The Tax Code Is Quietly Paying Companies to Replace Workers with AI

Most of the public conversation about AI and jobs is aimed at the wrong target.

Executives are accused of greed. Companies are accused of recklessness. CFOs are cast as villains for choosing software over people. The moral framing is satisfying, but it obscures the actual mechanism — and as long as the mechanism stays hidden, no amount of public pressure will change the outcome.

The mechanism is the tax code.

The U.S. tax code is quietly paying companies to replace workers with AI. Not metaphorically. Literally.

Let Me Show You Some Math

Run the numbers on a single $100,000 role.

If a company hires a person at $100,000, it also pays roughly $7,650 in employer payroll tax — the 6.2% Social Security contribution and 1.45% Medicare contribution.¹ Total company cost: $107,650. After applying the corporate tax deduction on that expense, the company is out roughly $85,000.

The federal government, meanwhile, collects roughly $28,500 in direct worker-linked revenue from that single job: $7,650 in employer payroll tax, $7,650 from the employee’s share of payroll tax, and about $13,170 in federal income tax for a single filer taking the 2026 standard deduction.² That figure does not include state income tax, which in most jurisdictions adds another $4,000 to $8,000.

Now replace that person with $100,000 of AI software that does the same work.

The company pays the $100,000 and deducts the full amount against income. Many purchased AI tools are currently deductible immediately: off-the-shelf software can qualify for Section 179 treatment, while SaaS subscriptions are typically deducted as operating expenses.³ Net cost after deduction: about $79,000. The government collects almost nothing — no payroll tax, no income tax on a worker who is no longer there. The only meaningful federal revenue is corporate income tax on the software vendor’s margin, which on a $100,000 sale at a typical software margin amounts to a few thousand dollars at most.

Same work. Same price tag. Different tax bill.

The company saves about $6,000 per substitution. The public loses on the order of $20,000 in direct worker-linked revenue, offset only partly by taxes collected elsewhere in the software supply chain.

Multiply across an enterprise replacing 500 roles. The company captures about $3 million in tax-driven savings. The public absorbs a revenue gap measured in the millions — money that funded Social Security, Medicare, and the general budget.

Don’t Blame the Companies. Blame the Incentive.

Here is where the conversation usually goes off the rails. The instinctive reaction to numbers like these is to blame the companies making the substitution. That is the wrong target.

CFOs are doing exactly what the tax code rewards them for doing. The system was written when labor and capital looked nothing like they look today. Payroll tax was designed in 1935 to fund a safety net by taxing the act of employing a person. Software depreciation rules were written to encourage investment in capital equipment. Neither was designed for a world in which capital and labor are interchangeable inputs producing the same output.

The result is an accidental but real subsidy: every time work moves from a person to a machine, the public loses revenue and the company captures the difference. No one designed it this way. But no one is fixing it either.

That revenue does not have to disappear. Tax the AI doing the work at the same rate we tax the labor it replaced, and the public revenue stays intact. The company still gets its productivity gain. The safety net still gets funded. No one has to lose a job to make the math work.

The framing of the problem matters. This is not a question of slowing AI down, or capping productivity, or punishing innovation. It is a question of removing an artificial discount that the tax code is currently extending to one side of a substitution decision. Fix the price signal, and the decisions that follow are undistorted decisions about where AI augments people versus where it replaces them — not decisions shaped by an outdated incentive.

What Government Should Actually Do

The instinct of policymakers in moments like this is to do nothing, or to propose sweeping legislation that never passes. Neither is the right move. There is a set of practical, targeted reforms that would correct the distortion without slowing AI investment. None of them require new agencies, new mandates, or new restrictions on what companies can build. They simply align the price signal with the public outcome.

Seven are worth considering now.

1. Make automation pay what labor pays.

Society exists to elevate being human, not penalize it. The current tax code rewards companies for moving work off people and onto software. That is backwards.

The most direct correction is to apply a payroll-equivalent levy to AI doing work a person used to do — roughly the 7.65% an employer currently pays on a human in that role. Companies still capture the productivity gain from automation. They simply stop getting a discount for choosing machines over people. The revenue replaces the payroll tax lost when the role disappeared, and flows to the same funds — Social Security, Medicare — that the original role would have helped underwrite.

Defining the levy precisely matters. It cannot apply to every software purchase or it becomes a tax on technology adoption broadly. It needs to attach to AI deployments that demonstrably displace defined workforce categories. The mechanism is closer to a payroll tax than to a sales tax: tied to the function being automated, not to the software itself.

2. Fund worker transition through outcome-based markets.

Government-run retraining programs have a poor track record. They tend to be expensive, slow, and disconnected from the actual labor market. The alternative is not more programs — it is markets.

Pay training providers only when a displaced worker lands in a comparable or better job. Make placement, not enrollment, the trigger for payment. The result is a private industry whose business model is getting people back to work, with capital flowing to the providers that actually deliver outcomes.

