending on how an agent is built and priced, it can be taxable in 22 states or only 4. That’s the current mess, before anyone adds a new tax on top of it.
So Where Does That Leave Us?
Tokens are the easy thing to count. But they can’t tell you who lost a job, or what that loss costs a community.
The tax code already treats people and software differently. Employers pay payroll taxes on wages. Software gets written off as a business expense. That gap rewards companies for replacing people, which is how we got here.
A better fix starts with how we tax labor and capital. But any fix runs into the same question: did this agent replace a person, or did it help a human accomplish more? A token meter can’t answer that.
Someone has to look at the work and money and make a judgment call. In most companies, a version of that call belonged to HR. That’s why I said HR sits at the intersection of work, power, politics, and money.
But that dynamic has completely shifted. If HR keeps shrinking inside companies, where does that skill go? That’s the next piece.