Agentic commerce may remove the checkout from sight, but it cannot remove the transaction from tax law.
The checkout is disappearing, but the tax event is not
Meta’s Muse is interesting for retail because it moves artificial intelligence one step beyond search, recommendation and conversation. Meta introduced Muse on September 8 as a personal AI agent that can act on behalf of a user across the web, and the company explicitly includes purchasing among the sensitive actions for which Muse asks for user approval. Stripe has integrated Link so that, for U.S. consumers, Muse can pay at more than one million businesses that accept Link and can use a purchase-scoped single-use virtual card at other merchants. Shopify, meanwhile, has created agentic storefronts through which eligible merchants can make products discoverable in AI channels and, on supported Meta surfaces such as Muse, allow customers to complete a Shopify-powered checkout without leaving the Meta experience.
That sounds like a revolution in shopping, and in many respects it is, but from a VAT and fiscalization perspective the most important observation is almost the opposite. An AI agent can become borderless, a checkout can disappear from the customer’s field of view, and the commercial journey can move from a retailer’s website into somebody else’s interface, yet the underlying supply still has a seller, a buyer, a product, a tax status, a place of supply, a payment, an invoice or receipt requirement, and in many countries a fiscal reporting obligation. Tax law does not disappear because the interface becomes conversational.
When we wrote our white paper “Fiscal Compliance in the Age of Agentic Commerce,” we separated the relatively simple case of an AI agent acting as a shopper from the more consequential cases in which an agent starts selling, brokering, negotiating or orchestrating transactions. The first real-world implementations of agentic commerce are now making that distinction much more concrete. The purchasing side is important, but the selling side is where the tax architecture begins to change.
The buying agent is usually the easier case
If a consumer in Germany asks an AI agent to buy a pair of shoes from an established retailer and the agent merely navigates the retailer’s existing checkout, enters authorized information and executes the purchase after approval, the agent has not automatically become the buyer, reseller or merchant of record. The commercial supply normally remains the retailer’s sale to the consumer, while the agent is a tool acting on the consumer’s behalf. The retailer’s existing tax and fiscal systems still have to determine the nature of the supply, the VAT treatment, the invoice or receipt requirement and any local reporting obligation.
This distinction matters because it prevents us from looking for a new tax rule simply because a new technology executes the click. In EU VAT, the place of taxation for intra-Community distance sales of goods is generally the Member State where dispatch or transport to the customer ends, subject to the EUR 10,000 cross-border threshold rules for qualifying suppliers and the possibility of using the One Stop Shop. The European Commission’s guidance is explicit that the destination of the goods, rather than the location of the browser or the software used to place the order, drives the basic place-of-taxation analysis for these supplies.
The same reasoning also shows why a sentence such as “the agent needs to know which country the buyer is in” is directionally useful but legally too simple. For a physical cross-border B2C sale, the relevant fact is often the destination of the goods; for particular services, the customer’s establishment, permanent address or usual residence can become decisive; for B2B transactions, the identity and VAT status of the legal entity and the place-of-supply rules for that category matter. An agent’s IP address, cloud location or the country in which the model happens to run may have little or no relevance to the VAT result.
Direct checkout changes the compliance architecture
The situation becomes more interesting when the purchase is completed inside the AI environment rather than on the retailer’s own website. Shopify’s current Meta integration gives us an unusually clear example. Shopify states that eligible customers can complete a Shopify-powered direct checkout on Meta surfaces, including Muse when available, without leaving the experience, while the merchant remains the merchant of record and continues to manage fulfillment, returns and customer service. Shopify also states that Meta receives order information including customer contact and physical address data for orders placed through direct checkout, while standard client-side analytics pixels do not fire because the checkout occurs on Meta rather than in the merchant’s online store.
There is one important geographic limitation that should not be overlooked in the current hype. Shopify’s Meta direct checkout is presently described as available to customers based in the United States, Canada and Mexico, and Stripe’s Muse integration is announced for U.S. consumers. The EU VAT problems discussed here are therefore not a description of today’s European Muse checkout flow; they are an analysis of the architecture that becomes relevant as agentic commerce expands into Europe and other multi-jurisdictional markets.
The architecture already tells us what will matter. A retailer can remain legally responsible for the sale while the interface that gathers customer intent, presents the offer and executes checkout belongs to another platform. That creates a new division between control of the experience and responsibility for compliance. The merchant still needs enough reliable transaction context to calculate tax, issue the correct document, determine whether a fiscal receipt must be generated, create an audit trail and report the transaction where required, even though a material part of the customer journey may have taken place elsewhere.
The agent has no tax country; the transaction does
Agentic commerce therefore needs a concept that conventional e-commerce could often hide inside a checkout page: explicit, machine-readable transaction context. A tax engine does not merely need a country field. It needs to know who is selling, who is buying, whether the buyer is acting as a consumer or a taxable business, where the goods start their journey, where they end it, what is being supplied, whether the seller has relevant establishments, which special scheme is being used, and which events have already occurred in the commercial flow.