This is not theoretical. Outcome-based contracting has been tested in workforce development pilots in the UK and in several U.S. states, with mixed but instructive results.⁴ The design challenges are real — defining “comparable or better job,” preventing cream-skimming, ensuring providers serve harder-to-place workers — but they are design challenges, not insurmountable obstacles. The current system pays for activity. An outcome-based system pays for results.

3. Reward augmentation, not just replacement.

The first three reforms address the substitution side of the ledger. The fourth addresses the alternative.

Give companies a tax credit when they deploy AI in ways that demonstrably increase employee productivity without reducing headcount in the affected function. Structure it similarly to the existing R&D tax credit under Section 41: self-certified, IRS-audited, tied to documented productivity and headcount metrics in defined cost centers.⁵ The credit would phase in over two to three years to allow companies to adjust their reporting.

The R&D credit is an imperfect instrument. Attribution is fuzzy. Companies that were already doing research capture a disproportionate share of the benefit. Yet it has been considered broadly effective for forty years and is one of the largest business tax subsidies in the code — the Joint Committee on Taxation estimates it will reduce federal revenue by roughly $189 billion from FY2025 to FY2029.⁶ An augmentation credit would face the same imperfections and produce similar results — tilting marginal decisions toward AI deployments that keep people employed.

Critics will point out that this rewards companies that were not going to cut anyway. That is a fair critique. But the goal of an incentive is not to coerce bad actors. It is to shift marginal decisions at the boundary, and to signal what the public values. Right now the code signals that the public values replacement. An augmentation credit signals the opposite.

4. Modernize the tax-free education benefit for the AI era.

The framework above corrects the asymmetry between labor and software. But the tax code also contains a quieter distortion: it underinvests in the one mechanism most directly aimed at keeping workers relevant as the technology shifts.

Section 127 of the tax code lets employers provide up to $5,250 per year in tax-free educational assistance to employees, a cap set in 1986 that remained unchanged for nearly forty years. The 2025 reconciliation act made the benefit permanent and scheduled it to begin inflation indexing after 2026, but the base amount itself is still far below where it would be if it had simply tracked inflation since the 1980s.⁷ Raise the base to $12,000, keep the indexing in place, and explicitly extend eligibility to AI literacy, prompt engineering, AI-augmented workflow training, and related certifications. The result is a tax-free reskilling benefit large enough to actually matter, and a tool employers can use to retain people rather than replace them.

This is a correction to an existing rule, not a new program. It also sends a clear signal: the tax code should reward investment in people, not penalize it.

5. Professionalize the trades and modernize the pipeline.

One of the strangest features of the current debate is that everyone worries about where displaced workers will go, while the economy is simultaneously short several million people in fields it desperately needs.

The skilled trades are facing a crisis that runs in the opposite direction of the AI conversation. The United States is projected to be short roughly 2.1 million skilled trades positions by 2030 — electricians, HVAC technicians, plumbers, pipefitters, construction equipment operators, and general maintenance workers, per JLL research drawing on Department of Education data.⁸ Electrician employment is projected to grow roughly three times as fast as the average for all occupations through 2034. Roughly 39% of U.S. facilities managers are over 55 and approaching retirement, compared with 28% across all occupations. The pipeline is not keeping up, and compensation has finally started to respond.⁹

This is the labor market hiding in plain sight. The same data centers that AI requires are the buildings that need electricians to wire them, HVAC technicians to cool them, and plumbers to plumb them. The infrastructure buildout the country is now committed to — for AI, for energy, for manufacturing reshoring — depends on a workforce that does not currently exist at sufficient scale.

The policy response should be structural, not cosmetic. Triple federal funding for registered apprenticeship programs, with a focus on trades aligned to infrastructure buildout. Create a federal tax credit for companies that establish or expand apprenticeship slots, modeled on the existing Work Opportunity Tax Credit. Tie federal infrastructure spending to apprenticeship participation, as the Inflation Reduction Act already does for clean energy projects. Fund community college trades programs at parity with four-year university support. Standardize credentialing across states so a credentialed electrician in Ohio is a credentialed electrician in Texas.

The trades are not a fallback. They are skilled, technically demanding work, increasingly involving networked systems and digital fluency, and they pay well. The cultural prejudice against them — the assumption that a four-year degree is the only legitimate path — is a luxury the next decade cannot afford.

6. Professionalize care work.

The other large category of work the country urgently needs filled is care — for children, for the elderly, for people with disabilities. The supply gap here is even larger than in the trades, and the structural problems are deeper.

The Department of Health and Human Services projects that the supply of direct care workers will fall short of 8.9 million projected job openings between 2022 and 2032.¹⁰ In January 2026, care-related sectors — at-home care services, hospitals, and long-term care facilities — added 124,000 of the 130,000 net jobs created economy-wide, with much of that growth concentrated in roles that support older Americans.¹¹ By 2030, all baby boomers will be older than 65; by 2040, roughly 78 million Americans are projected to be 65 or older.¹² The demand is not a forecast; it is already here.

