The 21p AI Minute Is a Price for the Wrong Thing

A proposal to charge artificial intelligence at Britain’s minimum-wage rate would turn today’s few-pence model calls into a potentially transformative levy: at £12.71 an hour, every notional minute of machine work would cost about 21p. The instinct behind it is serious — if automation erodes jobs and the payroll taxes that support public services, the gains should not accrue only to technology owners — but pricing inference as if it were a worker’s time confuses the cost of a tool with the value of the work it replaces.

A moral impulse in the wrong category

The taxi is an effective way to expose what is at stake. A passenger may choose a human driver even when a driverless car is cheaper, because the passenger prefers that the fare support another person’s livelihood. Someone else may take the autonomous ride because the lower price matters more. Neither choice is necessarily irrational. One is a consumer preference for a social outcome; the other is a decision to save money. A serious policy must acknowledge both without pretending that one can be removed by changing the number on the fare display.

The proposed AI base rate begins with an equally recognisable unease. A business can ask a model to classify documents, draft messages, research a question or write code for an amount that bears little resemblance to the cost of paying a person to do similar work. If this becomes a substitute for employment, why should the machine’s apparent labour be treated as almost free while the human worker’s labour is subject to a legal wage floor, tax and social-insurance contributions? Why should the public carry the costs of displacement while private firms retain the productivity gains?

Those are legitimate questions. The phrase “tax the AI like a worker” does not answer them. A model is not an employee who receives wages, needs housing, pays rent or makes a claim on social insurance. A company buys computation, access to software and often a package of other services. Its use may reduce demand for certain human tasks, complement existing staff, create entirely new work, or do several of these at once. Calling all of that “AI work” describes the system in a human metaphor; it does not establish that a minute of machine activity is a minute of labour displaced.

The proposal’s strongest version is therefore not that a chatbot should be put on a payroll. It is that governments may need a new way to share the gains when firms automate, and the tax base built around wages shrinks. The distinction matters. Once the state calls an inference charge a wage floor, it inherits every weakness of that analogy: no agreed measure of minutes, no stable definition of the worker, no natural employer, and no reliable account of which human job has disappeared. The political case for sharing gains deserves better than a meter built around a metaphor.

Not sure what this means for locally installed AI(LLM)? Would using local AI (unregistered) be like tax evasion through illegal employment?

That case also has a less comfortable side. A price high enough to discourage substitution is not merely a tax on a technology company. It is a tax on the customer who uses a model to translate a letter, a small firm that cannot afford a research department, a school testing a tutoring tool and a software team automating repetitive checks. The person whose livelihood policy means to protect may also be the person who benefits from a cheaper tool. The policy has to decide whether it values the worker, the user, the firm, the taxpayer or the activity — and cannot evade that choice by saying “make AI pay”.

A wage floor is not a tax

Britain’s National Living Wage for workers aged 21 and over is £12.71 an hour from 1 April 2026. Divide the hourly rate by 60, and you get 21.18 pence a minute; eight minutes of paid work would be about £1.70 in gross wages, before employer National Insurance, pension contributions, holiday pay, equipment, or supervision costs. The rate is a statutory wage floor for covered workers, not an all-in price tag on everything an employee does.¹

The example is rhetorically powerful because it supplies a neat comparison: a person’s eight minutes cost pounds, while a model’s short answer costs pennies. But governments do not usually collect income tax and social insurance by putting a legal minimum price on the activity that might compete with employment. PAYE is a mechanism for withholding tax and National Insurance from wages paid through employment. The employer’s payroll is the collection point because there is an employer, an employee, and remuneration for work. US federal employment taxes are similarly tied to employee wages and compensation. The proposed charge would instead tax the purchase or use of a computational service, even if its advocates describe that service as labour.²

That does not make a new levy illegitimate. Governments tax consumption, energy, property, transactions and corporate income for reasons other than the tax status of the item itself. They could choose to tax commercial AI use. But a charge on every model minute would not, in any meaningful administrative sense, be PAYE for a non-human employee. It would need its own legal basis, its own definition of the taxable event and its own rules about who owes it. Calling it “AI National Insurance” would make it sound familiar while hiding the actual policy decision.