A German retailer selling electronics to a French consumer illustrates the point. Once the EU distance-selling destination principle applies, French VAT becomes relevant because the goods are transported to France, while the EUR 10,000 threshold can keep qualifying small cross-border sellers within the supplier Member State’s VAT rules until the conditions cease to be met or the supplier opts into the destination principle. OSS can then simplify declaration and payment across Member States, but it does not remove the need to classify the transaction correctly before the sale is completed.[4][5]
An agent comparing several sellers cannot wait until the last millisecond of checkout to discover these facts if it wants to compare genuine final prices. A product shown at EUR 98 by one seller is not necessarily cheaper than a product shown at EUR 100 by another when VAT, shipping, customs, duties, product classification and the origin of the goods differ. The more an agent promises to optimize purchasing, the more tax determination moves upstream from an invisible checkout calculation into the decision-making process itself.
Tax-aware shopping will become part of shopping intelligence
This changes the role of tax technology in a subtle but important way. Traditional e-commerce frequently treats tax as a calculation invoked once the basket and destination are known. Agentic commerce can require a provisional tax determination while the agent is still searching, filtering and ranking alternatives, followed by deterministic recalculation as more facts become available. The white paper we published earlier anticipated this shift by arguing that VAT logic would have to become embedded in the transaction flow rather than remain a back-office step, with structured tax data and real-time interfaces supplying the facts needed for compliant decisions.[12]
That does not mean the large language model should become the tax engine. In fact, the opposite is safer. A generative agent is good at understanding intent, extracting context and orchestrating actions, but tax outcomes should be derived from authoritative rules, validated transaction facts and deterministic calculation services, with the AI receiving the result and the reasons behind it. The OECD’s work on tax-administration digitalisation points in the same general direction: digital identity, APIs, e-invoicing, machine-readable rules and AI are becoming parts of increasingly integrated tax systems, while governance, objectivity, privacy and trust remain central requirements.[10][11]
For retail technology providers, the strategic implication is that “shopping intelligence” will eventually need a compliance layer. An agent that can recommend the best product but cannot determine the legally correct total price, invoice treatment or fiscal path is not yet a complete commerce agent.
The deemed-supplier boundary deserves particular attention
A second issue becomes more important as AI platforms move deeper into the sale. EU VAT already contains rules under which an electronic interface can, in defined circumstances, be treated as if it had received and supplied the goods itself. Article 14a of the VAT Directive applies deemed-supplier treatment to specified transactions facilitated through an electronic interface, including certain imported distance sales and certain supplies within the EU made by non-EU suppliers. The implementing rules also define “facilitates” and look at factors such as involvement in the terms and conditions, authorization of payment, ordering and delivery.[7][8]
This should not be misread as a conclusion that Meta or Muse is currently a deemed supplier in the EU. The present Shopify implementation expressly keeps the merchant as merchant of record, its direct checkout is currently limited geographically, and the deemed-supplier provisions have specific legal conditions that require transaction-by-transaction analysis. The important point is architectural: as an AI platform moves from recommendation to ordering, payment authorization, transaction terms, supplier selection, bundling or fulfillment coordination, the tax classification of its role becomes more significant, not less.
The next generation of agentic commerce may therefore create a spectrum rather than a clean line between “assistant” and “seller.” At one end sits an agent that merely navigates a merchant’s existing site. In the middle sits an agentic channel that provides discovery and checkout while the underlying merchant remains responsible for the supply. At the other end lies a platform that shapes the commercial terms so extensively that existing marketplace and deemed-supplier concepts, or future rules built for autonomous commerce, become increasingly relevant.
Fiscalization may be harder than the VAT rate
VAT is only part of the problem because fiscalization is often more operational and more country-specific. A VAT engine can determine the applicable tax treatment, while a fiscal system may additionally need to determine when a legally relevant sale has occurred, which entity must issue the receipt or invoice, which numbering sequence applies, whether the transaction must be signed, which data must be stored, which authority must receive it and how corrections, cancellations or refunds must later be linked to the original transaction.
Agentic commerce complicates this because commercial actions that were once visibly separated by screens can become machine-to-machine events executed in seconds. The consumer may approve an offer, the agent may submit the order, a merchant may accept it, payment may be authorized, stock may be reserved, an invoice may be created and fulfillment may start almost immediately, yet different legal systems can attach fiscal consequences to different moments. A robust architecture therefore needs a canonical event model that can distinguish intent, offer, acceptance, order submission, order acceptance, payment authorization, supply, invoicing and fiscal issuance instead of collapsing everything into a generic “purchase completed” signal.
