Employer of Record for Robots? Why AI Labs Might Become the New Payroll Problem

In March 2026, Peter Bailis left his role as Chief Technology Officer at Workday to join Anthropic as a member of technical staff. He’d been in the CTO seat for less than a year. Workday, for context, is an $8 billion revenue company with 18,000 employees. The move wasn’t a redundancy or a restructure. It was a deliberate choice to leave enterprise HR software’s top table and walk into a frontier AI lab that is now openly building the kind of HR applications Workday sells.

The competitive irony writes itself. Workday’s own CEO had publicly stated that Anthropic, Google, and OpenAI all use Workday’s software internally. The company that bought the licences has now hired the CTO to build the replacement.

But this article isn’t about a CTO changing jobs. That’s a talent signal, and an interesting one, but the bigger story is what happens when AI labs stop selling tools to HR departments and start becoming the employment infrastructure itself. Because if you follow the trajectory far enough, you arrive at a question that governments, tax authorities, and workforce planners will have to answer sooner than they think: who employs the agent?

From Tool Vendor to Employer Proxy

The current generation of AI in HR is still largely assistive. Draft a job description. Summarise a CV. Generate an offer letter template. Useful, but firmly in the “copilot” category. A human recruiter or HR professional reviews, edits, approves, and acts.

That is changing fast. Anthropic has launched Claude plugins for HR use cases including generating job descriptions, onboarding materials, and offer letters. It has posted roles tied to building “people products” that support hiring, training, and promotions. These aren’t peripheral experiments. They’re products. And the direction of travel is clear: from assisted workflows to autonomous execution.

Now imagine the near-term future. An AI agent receives a hiring requisition, writes and posts the job advertisement, screens and ranks candidates, schedules interviews, generates interview scorecards, drafts the offer, processes the onboarding paperwork, and enrols the new hire in benefits and payroll systems. No human touches the process end to end. The recruiter role hasn’t been augmented. It has been replaced by a software agent running on infrastructure owned and operated by an AI lab.

At that point, the platform provider starts to look structurally identical to an Employer of Record.

For those unfamiliar with the term, an Employer of Record (EOR) is a third-party entity that legally employs workers on behalf of another company. The EOR handles payroll, tax filings, benefits, and compliance with local labour laws, while the client company directs the day-to-day work. It’s a model that has grown significantly in recent years, driven by remote work and global hiring. Companies like Deel, Remote, and Globalization Partners have built billion-dollar businesses on this framework.

The parallel isn’t perfect, but it’s closer than it first appears. In both cases, a third party sits between the organisation and the employment relationship. In both cases, the third party handles the mechanics of employment. The difference is that with an EOR, the worker is a human being with legal rights, tax obligations, and a pension contribution. With an AI agent, there is no worker. There is only the work.

I explored a version of this trajectory back in 2024 when I wrote about the rise of AI-crafted digital twins for the TIC Digest. That piece examined what happens when AI can construct an autonomous replica of a professional’s working persona, one capable of attending meetings, drafting documents, and executing tasks across multiple organisations simultaneously. I raised the question of whether such digital twins might eventually demand workplace protections, stake claims to earnings, or even collectively bargain. At the time it felt speculative. Two years on, with Anthropic actively building “people products” and AI agents executing end-to-end HR workflows, the distance between speculation and reality has narrowed considerably. The question I posed then about who is “clocked in” when a digital twin is working is essentially the same question we now face with agentic AI: who is the employer, and who pays the tax?

And that’s where the fiscal problem begins.

The Fiscal Black Hole

Modern tax systems are built on a simple foundation: people work, they earn money, and governments tax that income. In the US, roughly three-quarters of all federal tax revenue comes from labour income. In the UK, income tax and National Insurance contributions together represent the single largest source of government revenue. Social security systems, pension pots, and public services across the developed world are funded on the assumption that most working-age adults will be employed and paying into the system.

AI threatens to hollow out that assumption. Not overnight, and not uniformly, but progressively and at scale.

Consider a straightforward example. A UK employer hires a recruiter at £45,000 a year. The employer pays employer’s National Insurance at 15% (roughly £5,500). The employee pays employee’s NI and income tax (roughly £9,500 combined). The Treasury receives approximately £15,000 a year from that single employment relationship. The employee also contributes to a workplace pension, building the savings that will eventually reduce their dependence on the state pension.