Even the wage selected as the benchmark is not a neutral fact. Britain’s minimum wage is a policy rate, set through a process that weighs a target relative to median earnings alongside employment conditions, business effects and the cost of living. Germany’s statutory minimum wage is €13.90 an hour in 2026. The US federal floor remains $7.25, but many state and local rates are higher; the Department of Labor’s July 2026 table lists 30 states, as well as the District of Columbia, Guam, Puerto Rico and the Virgin Islands, above the federal level.³⁴

In the transition, the customer could, for ideological reasons, choose to enrich a human rather than enrich the billionaire owner of a big tech company.

A transatlantic “Western rate” would have to choose among these benchmarks, or invent a basket of them. Then it would have to decide whether the conversion is based on market exchange rates, purchasing power, local wages, or where a customer happens to be. Those options do not converge. A local rate might meet one country’s wage policy while disadvantaging another country’s firms; a common rate could become a bargaining contest between governments. Minimum wages have always been tied to local institutions and labour markets. A globally uniform AI minute would be a different instrument, not the international extension of an existing one.

The arithmetic is shocking, but it is not the model

The cost contrast is real. Anthropic’s published price for its Claude Opus 5.5 API is $4 per million input tokens and $20 per million output tokens. Google’s Gemini API lists much cheaper models alongside more expensive options; Gemini 3.1 Flash-Lite, for example, costs $0.25 per million input tokens and $1.50 per million output tokens. A simple illustrative exchange of 2,000 input tokens and 800 output tokens would cost roughly 2.4 US cents on the former rates, before any special features or other service charges, and a fraction of a cent on the latter. The precise bill depends on the model, the prompt, the answer, caching, reasoning tokens, tools and the provider’s billing rules.⁵⁶

This helps explain why the comparison feels almost indecent. A worker paid at the UK adult minimum wage earns around £1.70 gross for eight minutes. A short API exchange may cost less than a small coin. But the figures describe different things. The wage covers a person’s time, including reading, deciding, correcting, and taking responsibility. The token bill prices a narrow slice of a technical service. It does not include the customer’s staff time spent shaping a request, checking the answer, integrating it, handling an error or accepting liability when the result is wrong. It does not even reveal whether the model completed the task in eight minutes, eight seconds or eight separate attempts.

A business comparing the two should ask what the model replaces, not how many seconds its servers were active. If an employee spends eight minutes copying a column between spreadsheets and a model does the same task in a fraction of a second, the AI charge could be minimal while the time saved is real. If a legal assistant spends eight minutes locating a relevant clause, the model may return an answer at a similar API cost, but the human review needed to trust that answer may still be required. If the request is for a poem, translation or first draft that no one would otherwise have commissioned, it may replace no paid job at all. The inference meter charges each use alike. The labour effect does not.

Nor does an API list price equal the provider’s full production cost or social value. It is a commercial price at one point in a competitive market. It may reflect hardware, power, model development, sales and support, but a published token rate is not a transparent accounting of each component. The rates also change as models improve, providers discount, and hardware becomes more efficient. A government pegging a statutory AI price to a human wage would be joining two moving targets that respond to entirely different forces.

At the proposed UK floor, the issue becomes stark. A per-minute charge of 21p would make every eight minutes of billable AI activity cost £1.70 before any other fees. For short written outputs, even premium inference can be priced far below that amount. The floor would not be a modest contribution on the margin; for many text tasks it would overwhelm the present model charge. It might make certain uses uneconomic, but it would not selectively stop a firm from replacing an employee. It would also make a person’s low-cost productivity tool, a researcher’s analysis and a business’s bulk classification more expensive, whether or not any job was lost.

Tokens measure activity, not the labour displaced

The crucial defect is hidden in the word “minute”. Human work is not a continuous tap that elapsed clock time alone can measure. People pause, switch tasks, consult colleagues and carry context between assignments. AI work is not a continuous shift either. A model may generate an answer quickly, wait while a search tool runs, receive more context, make a mistake, and be prompted to try again. An agent may sit idle while a test runs or a person decides whether to approve its proposed change. Which of those minutes count as work? Which count as the equivalent of a minimum-wage employee’s paid time?

