New Hampshire Taxes Payroll. AI Is Coming for the Payroll.

Independent New Hampshire Journalism · Northfield, NH

The state’s payroll count is flat across a year in which residents kept finding work. Its signature business tax runs on those wages, and nobody in Concord is measuring the gap.

New Hampshire collects a tax almost nobody outside a Concord accounting office thinks about, and it is the one most exposed to artificial intelligence. The Business Enterprise Tax does not tax profit. It taxes what a company pays out: compensation, interest, and dividends, at 0.55 percent, whether the business earned a dollar that year or not. In Tax Year 2023 that base was roughly $45.9 billion, in a state whose entire economy measured $114.4 billion.

Take out the interest and the dividends and what is left is wages. The New Hampshire Fiscal Policy Institute published the most detailed public analysis of the tax in January 2026, describing it as “fundamentally a tax on compensation paid by businesses” and reporting that no other state funds itself this way.

Here is the arithmetic nobody in Concord has done out loud. Every job a New Hampshire employer converts from a salary line into a software subscription shrinks the base of one of the two taxes that carry the state budget. That is not a prediction about robots. That is how the statute is written.

Two groups are making the same mistake from opposite ends. Workers who are certain a machine cannot do what they do are guessing. Executives who are certain they can subtract the payroll and keep the revenue are guessing too. New Hampshire is the third party to that bet, and the state never placed it.

What the evidence shows, and what it doesn’t

Start with what is measurable. Challenger, Gray & Christmas tracks announced job cuts and the reasons employers give for them. Through June 2026, American employers attributed 101,743 cuts to artificial intelligence, roughly 23 percent of all cuts announced in 2026 and the leading stated reason for four straight months. The firm has logged 173,568 AI-attributed cuts since it created the category in 2023. Technology alone accounted for 139,156 cuts in the first half of 2026, up 83 percent over the same stretch of 2025. Andy Challenger, the firm’s chief revenue officer, says AI “continues to reshape how companies think about headcount.”

That is a survey of what executives say, not a measurement of what happened. Daniel Zhao, chief economist at Glassdoor, cautions that a company citing AI for a layoff is not proof AI caused it. Fabian Stephany of the Oxford Internet Institute makes the same point: some firms use the technology as cover for ordinary cost-cutting. Blaming a machine is easier than admitting you overhired.

The hard research is narrower. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at the Stanford Digital Economy Lab went through ADP payroll records for millions of workers. Controlling for firm-wide shocks, they found entry-level hiring in AI-exposed jobs fell 13 percent relative to less-exposed jobs at the same companies, with the effect appearing only after large language models spread. Older workers showed no statistically significant hit.

Chandar is blunt about the ceiling on his own finding. “The overall impact of AI on aggregate employment is likely small right now,” he wrote in his survey of the evidence. A Danish study by Anders Humlum and Emilie Vestergaard found no hiring gap at all between adopting firms and the rest. Work out of the United Kingdom found the opposite. The researchers doing this for a living disagree in print, and say so.

Anyone telling Granite Staters this question is settled, in either direction, is selling something.

The mistake on the other side of the table

I once made a cruder version of this argument to a friend. It has a formal literature behind it. The version people mock is simple: if you strip a hundred thousand dollars a year out of a household and hand it to shareholders, you have removed a customer. Do it a million times and you have removed the market.

Xupeng Chen, a researcher at New York University, put that intuition into a model in March 2026. His paper describes a displacement spiral, in which each firm’s rational decision to swap software for staff reduces total wage income, which reduces demand, which squeezes margins, which pushes the next firm to automate faster. He pairs it with what he calls Ghost GDP: output keeps rising while the money circulating through wage earners falls, so the headline numbers look fine until they don’t. The load-bearing fact is concentration. The top fifth of American earners drive between 47 and 65 percent of all consumer spending, and those are the same households sitting in the most AI-exposed occupations.

Now weigh the source. Chen’s paper is an unreviewed preprint, and he sits in an engineering department, not an economics faculty. His own simulations undercut the doom reading: the median run is benign, with about 14 percent producing a crisis-scale shortfall. He says plainly that “this paper is a stress test, not a forecast.”

The heavier source cuts the other way. Daron Acemoglu of MIT, who won the Nobel in economics in 2024, modeled the same question and found the macroeconomic effect modest: no more than a 0.66 percent gain in total factor productivity over a decade, and likely less. Acemoglu is peer reviewed, published in Economic Policy, and working in his own field. On the size of the shock, he is the stronger authority, and his answer is that the shock is small.

What the two agree on is the part that should worry a state legislature. Acemoglu finds no evidence AI will narrow the gap between labor income and capital income, and predicts it will widen. Chen builds his whole model on that widening. Even the optimistic reading says the money moves from wages toward owners. In most states that is an inequality story. In New Hampshire it is also a revenue story.

Why the arithmetic bites harder here

Between 2008 and 2015, New Hampshire’s BET receipts tracked total private-sector wages paid in the state closely enough that the Fiscal Policy Institute uses the relationship as its baseline for estimating what rate cuts have cost. Wages up, BET up. The correlation is the point of the tax, and the vulnerability. Software licenses do not appear in the enterprise value tax base. Salaries do.