And yet care workers are paid little enough that many qualify for the federal safety net themselves. The median home health and personal care aide earns about $34,900 per year, or roughly $16.78 per hour — barely above the 2026 federal poverty guideline for a family of four in the contiguous United States, and well below the $49,500 median for all U.S. workers.¹³ Because part-time work is common in the field, many direct-care workers earn meaningfully less than the median on an annual basis. Child care workers earn less still, with a median hourly wage of $15.41.¹⁴ Roughly 87% of direct care workers are women; more than a quarter are immigrants. This is essential work, performed disproportionately by people the rest of the economy has historically underpaid and undervalued.

The policy response has to do more than raise wages, although raising wages is part of it. The deeper problem is that care work is structured as a low-status, low-credential field with no clear professional ladder. Treating it as a serious profession would change that. Establish a federal credentialing framework for elder care, child care, and disability care, with defined skill levels, portable credentials, and clear pathways to advancement. Fund the training that lets workers move up that ladder. Tie federal reimbursement rates — particularly through Medicaid, which funds a large share of direct care — to wage floors that actually reflect the work’s value. Make these jobs the kind of careers that attract people, retain them, and pay them like the essential work they are.

The framing matters. If AI displaces millions of routine cognitive jobs in the next decade, the question of where those workers go is not abstract. Some will move into the trades. Some will move into care. Both fields will absorb significant numbers of people only if the work is structured to make that movement worth it — with real wages, real credentials, and real professional standing.

7. Build a federal data foundation for AI workforce impact.

This is the least visible recommendation on the list and arguably the most important.

The United States measures unemployment, inflation, and GDP in close to real time. It maintains detailed industry-level employment data through the Bureau of Labor Statistics, with quarterly updates and monthly establishment surveys. None of that infrastructure is currently tuned to capture AI’s effects with any precision. We have anecdotes from companies and projections from consultants. We do not have a federal data series that tells us, function by function, which jobs are being augmented, displaced, or created by AI deployment.

BLS should be funded to build it. The methodology challenges are real but solvable: establishment surveys can be extended to capture AI deployment by function, occupational employment statistics can be cross-referenced against firm-level AI investment, and longitudinal data on displaced workers can be assembled from existing administrative records. The cost is modest in federal terms. The payoff is the ability to make policy on evidence rather than impression.

You cannot make smart policy on a system you cannot see. Every other recommendation on this list depends on data that does not currently exist. Building it is the foundation that makes the rest of the agenda credible.

Why This Is Different From Past Technology Shifts

One reasonable objection to all of this is historical. Every previous wave of automation — mechanization in agriculture, electrification in factories, robotics in manufacturing, the internet in services — produced eventual employment growth despite displacing specific roles. Why intervene this time?

The answer is that we may not know the full shape of AI-driven displacement until it has already happened. The pace of substitution is faster than in previous waves. The breadth is wider — not confined to manufacturing or to specific occupational categories. And the substitution is happening in service and knowledge work, where the worker-protection infrastructure of the last century never fully developed in the first place. There is no AFL-CIO equivalent for displaced paralegals, customer service representatives, or coders.

It is possible that, as in previous waves, AI ultimately produces net job creation. It is also possible that the transition takes a generation, that the new jobs require skills the displaced workers do not have, and that the cost of waiting to find out falls disproportionately on people who can least afford to absorb it.

Building the infrastructure to manage the transition is not a bet against AI. It is a bet against being wrong about how it plays out.

The cost of the proposed framework, if AI proves benign for employment, is modest. The cost of doing nothing, if it proves disruptive, is enormous.

The Choice in Front of Us

None of this is about slowing AI down. None of it caps investment, restricts deployment, or tells companies what to build. The framework above leaves every productivity gain from AI in place. What it changes is the question of where some of that gain flows.

Right now the answer is: entirely to the companies that deploy the technology, with the public absorbing the displacement costs through reduced tax revenue, weakened safety-net funding, and a workforce navigating a transition without the supports that previous transitions built over decades. That is not a sustainable arrangement, politically or economically. And it is not the result of any deliberate policy choice. It is the result of an accidental subsidy embedded in tax rules written for a different economy.

Fixing it requires a clear-eyed account of how the current system actually works, the willingness to act on that account, and a realistic view of what government can and cannot do well. Government cannot decide which AI deployments are good and which are bad. It cannot pick winners among technology vendors. It cannot run a successful retraining program by direct provision.

What it can do is fix the price signal. Make the costs of substitution visible. Fund the institutions that help workers move. Reward the deployments that keep people in jobs alongside the technology. Build the pipelines into the work the country actually needs done. Build the data that lets the rest of us see what is actually happening.

That is the agenda. And it is achievable inside the next Congress, if anyone in Washington is paying attention.

The companies are not the problem. The rules are. And the rules can be rewritten.

I fully acknowledge that this is a complicated issue, and that thoughtful people will land in different places on it. But even if you disagree with the specific reforms above, we should at least be willing to have the conversation openly and constructively. The tax code is already making the choice for us, every day, in the absence of any debate at all. Let’s be bold enough to have the debate.