Tokens are no better proxy. They are units used to meter text and other inputs and outputs; they are not units of economic value. Two requests with the same token count can vary enormously in difficulty. One may ask for a routine reformatting job; another may ask for a high-stakes decision involving incomplete evidence. Some models produce concise output after substantial internal reasoning, while others use more tokens to produce a less reliable result. In multimodal systems, the relevant use may involve images, audio, video, external search, code execution or computer control. A tax based on text tokens would privilege or penalise architectures by how they consume computation, not by the jobs they affect.

Trying to convert output into a standard “human-equivalent minute” only moves the argument to a new official formula. A speech model might be assigned a word-per-minute rate; an image model would need another conversion; a coding agent would need some measure of the task’s difficulty. Each formula would favour the things it can count. If the price is based on words, a terse model is cheaper than a verbose one even when their usefulness is identical. If it is based on wall-clock time, a provider can run the same work faster and be rewarded for efficiency. If it is based on compute, firms with custom chips or optimised models receive a different burden for producing the same outcome.

The policy would also need to decide where one unit of work ends. Is an AI call to draft a customer reply one taxable event, or is it part of a single customer-service session? If a spreadsheet formula is generated inside a product, does the software vendor remit the charge, the company that bought the licence, or the individual who clicks the feature? If a human asks one model to search, another to summarise and a third to check the answer, is that three workers, one task or one system? The more AI is embedded in ordinary software, the less plausible it becomes to treat each visible prompt as the complete unit of use.

Research that tries to measure autonomy makes the same distinction more rigorously. METR’s task-completion time horizon is based on the amount of time a human expert is estimated to need for a selected task, and the probability that an AI agent completes it reliably. The organisation explicitly cautions that this measures task difficulty, not a claim that an agent can work continuously for that duration or perform any job of equivalent length. Its test suite is weighted toward self-contained software, machine-learning, and cybersecurity tasks with defined success criteria; most workplace activity has more context, messier objectives, and human consequences.⁷

That is why “an AI minute replaces a human minute” cannot be the premise of a tax. The replacement relationship is not in the meter. It must be established through the task, the quality threshold, the need for human review, and the counterfactual: what would the employer have done if the model were unavailable? Any government that skips those questions will tax visible machine activity because it is measurable, not because it has identified a lost job.

More code is not more value

Software development offers a useful test because it exposes the distance between output that can be counted and work that matters. A June 2026 National Bureau of Economic Research working paper by Mert Demirer, Leon Musolff and Liyuan Yang analysed more than 500,000 GitHub developers alongside AI-use data. It found that coding activity rose substantially across successive tools: the estimated cumulative effect reached 240 per cent for commits when autonomous agents were included. Yet the gain narrowed further along the production chain: the number of projects rose by 80 per cent and actual releases by 30 per cent. The authors estimated an elasticity of substitution between AI and human effort of 0.23, a result they interpret as strong complementarity. The paper is a working paper, not a final verdict on every firm or labour market, but it gives a precise warning against treating code volume as headcount displaced.⁸

That gap is not an academic curiosity. A commit is an entry in a version-control history. A release is a change made available to users. A release that users do not adopt, a feature that does not solve a problem, or a fix that triggers a costly incident may have little value. The measure improves again only when the change helps the product earn revenue, retain customers, reduce risk or deliver a public service more effectively. A system that produces three times as many candidate changes but only 30 per cent more releases is not automatically three times as productive. It may have increased the workload of the people who must inspect and integrate those changes.

Consider the six-person product team in the proposal. It has engineers, a backlog of small tickets, a coding agent that attempts fixes and a review process before changes go live. One team might use the system to clear routine defects faster while leaving its staff in place to make architectural decisions, understand customer requirements and own production. Another might use the increase in machine-generated pull requests to reduce headcount. The AI product can be identical; management makes the labour choice under incentives the per-token rate does not control.