Automation converts a taxable compensation line into an untaxed vendor expense, while whatever margin it produces lands in profit. Some of that reappears through the Business Profits Tax. Not all of it, and not reliably. The BPT swings with national corporate earnings, the more volatile of the pair, which is why the BET was invented in 1993 to steady the base. Trading the stable tax for the unstable one is a downgrade even when the dollars come out even.

New Hampshire is running this experiment with no cushion. Costs outpaced revenues badly enough this budget cycle that lawmakers pulled $67.3 million from the Rainy Day Fund. The state has no income tax and no sales tax. When the business taxes soften, the pressure lands on local property taxpayers, the machinery I traced through Concord’s school bonding fight in April 2026.

To be exact about the limits: no one has modeled it. Not the Department of Revenue Administration, not the Fiscal Policy Institute, not Employment Security. The direction is clear from the statute. The magnitude is unknown. That gap is the story.

The absence you can’t see

Chen’s paper makes one observation worth pinning to a Concord bulletin board. Displacement shows up three ways: layoffs, wage compression, and jobs that simply never get posted. The third arrives first and is hardest to spot, registering as an absence rather than an event. He calls the result an observation gap: the trend runs well ahead of the indicators meant to catch it.

Now look at New Hampshire’s numbers. Through the spring, the state’s two surveys pulled apart. In May, Employment Security put the seasonally adjusted unemployment rate at 3.0 percent, down two-tenths on the year, and employed residents rose to 753,300. Nonfarm payrolls came in at 700,000, down 3,200 from May 2025. March ran 5,800 below the prior year; February, 4,200 below. Then June snapped back. The July 21 federal jobs release showed New Hampshire adding 4,500 payroll jobs in a month, a 0.6 percent jump, the largest percentage gain in the country and one of only three statistically significant increases. The bounce pulled the year-over-year change to what the Bureau of Labor Statistics calls essentially unchanged.

Flat, in other words. More residents working; a payroll count ending a year of erosion where it started; construction shedding jobs faster than any state in the nation, down 2.5 percent on the year in an industry group’s read of the federal data. I am not going to tell you AI did that, because I cannot prove it and neither can anyone else. Payroll and household surveys diverge for ordinary reasons, including self-employment and contract work, and one strong month is one strong month. But a compensation base that has stopped growing is the exact shape researchers say to watch for, in a state that funds itself off that very line item, and the state is not looking.

What Concord did about it

It had the chance. Senate Bill 657, filed in January by Senators Rebecca Perkins Kwoka, Cindy Rosenwald, Suzanne Prentiss, Patrick Long, and David Watters with Representatives Alexis Simpson and Laura Telerski, would have created a standing commission to study, among other things, how AI development affects workforce composition and employment here. It cleared the Senate on March 26.

The House rewrote it. The version that came back was state-agency information technology governance: guidance from the Information Technology Council, an advisory committee, an annual report on how agencies use these systems. Useful housekeeping. Not a study of what the technology is doing to Granite State workers. On May 21, Senator Bill Gannon moved that the Senate refuse to concur with the House amendment. It carried on a voice vote, and the bill died there. A separate AI regulation bill, House Bill 1725, died in February.

The public is well ahead of the building. In a Granite State Poll fielded in late May, the University of New Hampshire Survey Center found nearly two-thirds of residents expect AI to hurt the country over the next decade. One in five working Granite Staters thinks it likely they will lose their job to it within ten years. A plurality would halt new AI data center construction in the state outright, and two-thirds would oppose one in their own town.

Voters worried about their paychecks. A legislature that produced an IT reporting requirement.

Three things that should happen before the 2027 session

None of this requires believing the worst-case scenario. It requires measuring.

First, Employment Security’s Economic and Labor Market Information Bureau should publish a breakdown of New Hampshire occupational employment by AI exposure. The exposure indices are public and validated against real hiring outcomes; the occupational data already sits on the bureau’s shelf. A report, not a research program.

Second, the Department of Revenue Administration should model what happens to the enterprise value tax base under a range of compensation-growth scenarios, including flat and declining. The agency models proposed tax changes for the Governor’s office as routine. The same exercise, pointed at a harder question.

Third, the 2027 session should refile the workforce study as its own bill, with its own hearing, not as language riding inside a technology-governance vehicle that a second chamber can quietly strip. If a commission is worth having, it is worth a standalone vote.

The man who thinks no machine can do his job and the executive who thinks he can delete the payroll and keep the customers are both certain, and both certainties are unearned. New Hampshire does not get to sit that argument out. It wrote a tax that only works when Granite Staters get paid.

— Dexter Dow, Granite State Report

Fact check

Editor’s note. Every factual claim above was checked against primary sources before publication; see the fact-check table. Three limitations, stated plainly rather than buried. First, no interviews were conducted for this piece; it rests on documents, agency data, and published research. Second, New Hampshire Employment Security’s monthly figures are quoted from the agency’s release pages and cross-checked against the U.S. Bureau of Labor Statistics’ July 21 state employment release. Third, the connection drawn here between AI adoption and New Hampshire’s tax base is a structural argument about how RSA 77-E is written, not a modeled forecast; no agency or institute has quantified it, which is the reason this article calls for that work. Where sources conflict on the size of AI’s labor market effect, this piece treats Acemoglu’s peer-reviewed estimate as stronger than Chen’s unreviewed preprint and says so in the text. Corrections: Granite State Report corrects verified errors promptly and appends a dated note identifying what changed and when. The corrections policy and the corrections register are public, and correction notes are never removed.