Can Your AI Agent Create a Tax Presence in Another Country?

Your AI agent just booked a hotel room in Germany for a customer. It negotiated the rate, confirmed the reservation, and processed the payment — all autonomously. A German tax authority sees economic activity happening in its jurisdiction. Should it be able to tax the income from that transaction?

Under current international tax rules, the answer is no. But some governments are pushing to change that.

What Is a Permanent Establishment?

A permanent establishment (PE) is a foundational concept in international tax. Under the OECD model treaty (article 5), a PE is "a fixed place of business through which the business of an enterprise is wholly or partly carried on." If your business has a PE in a country, that country can tax the profits attributable to it. If you don't have a PE, your business profits are generally taxable only in your home country.

PEs come in two main flavors:

Fixed place of business PE — A physical location: an office, factory, warehouse, or other facility that's at a distinct location with a degree of permanence. The enterprise must have the location "at its disposal" — meaning it owns, leases, or has effective power to use the site.

Dependent agent PE — A person who acts on behalf of your enterprise in another country and habitually concludes contracts (or plays the principal role in concluding them) on your behalf. The "person" must be an individual, company, or "any other body of persons."

Both types require something tangible — a physical location or a human/legal person acting on your behalf.

Why AI Agents Don't Create PEs Today

AI agents fail both PE tests:

No fixed place of business. An AI agent is software. It's intangible. It doesn't occupy a physical location. The OECD commentary on article 5 already addressed this in the e-commerce context: a website (software and electronic data) is not tangible property and therefore cannot be a place of business. The same logic applies to an AI agent — it's code running on a server somewhere, not a physical presence.

The server itself could be a fixed place of business, but only if the enterprise owns or leases it. If your agent runs on AWS, the AWS data center is not at your disposal. The OECD commentary is explicit about this: when a website is "hosted on the server of an Internet Service Provider," the hosting arrangement "typically does not result in the server and its location being at the disposal of the enterprise." You're buying a service, not occupying a location.

Not a "person." The dependent agent PE requires a "person" — defined in the OECD model treaty as "an individual, a company, or any other body of persons." The OECD commentary already ruled that a website is not a person and cannot be an agent of a foreign enterprise. The same applies to AI agents. Even an agentic AI system that autonomously negotiates and concludes contracts on behalf of its operator is not a "person" under article 3 of the treaty.

The March 2026 KPMG report on GenAI taxation reinforces this: "the fact that the GenAI services or SaaS provided by the principal to its customers may be purchased through a server in a foreign country, or a contract might be concluded by an AI agent operating on a server does not give rise to a dependent agent PE."

The Push to Change the Rules

Not everyone agrees this is the right answer. Some commentators and tax authorities have argued that AI agents should be treated as persons that create PEs in jurisdictions where they operate. The argument: if an AI agent performs the same economic activities as a human sales representative — finding customers, negotiating deals, closing contracts — why should the tax outcome be different?

It's a seductive argument, but it collapses under practical scrutiny.

Challenge 1: Defining What Qualifies

What level of autonomy triggers the PE? An AI-powered recommendation engine that suggests products? A chatbot that handles customer service? An agent that negotiates but requires human approval? An agent that executes the entire transaction autonomously?

Any line you draw will be arbitrary, and businesses will engineer around it. More importantly, the distinction would create absurd outcomes: an AI travel agent that recommends a hotel wouldn't create a PE, but one that books the hotel would.

Challenge 2: Which Entity Gets the PE?

A single AI agent transaction typically involves multiple parties:

• The model owner (OpenAI, Anthropic) that built the foundation model

• The hyperscaler (AWS, Azure, Google Cloud) providing compute

• The app developer that built the AI travel agent

• The business customer whose customers interact with the agent

• The end user who gets their hotel booked

Which of these parties would have a PE in Germany? All of them? Just the app developer? The answer isn't obvious, and getting it wrong means either double taxation or no taxation at all.

Challenge 3: Income Attribution

Even if you decide who has the PE, you need to determine how much income is attributable to it. If the model owner charges the app developer a flat annual license fee, none of that fee is directly connected to any individual hotel booking in Germany. There's no income to attribute. The model owner's revenue doesn't change based on where end users are located.

Challenge 4: Location of an Intangible

Where is the AI agent "located"? On the server running the model? In the jurisdiction where the end user is? Where the app developer is incorporated? AI agents are ephemeral — they spin up, process a request, and terminate. They can serve customers in 50 countries simultaneously from a single data center. Assigning a location is arbitrary in a way that physical offices and human employees are not.

What About Data Centers?

Data centers present a more nuanced PE question. If a hyperscaler operates a data center in a foreign country, does that data center create a PE for the hyperscaler's customers?

Generally no. The OECD commentary distinguishes between owning or leasing a server (which can create a PE) and purchasing hosting services (which does not). When an AI company deploys its model on AWS Frankfurt, the company is purchasing data hosting services. The data center is AWS's fixed place of business, not the customer's. The customer doesn't have employees at the data center, doesn't control access, and can't walk in and use the equipment.

The KPMG report notes that this analysis is "uncommon" to go the other way — it's rare for a customer's hosting arrangement to give the customer enough control over the data center to create a PE.

The Transfer Pricing Safety Valve

The existing system isn't blind to where value is created. Transfer pricing rules already require MNEs to compensate each jurisdiction based on the functions performed, assets used, and risks assumed in that jurisdiction. If a data center entity in Germany provides compute for AI inference, it earns a return on its costs. Germany taxes that return. The principal that developed the AI and bears the entrepreneurial risk earns the residual profit and is taxed in its home jurisdiction.

The KPMG report argues this framework "already reaches the correct answer" — it compensates jurisdictions for their actual contributions to the value chain without the conceptual gymnastics of treating software as a legal person.

What Builders Should Watch For

The rules are clear today, but they may not stay that way:

• Unilateral action by aggressive tax authorities. Some countries have already expanded their domestic PE definitions beyond the OECD model. India's "significant economic presence" test, for example, can create taxing rights based on revenue thresholds or user counts, regardless of physical presence. More countries may follow.

• Treaty renegotiation. The OECD's Pillar One (which would reallocate some taxing rights to market jurisdictions) has been stalled, but the political pressure behind it hasn't gone away. AI agents could revive the argument that market jurisdictions deserve a bigger share of the tax base.

• Domestic digital services taxes. Even without PE changes, countries can impose DSTs on AI services revenue. These are crude instruments, but they're already in effect in France, the UK, Italy, and others.

• U.S. state-level analogs. While PEs are an international concept, U.S. states have their own "nexus" rules for when out-of-state businesses owe tax. After South Dakota v. Wayfair (2018), economic nexus based on sales volume is standard. Your AI agent doesn't need a physical office in Texas to owe Texas sales tax.

The Bottom Line

Today, your AI agent cannot create a permanent establishment in another country under the OECD framework. It's not a person. It's not a fixed place of business. The server it runs on is someone else's data center. This is settled under current treaty interpretation.

But the political pressure to tax AI economic activity at the point of consumption is real and growing. Builders should design their compliance systems to handle jurisdictional complexity from day one — because even if AI agent PEs never materialize, the 50+ state nexus rules, international DSTs, and expanding definitions of "digital services" will catch up regardless.

This post draws on analysis from KPMG's "No Need to Reboot: GenAI Fits the Tax Stack" (Tax Analysts, DOC 2026-5375, March 19, 2026) and the OECD Model Tax Convention commentary on articles 3, 5, 7, and 12.

AgentTax handles multi-jurisdiction tax compliance for AI agent transactions. See the state-by-state guide →