The team still needs someone to decide whether the ticket is worth solving, whether the proposed change is correct and whether passing tests are enough to justify a release. A second automated reviewer does not remove that responsibility; it adds another model’s output to a process that must be monitored. If the queue of generated changes expands faster than human review capacity, the scarce resource shifts from writing code to deciding which code to trust. An organisation that cuts reviewers because agents can write more code may discover it has bought a larger supply of unexamined risk.

It is tempting to infer a staffing formula — keep three of six engineers, say — from a productivity study. That would be false precision. The study measures developers’ activity and downstream output, not the staffing choice of every six-person engineering group. But it does make one assertion indefensible: that a 240 per cent rise in commits demonstrates a 240 per cent rise in finished value or an equivalent collapse in human labour. A minimum-price policy built on that premise would be both economically crude and socially misdirected.

A firm may choose to keep six people and produce more. It may reduce roles after real workflow redesign. It may use the tool to tackle work that had never been funded. A salary saved is only one possible outcome. The effects depend on product demand, quality, review costs, customer adoption, and whether the remaining work is still commercially worthwhile. A pricing policy that treats all three responses as a single outcome called “AI employment” cannot distinguish genuine substitution from an expansion in output.

The work after launch changes the test

The story of software is often told as a launch: a team writes code, ships a product and moves on to the next invention. In reality, a live system becomes a standing obligation. Its dependencies change, operating systems and cloud services are updated, security weaknesses need attention, regulations evolve, users ask for changes and the business model shifts. A database migration can be routine in one company and a serious risk in another because the software handles different customers, data and revenue. The product’s creation is visible; its continued usefulness is often the less glamorous work that keeps it from becoming obsolete.

That work falls into different kinds of change. Corrective work addresses defects and failures. Adaptive work keeps a product compatible with new platforms, vendors, standards and legal requirements. Evolutionary work changes the product so it continues to meet demand or defend its position. The categories overlap, but the distinction is useful. A developer updating an expired dependency, a security engineer fixing a vulnerability, and a product team redesigning checkout may all produce code, but counting pull requests doesn’t capture their impact.

Agents may be particularly attractive for repetitive work: dependency upgrades, boilerplate, test generation, targeted bug fixes, and migrations with clear before-and-after criteria. They can make previously neglected work cheaper to attempt. But the long tail of a system is not just a collection of small tickets. It includes tacit knowledge: why a scheduled process runs in a certain order, which customer contract depends on an old interface, what a “temporary” workaround was designed to protect, or which data cannot be reconstructed after an error. That knowledge may exist in a few people’s heads rather than in the repository the agent can read.

Automation can therefore reduce the cost of making a change while increasing the number of changes a business has to evaluate. A model may propose a tidier abstraction that makes the immediate test pass but leaves a more complex system for the next engineer. It may modernise a component without understanding the commercial or operational reason it was left alone. An engineer who remembers the history can reject a technically plausible change because it creates a larger risk elsewhere. That is not resistance to progress. It is the difference between local correctness and stewardship of a system over time.

You don’t need to claim that maintenance always takes a particular percentage of a software budget to see the policy implication. The relevant point is that a product’s value is not settled when it is first built. A firm that uses agents to touch more of its estate might preserve revenue, meet a compliance obligation or reduce a future failure. Or it might generate activity that leaves users no better off and creates more future debt. Either way, charging for machine minutes cannot distinguish the socially useful work from churn.

This changes the employment question too. If a company removes the engineers who know how its older systems behave, it may save wages today while increasing the chance that later AI-generated changes damage the service. An agent can propose, test and document changes; it cannot assume the firm’s legal responsibility for the outcome. A minimum AI charge might slow the number of proposals. It would not ensure that someone remains who understands the system, that a junior engineer can learn it, or that the next migration preserves the product’s value.

A blunt floor invites the wrong exits

A charge based on minutes is designed to stop firms buying cheap artificial labour. Yet the first response would be to change what gets counted. Providers could price by task, bundle use into a subscription, classify models as software features, offer on-premises systems or route requests through a cloud provider in another jurisdiction. A business with enough scale could train or host its own open-weight model and pay no external API bill. An individual could run a small model on a device. The law would need to decide whether it taxes a provider’s invoice, a user’s prompt, a volume of computation or the capability of the model itself.