This is where fiscal middleware becomes much more than an integration component. It becomes the translation layer between the language of an AI agent, which deals in intentions such as buy, compare, reserve, ship and return, and the language of law, which deals in taxable supplies, invoice requirements, receipt issuance, sequencing, signatures, reporting and retention. Our earlier white paper described precisely this interoperability gap and proposed fiscal middleware as the neutral bridge between autonomous commercial behavior and national compliance systems.[12]
B2B agentic commerce will raise the stakes again
Most consumer excitement around Muse concerns personal shopping, but B2B purchasing will eventually create an even more demanding compliance problem. Shopify’s current Meta direct-checkout documentation explicitly excludes B2B-only products, which is a useful reminder that the simplest consumer use cases are only the first layer of the market.
A B2B agent cannot merely know that somebody wants to buy. It needs to know which legal entity is purchasing, whether the person or agent has authority to bind that entity, which VAT identification number is relevant, whether that number is valid, where the supplier and customer are established, whether reverse charge or another treatment applies, what invoice content is mandatory and which digital reporting obligations are triggered. These are not conversational details; they are legally relevant transaction facts that must be authenticated, preserved and made auditable.
This is also where digital identity and authorization become essential. The OECD’s 2025 report on tax-administration digitalisation notes that most participating tax administrations already use digital identity for access to online services and highlights identity as a prerequisite for reliable system-to-system interaction and real-time data exchange. Agentic commerce will need an equivalent discipline on the commercial side, because an autonomous instruction is only useful for compliance when the system can establish who the agent represents and under which authority it is acting.
Europe is moving toward the same technical direction from the regulatory side
The timing is significant because European VAT administration is already moving toward more structured, digital and machine-readable processes. The VAT in the Digital Age package was adopted in March 2025 and is being introduced progressively through 2035. Among other measures, it enables wider use of mandatory e-invoicing, introduces changes to OSS and IOSS, and from July 2030 brings digital reporting requirements for cross-border B2B transactions based on e-invoicing, with national real-time reporting systems required to converge with the EU model by 2035.
ViDA was not designed specifically for Muse or agentic commerce, but the direction is compatible with what autonomous commerce requires. A world in which machines initiate, negotiate and execute transactions cannot scale on tax processes that depend on a human reading a PDF, interpreting an unstructured rule and correcting the result weeks later. The more commerce becomes machine-to-machine, the more compliance also needs structured identities, structured events, structured invoices, deterministic rules and verifiable audit trails.
The OECD reports the same broader transformation among tax administrations, where e-filing, e-payment, APIs, artificial intelligence and digital identity have become increasingly mainstream, while more than 70 percent of the administrations covered by its 2025 digitalisation report said they were using AI in some form. The technology on the authority side is therefore developing at the same time as the commercial side is beginning to delegate transactions to agents.
What retailers should build before agentic commerce arrives at scale
Retailers do not need to replace their tax stack because Muse exists, and rushing to put VAT logic directly inside an AI agent would be the wrong response. They do, however, need to make their existing compliance capabilities callable, explainable and channel-independent. Product tax classification, destination logic, customer tax status, VAT calculation, invoice generation, fiscalization and reporting should increasingly be exposed as governed services that can be invoked from a website, mobile app, marketplace, point of sale or autonomous agent without duplicating legal logic inside every channel.
The transaction record should also preserve the provenance of decisions. If an agent selected a product, if the customer approved a purchase, if a platform supplied the delivery address and if a tax engine applied a destination VAT rate, the retailer should be able to reconstruct those facts later. This is not only an audit requirement; it is the basis for correcting transactions, resolving disputes and proving why a particular tax treatment was applied when the original customer journey may have happened outside the retailer’s own interface.
The same principle applies to technology vendors. The valuable component will not be an AI model that claims to “know VAT,” but a compliance service capable of receiving a normalized transaction context, applying authoritative rules, returning a deterministic result, documenting the legal basis and triggering the appropriate local fiscal process. In practical terms, the future architecture looks less like an intelligent chatbot connected directly to a tax authority and more like an intelligent agent connected to a controlled compliance layer that mediates between commercial intent and legal execution.
The real change is not who clicks the button
Muse makes agentic commerce visible because consumers can finally watch an agent move from conversation to action, but the deepest retail impact will emerge when that action becomes a mainstream sales channel. The change is not that software can click “buy.” E-commerce systems have automated actions for decades. The change is that a third-party agent can increasingly own discovery, selection and checkout while the retailer continues to carry significant obligations for the legally relevant transaction.
For tax and fiscalization, that produces a simple principle with far-reaching consequences: the AI agent may be borderless, but the transaction is not. Every autonomous purchase still needs a legally meaningful identity, destination, classification, value, time, supplier and audit trail, and every seller-side agent needs a reliable way to translate commercial intent into the fiscal language of the jurisdiction in which the transaction becomes taxable.
That is why I would slightly update the line we used in our earlier white paper. We wrote that if an AI can buy or sell, it must also fiscalize. The more precise formulation now is that if an AI can sell, the transaction must know how to fiscalize itself. The agentic-commerce architecture should be able to make the legal transaction automatic without making it any less correct, traceable or accountable.