Now replace that recruiter with an AI agent. The employer pays a software licence fee to an AI lab. No income tax. No National Insurance. No pension contribution. No employer NI. The Treasury loses £15,000 a year. The pension system loses a contributor. The employee joins the unemployment figures, potentially claiming benefits rather than paying taxes. The fiscal swing from a single role replacement is closer to £25,000 than £15,000 when you factor in the benefits cost.

Multiply that across hundreds of thousands of roles and the numbers become genuinely alarming. And this isn’t a problem unique to recruitment. The same dynamic applies to any function where AI agents can execute work autonomously: customer service, content production, data analysis, financial processing, legal research, procurement. Every role displaced is a tax contribution lost.

OpenAI itself has acknowledged this. In April 2026, the company released a set of policy proposals that included public wealth funds, expanded social safety nets, and what many have termed a “robot tax,” proposing levies on automated labour to capture a share of the productivity gains that would otherwise flow exclusively to capital owners. The company explicitly warned that as AI automates more work, the wage and payroll tax revenue that funds Social Security, Medicaid, and housing assistance could collapse.

When the companies building the technology are telling governments the tax base is at risk, it’s probably time to listen.

This isn’t a new concern for those of us watching the macro picture. In early 2025, I wrote about the case for strategic state intervention in an automated economy, arguing that shareholder capitalism’s relentless pursuit of efficiency through automation risks destroying its own consumer base. The mathematics of that piece still hold: when machines displace workers faster than new roles emerge, and when automation’s gains flow solely to shareholders, the economic model eats itself. What’s changed since then is the speed. The automation I was writing about in general terms is now arriving in specific, named products from specific, named companies. The AI-EOR concept outlined in this article is, in a sense, a more commercially palatable companion to the strategic nationalisation argument I made then. Both address the same underlying problem. This one offers a mechanism that works within existing market structures rather than against them.

The Demographic Amplifier

The fiscal challenge from AI-driven labour displacement doesn’t arrive in isolation. It lands on top of a demographic crisis that is already well underway.

Across the OECD, populations are ageing. Birth rates are falling. Dependency ratios are climbing. The number of working-age adults supporting each retiree is shrinking in virtually every developed economy. Japan is the most cited example, but the UK, Germany, South Korea, Italy, and increasingly the United States face the same structural pressure.

Pension systems designed around the assumption that each generation of workers would be larger than the last are already strained. Add AI-driven job displacement into this picture and you get a triple squeeze: fewer workers, more retirees, and AI agents doing work that used to generate the tax revenue needed to fund the gap.

This isn’t a problem that can be solved by retraining alone. Retraining assumes there are roles to retrain into. If AI agents continue to expand their capability envelope, absorbing not just routine tasks but increasingly complex professional work, the supply of roles that require human labour may contract faster than new categories of work emerge. That’s not a certainty, but it’s a scenario that responsible workforce planning needs to account for.

Could AI Labs Become Employers of Record for Agents?

Here’s the provocation at the heart of this article. Rather than inventing entirely new tax frameworks from scratch, what if governments adapted an existing one?

The Employer of Record model already provides a legal and commercial framework for situations where a third party sits between an organisation and an employment relationship. It handles payroll, tax, compliance, and benefits administration. Governments already understand it. Tax authorities already regulate it. Companies already use it.

What if AI labs were required to act as Employers of Record for their agents?

The concept would work something like this. When Company A deploys an AI agent from Lab X to perform work that would otherwise be done by a human employee, Lab X becomes the “Employer of Record” for that agent. Lab X pays an equivalent of payroll tax or National Insurance based on the economic value of the labour displaced. The tax revenue flows to the Treasury in the same way it would if a human were doing the work. The pension contribution gap is plugged, at least partially, through a mandatory equivalent contribution to a public fund.

There are several reasons this framework has appeal.

First, AI labs already have the telemetry. They know exactly how many agents are deployed, what tasks those agents perform, how many API calls are made, and in many cases they can estimate the human-equivalent labour hours being displaced. The data infrastructure for measurement already exists, which is more than can be said for most proposed robot tax schemes.

Second, it maps onto existing legal and regulatory architecture. Governments don’t need to invent an entirely new category of taxation. They can extend and adapt a framework that already has established precedents, case law, and compliance infrastructure.