Each boundary produces a different loophole. If the rule applies to a company that sells model access, it may miss AI built into business software or inference performed by the user. If it applies to the server’s location, firms can move workloads to places outside the scheme. If it applies to the user’s location, regulators need a rule for mobile workers, multinational operations, and software that quietly calls models in the background. If open-weight models are exempt because no commercial provider bills per token, the policy may give firms an incentive to internalise inference rather than buy it from a local service. Tax the electricity or compute hardware instead, and the charge becomes a broader industrial policy that also affects non-AI activity.

The international benchmark problem is not just a paperwork issue. In the United States, the federal floor is one rate, but state and local rates differ; in Europe, national systems and currencies differ. A business choosing where to process data can compare the tax attached to inference alongside energy, land, regulation and staff costs. A unilateral levy may raise a little money from activity that cannot move and push more elastic work elsewhere. The firms most able to relocate are usually not the small local customers the policy claims to protect.

There is also a practical paradox. The expensive long-running agent loop may be a better target for a spending control than the ordinary text prompt, but a per-minute floor would charge both in the same currency. If an agent makes hundreds of calls, searches a large corpus, launches tests and retries failures overnight, the total cost can become significant even at low token rates. The remedy is not necessarily to make every useful exchange cost 21p per minute. Providers can offer hard spending limits, execution caps, approval gates and visible reporting; employers can specify who may run agents and when. These measures address runaway use directly, rather than relying on a wage comparison that has little to do with the incident.

A policy can, of course, accept that some migration will occur and set a levy anyway. But it should not promise that a national fee will keep AI spending at home. It may instead protect incumbent providers that can absorb compliance costs, raise the entry barrier for smaller competitors, and make local businesses pay more for access to a tool their competitors can obtain elsewhere. “Keep the work in the country” sounds like a straightforward goal. With remote computation, open models and global software distribution, it needs an enforcement architecture, not just an invoice line.

The result could be a familiar regulatory pattern: the largest companies build compliance departments, vertically integrate, and continue; small suppliers exit or sell; users who can afford a local deployment switch to it; and the people the fee was meant to help see little revenue because the taxable base has shifted. That is a risk, not a certainty. But any plan that relies on providers voluntarily reporting every use and on all firms staying inside one jurisdiction has not yet answered the central question of enforcement.

Self-reported productivity is not a payroll forecast

People working with coding and research agents often say they are much faster. That report should not be dismissed: experienced users know where the tool saves them time, and first-hand accounts can reveal uses that formal studies have not yet measured. But speed, output and value are different quantities. A task that once took an afternoon may now take an hour; the saved time could produce more useful work, more low-value work, or simply a larger stack of drafts that still need checking.

A May 2026 METR survey of 349 technical workers found a median self-reported increase of between 1.4 and two times in the value of their work attributed to AI tools, while the median self-reported speed increase was three times. METR itself flags reasons for caution: respondents may overestimate counterfactual gains, the sample was a convenience sample, and survey beliefs are not a direct measurement of output. That is useful evidence about how workers perceive the technology, not a forecast that three quarters of a team can be dismissed or a demonstration that a firm’s product value has tripled.⁹

METR’s time-horizon work reinforces the caution from another direction. The benchmark estimates how difficult a defined task is for an agent to complete at a given reliability level. It is not a measure of how many minutes a model is “employed”, and its benchmark tasks do not reproduce the full context of a workplace. A company’s most valuable work may depend on knowing a customer, negotiating with a colleague, spotting that a request is misguided or deciding what failure would mean. None of those skills becomes a token count because an agent is part of the workflow.

The policy problem is therefore not to deny that AI can replace work. It is to refuse the easy substitution of a claim about productivity for evidence about employment. A firm can use AI to increase each employee’s output, avoid hiring additional staff, remove positions after reorganisation, or bring tasks in-house that were previously outsourced. The same headcount outcome can mask different effects on wages, job quality and future hiring. A payroll-tax gap may emerge in some sectors even when aggregate employment does not collapse; conversely, productivity gains do not prove that a particular worker’s job has been displaced.

The public-finance concern is real. When income moves away from wages, tax systems that rely heavily on labour income and social contributions may collect differently. Yet the right response depends on what has moved: a job, a contract, a profit margin or simply the organisation of a task. A levy on all AI consumption may tax firms that are using AI to make existing workers more productive, while leaving a company that moves work offshore or changes its corporate structure untouched. A tax instrument should follow the economic event it is meant to address, not the most visible interface through which it passes.

Make the tax base fair, then help people through change

The OECD’s work on employment-form taxation offers a useful, if limited, analogy. Its 2019 study examined how eight countries taxed standard employees and self-employed workers, including social contributions. It warned that different tax treatment can create incentives to choose one form of labour contract over another, and it evaluated the issue through equity and neutrality. It was not a proposal for an AI tax, and its rules are not a template for one. Its relevance is more basic: even when two people perform similar tasks, tax policy must decide whether the legal and economic differences between their situations justify different treatment. The AI case is harder because one side of the comparison is not a person earning income at all.¹⁰

If governments want to protect the revenue that funds public services, they can begin by stating that objective directly. They can review how business profits, capital income and labour income are taxed when firms replace payroll with automation. They can examine whether existing tax rules subsidise investment in machines more generously than investment in people. They can enforce the tax laws already on the books, scrutinise how multinational companies allocate profits, and consider how to finance social insurance as work arrangements change. None of those measures is painless or simple; the point is that they address the tax base rather than pretending the model has become an employee.

A tempting answer is to place a charge on AI-generated value and use the proceeds to pay people displaced by automation. But “value generated by AI” is even harder to measure than minutes. A new product might depend on software, data, electricity, human design, customer relationships and a model service; assigning its profit to the AI contribution would involve contestable accounting. A charge on the API bill is administratively cleaner but bears little relationship to the value created. A levy on profits better follows an ability to pay, but firms can shift profits, and successful companies may create value through many inputs. A levy on headcount reductions can be gamed through contractors, attrition or reorganised subsidiaries. Each option carries a different cost and loophole.

Still, governments have choices short of a universal AI wage. They can make transition assistance automatic and portable rather than waiting for a worker to prove that a particular model caused a redundancy. They can fund wage insurance, retraining with a direct link to actual vacancies, unemployment support and apprenticeships that preserve routes into skilled work. They can require large employers to disclose how they deploy automation, including changes in staffing, hours, pay, output, and service quality. Such disclosure would not by itself redistribute anything, but it would improve the evidence on which a targeted intervention could rest.

A support system should not be sold as compensation for every technological change. Some tools create new services, improve safety or remove work that people reasonably do not want to do. Some displace tasks without eliminating the job. Others do eliminate roles, concentrate bargaining power or transfer productivity gains from workers to shareholders. A serious policy should separate these cases where possible, while admitting that no public dataset will tell us the counterfactual perfectly. It should not promise that a fee can freeze a specific occupation in place.

There is also a case for looking at the tax treatment of employment itself. If a firm pays a worker, it incurs wages, employer contributions and legal duties. If it substitutes a service contract or a piece of software, those costs may be different. That can create an incentive to organise activity in one legal form rather than another. The OECD’s analysis shows how this issue arises even between employees and self-employed people, who share a human capacity to earn and spend income but bear different rights and risks. It is a more useful guide to the AI question than treating a model as an employee: identify the distortions, decide whether they are justified, and legislate for the actual legal forms at issue.

If the aim is restraint, measure outcomes

The proposal can be rescued only by changing its purpose. If the aim is to stop every AI substitution by making machine use as expensive as low-wage human time, then the proposal is a direct attempt to suppress automation. Supporters should say so. They should also explain why this is preferable to raising worker incomes, supporting transitions or bargaining over how productivity gains are shared. If the aim is to collect revenue as payroll’s share falls, a minimum-wage meter is poorly fitted to that job. If the aim is to limit runaway agents, a wage-linked per-minute charge is indiscriminate. If the aim is to preserve domestic AI businesses, a tax that drives inference elsewhere may do the opposite.