Third, it creates a commercial incentive for honest pricing. If AI labs bear an employment-equivalent tax burden for their agents, the cost of deploying an AI agent begins to reflect its true economic impact rather than simply the marginal cost of compute. This doesn’t eliminate the cost advantage of automation, but it narrows the gap to a level that reflects genuine productivity gains rather than tax arbitrage.

Fourth, and perhaps most importantly, the legal groundwork is already being laid. This one deserves its own section.

The Legal Bridge: Mobley v. Workday

Mobley v. Workday, Inc. is not an EOR case. It’s an employment discrimination lawsuit. But the legal principle it is establishing may turn out to be the most important building block for any future AI-EOR framework.

The facts are straightforward. Derek Mobley, an IT professional and African American man over 40 with a disability, applied for more than 100 positions through companies using Workday’s AI-powered applicant screening tools. He was rejected every time. When he received a rejection email at 1:50 in the morning, less than an hour after submitting his application, he realised no human being had ever looked at his CV. The decision to reject him had been made entirely by software.

Mobley sued Workday, not the employers. His argument was that Workday’s AI screening tools systematically discriminated against older applicants, candidates with disabilities, and Black applicants. Workday’s defence was what you’d expect from any software vendor: we just provide the tool, the customer makes the decision.

The court disagreed. In July 2024, Judge Rita Lin of the Northern District of California ruled that Workday could face direct liability under an “agency” theory. The court held that Workday’s software was not simply implementing employer criteria in a rote fashion but was actively participating in the decision-making process, recommending some candidates and rejecting others. The court drew a clear distinction between Workday’s role and that of a simple spreadsheet or email tool. The degree of automation and decision-making authority mattered.

The case has accelerated since then. In May 2025, the court granted preliminary certification of a nationwide collective action under the Age Discrimination in Employment Act. In early 2026, formal opt-in notices went out, inviting thousands of applicants to join. In March 2026, Judge Lin rejected Workday’s attempt to have the ADEA claims dismissed, ruling that the law covers job applicants as well as employees. The EEOC filed an amicus brief supporting the plaintiff’s theories of direct AI vendor liability. Legal commentators now describe Mobley as one of the most significant AI employment cases in progress, with a decision on the merits potentially arriving later this year.

Why does a discrimination case matter for the AI-EOR argument? Because of the legal principle at its core. The court has accepted that when a company delegates employment decisions to an AI vendor’s software, that vendor can be treated as an agent of the employer. It is not just a tool provider. It is a participant in the employment relationship with corresponding legal obligations.

That is a profound shift. If an AI vendor can be held liable as an agent for discriminatory outcomes, the logical extension is that an AI vendor (or AI lab) can bear other employer-equivalent obligations too, including fiscal ones. The jump from “Workday is an agent of the employer for discrimination liability purposes” to “Anthropic is the Employer of Record for its AI agents for taxation purposes” is not as large as it first appears. Both rest on the same foundational idea: when software autonomously executes functions that used to be performed by humans in an employment context, the provider of that software bears a share of the obligations that come with it.

The case also has a delicious irony in the context of this article. Workday, the company whose AI tools are being challenged as a de facto employment agent, has just lost its CTO to Anthropic, the company that is building the next generation of AI-powered HR products. The legal framework being established around Workday’s current products may end up applying to the products Anthropic is building with the expertise of Workday’s former technical leader.

Why It’s Hard

The AI-EOR concept is deliberately provocative, and there are genuine obstacles.

Defining “human-equivalent work displaced” is the most obvious challenge. Not every AI deployment replaces a full-time employee. Many augment existing workers, making them more productive rather than making them redundant. Drawing a clear line between augmentation and displacement is genuinely difficult, and getting the measurement wrong in either direction creates problems. Too aggressive and you penalise productivity gains. Too lenient and you miss the tax revenue.

Jurisdictional complexity is another headache. Where is the agent “employed”? Is it where the AI lab is headquartered? Where the client company operates? Where the displaced worker would have been based? In a world of cloud-based AI services deployed globally, the answer could be all three simultaneously. This mirrors the challenges that already exist with digital services taxation, but adds a labour market dimension that makes it even more complex.

AI labs would resist the classification, and with good reason from their perspective. Being designated as an employer, even a fictional one for tax purposes, carries liability implications that extend well beyond taxation. Employment law, discrimination law, health and safety obligations, pension regulation: the legal surface area of being an “employer” is vast, and AI labs would argue (not unreasonably) that applying it to software agents stretches the concept beyond recognition.