A better first instrument is not a new slogan but a record. Large employers using AI at scale could report categories of use, whether systems are internal or supplied by a vendor, how many human tasks they augment or replace, what happened to hours and headcount, how output changed and what new review work was required. Reporting should be designed to avoid false precision: firms cannot always know which role would have existed without a tool. But longitudinal data can show where automation is associated with fewer hires, lower labour demand, higher output or a change in job quality. Without such information, the debate will be forced to rely on anecdotes from workers, vendor claims and guesses about the counterfactual.

The performance measure should also sit closer to the result. In coding, that means not just tokens or commits, but changes that are accepted, released, used and maintained without an increase in failures. In customer service, it means resolved cases, repeat contacts, satisfaction and the share of problems passed back to a human. In research, it means an answer that survives review and informs a decision, not merely a large bundle of generated pages. A government cannot tax every such outcome individually, nor should it necessarily try. But those measures help distinguish productivity gains from activity generation.

For a six-person engineering team, a fair account would track the number of people doing specification, review, security and product ownership; the number of releases; how often changes are reverted; how much time staff spend checking machine output; and whether the product’s reliability and user adoption improve. It would also ask what happened to junior roles. Firms may save money in the short term by using agents for work once assigned to beginners. But if no one is trained to become the next senior engineer, the company may quietly remove the apprenticeship path on which its future expertise depends. A charge on tokens cannot make that decision for the employer.

This does not imply that every firm should retain its old staffing level. It means assessing a headcount cut against the total operating model, not the tool’s invoice. If employees are removed while review work, risk and infrastructure costs increase, the business may have shifted rather than eliminated labour. If the tool lets the same staff deliver more useful work and maintain a more reliable service, the gain is real even if no one is displaced. A tax regime that treats both outcomes alike rewards neither responsible adoption nor responsible employment.

The next question is how to design any future levy. A broad fee on commercial inference could be tested at a low rate, made visible, and paired with an independent review of effects on small firms, public services, local providers and offshore activity. If the purpose is worker transition, revenues should be ring-fenced or transparently connected to a durable support fund rather than promised in the abstract. If evidence later shows a specific pattern of large-scale displacement or a clear tax advantage favouring automation, a more targeted policy could be debated. A mechanism should be revised when it fails; a “minimum AI wage” cannot be treated as a moral principle immune to its own outcomes.

No one needs to pretend that AI’s cost advantage is an illusion. A few cents can buy a useful draft or a first pass through information that would otherwise take a person longer. The gap is exactly why businesses are adopting these systems and why workers fear what comes next. But a cheap tool is not automatically a cheap substitute. The real price includes the people who set the goal, test the answer, absorb the failure, maintain the system and remain accountable to customers. It also includes what society loses when someone’s wage, experience or route into a profession disappears.

At £12.71 an hour, Britain’s adult minimum wage converts neatly to about 21p a minute. That figure makes for a sharp political comparison. It does not tell a tax authority what an AI minute is, what job it replaced, what value it produced or who should pay. The arithmetic is a useful alarm; it is not a workable tax code. Governments will have to decide how to share the gains and insure the people who bear the costs. They should do so by measuring economic outcomes and taxing actual income or use under clear rules, not by pretending that a machine’s inference clock is a human payslip.

References

1. https://www.gov.uk/government/publications/the-national-minimum-wage-in-2026/the-national-minimum-wage-in-2026
2. https://www.gov.uk/paye-for-employers; https://www.irs.gov/businesses/small-businesses-self-employed/understanding-employment-taxes
3. https://www.dol.gov/agencies/whd/mw-consolidated
4. https://www.bundesregierung.de/breg-de/aktuelles/mindestlohn-steigt-2391010
5. https://www.anthropic.com/claude-opus-5-5
6. https://ai.google.dev/gemini-api/docs/pricing
7. https://metr.org/time-horizons
8. https://www.nber.org/papers/w35275
9. https://metr.org/blog/2026-05-11-ai-usage-survey
10. https://www.oecd.org/content/dam/oecd/en/publications/reports/2019/03/taxation-and-the-future-of-work_3a1fe35a/20f7164a-en.pdf