There’s also the innovation argument. Any tax on AI deployment risks slowing adoption, and governments that move first may simply push AI activity to jurisdictions that don’t. This is the standard objection to any form of technology taxation, and it has some validity. The counter-argument is that a well-designed levy captures value without killing the value creation, and that the cost of inaction (a collapsing tax base, an unfunded pension crisis, rising inequality) is far higher than the cost of slightly slower AI adoption.

The Policy Landscape Is Moving Fast

Whatever you think of the AI-EOR concept specifically, the broader policy conversation is accelerating.

OpenAI’s April 2026 policy paper called for shifting the tax base away from payroll and labour income and toward corporate income and capital gains, alongside the “robot tax” proposals. Andrew Yang has been arguing publicly that the US should stop taxing labour entirely and shift the burden to AI companies. Dario Amodei, CEO of Anthropic, has himself advocated for AI-related taxes. Brookings has published detailed proposals for token taxes on AI-generated content and robot services taxes on automated services. And serious academic work is underway on the question of whether AI systems should be granted a form of legal personality to create a structured basis for taxation and accountability.

The direction of travel is clear. The question is whether governments will land on something coherent or end up with a patchwork of conflicting national approaches that creates compliance chaos for multinational employers and AI providers alike.

What This Means for Talent Intelligence

If you work in talent intelligence, workforce planning, or talent acquisition leadership, this isn’t a distant policy debate. It has practical implications for how you model workforce costs, plan headcount, and advise your business.

Workforce planning models need to evolve. Today, most models treat AI deployment as a cost reduction: fewer FTEs, lower headcount cost, improved productivity ratios. That’s accurate as far as it goes, but it ignores the potential for AI-related levies or agent taxes that could change the cost equation. If governments introduce an AI-EOR levy or equivalent, the total cost of an AI agent stops being just the licence fee and becomes the licence fee plus the tax equivalent. Planning models that don’t account for that scenario are incomplete.

Total cost of workforce calculations need updating. The traditional TCOW model includes salary, benefits, employer taxes, overhead, and contingent labour costs. A future version may need to include “AI agent deployment costs” as a distinct category, with its own tax and compliance implications. For organisations deploying agents at scale, this could be a material budget line.

Competitive intelligence should be tracking which AI labs are building toward autonomous HR execution. Anthropic is the most explicit right now, but OpenAI, Google, and Microsoft are all investing in adjacent capabilities. Understanding which platforms are likely to become employment infrastructure, rather than just productivity tools, is a strategic intelligence question.

Scenario planning for regulatory divergence is essential. The EU, UK, US, and major Asian economies will approach AI taxation differently, creating both arbitrage opportunities and compliance complexity. Organisations operating across multiple jurisdictions will need talent intelligence teams that can model the workforce cost implications of different regulatory scenarios.

And perhaps most importantly, the role of talent intelligence itself is changing. This kind of analysis, connecting labour market dynamics, fiscal policy, technology adoption, and competitive strategy, is exactly the kind of strategic advisory work that positions TI teams as essential partners to the CFO and the board. Not sourcing support. Not recruitment analytics. Strategic workforce intelligence that shapes capital allocation decisions.

Back to the CTO

When a Workday CTO leaves an $8 billion enterprise software company to build HR products at a frontier AI lab, it’s tempting to read it as just another executive move in a hot market. But talent signals are only useful if you read them in context.

The context here is an AI industry that is moving from building tools for HR to building systems that replace HR. An AI lab that is openly recruiting enterprise software expertise to compete for HR budgets. A policy landscape where the companies building the technology are themselves calling for new tax frameworks to address the labour displacement they expect to cause. And a fiscal reality where the tax systems that fund public services, pensions, and social safety nets are built on a foundation of human employment that AI is actively eroding.

The question isn’t whether AI agents will do work that used to generate tax revenue. They already do. The question is whether governments will figure out how to capture that value before the fiscal hole becomes unmanageable. The EOR model might not be the final answer, but it offers something that most robot tax proposals lack: a framework that already exists, that governments already understand, and that maps onto the actual commercial relationship between AI labs, their clients, and the work being performed.

Talent intelligence practitioners should be watching this space closely. Because when the tax treatment of AI agents changes, and it will change, the workforce planning models, cost structures, and competitive dynamics that underpin our work will change with it.