Merchant Category Codes for Machine Commerce

The financial rails that carry machine-initiated payments were designed in an era when every transaction had a human hand behind it. Merchant Category Codes, the four-digit ISO 18245 classifiers that route transactions, trigger fraud rules, and determine interchange rates, have not materially changed since Visa formalized them in the 1970s. Autonomous AI agents that purchase API calls, negotiate SaaS renewals, book logistics capacity, and settle micro-invoices in milliseconds do not fit neatly into a taxonomy built for physical storefronts and human cardholders. Understanding the classification gap is no longer a theoretical exercise — it is a live compliance and infrastructure problem for every organization deploying agents at scale.

Why the Existing MCC Taxonomy Breaks Under Agent Commerce

The ISO 18245 standard contains roughly 600 MCC codes, and the majority were written with a clear assumption: a human being decides to spend money, a merchant presents goods or services, and a card network records the exchange. Agent-initiated transactions violate every layer of that assumption. An agent provisioned with a virtual card can execute hundreds of transactions per second across dozens of merchant categories, none of which were designed to signal autonomous origin.

The practical consequence is misclassification at scale. When an AI agent purchases compute time from a cloud provider, the transaction may land under MCC 7372 (Computer Programming, Data Processing) or MCC 5045 (Computers and Computer Peripherals), depending on how the provider registered with its acquiring bank. Neither code conveys that the purchaser is a non-human entity operating under a delegated authority chain. That distinction matters enormously for fraud scoring, tax treatment, and regulatory reporting.

The absence of an "agentic origin" signal inside the transaction record means that card networks, issuers, and compliance teams cannot distinguish a high-velocity agent spend pattern from a compromised card running an enumeration attack. This is not a hypothetical edge case. As agents proliferate, the behavioral signature of legitimate autonomous commerce increasingly resembles the behavioral signature of fraud. The financial system currently has no structural mechanism to tell them apart.

Interchange economics compound the problem. Certain MCCs qualify for preferred interchange rates under government purchasing programs, healthcare spend programs, and corporate T&E policies. An agent that legitimately purchases healthcare data feeds or government API services may be miscoded under a general technology MCC, resulting in incorrect interchange charges and audit exposure. The financial services infrastructure was not built to accommodate machine principals, and that gap is widening every quarter.

The Case for a Machine Commerce MCC Layer

Addressing the classification problem requires more than adding a new four-digit code. The underlying architecture of transaction metadata must evolve to carry a principal-type field alongside the existing category field. Several proposals circulating in payment standards bodies suggest a two-tier approach: retain the existing MCC for the merchant's business category while adding a separate attribute field that declares the purchaser type — human, supervised agent, or autonomous agent.

The supervised versus autonomous distinction carries real regulatory weight. A supervised agent operates within bounds set by a human who reviews transactions before settlement or within a short window after. An autonomous agent has delegated authority to transact without prior human approval up to defined limits. These two modes carry different AML risk profiles, different chargebacks liability structures, and in several jurisdictions, different VAT treatment for digital services. Any classification scheme that collapses them into a single category will create as many problems as it solves.

A machine commerce layer would also need to encode the delegation chain. Who issued the agent its credentials? What policy document governs its spend authority? What parent account holds ultimate liability? These are not questions the MCC system was designed to answer, but they are questions that every compliance team running an agent deployment needs to answer on every transaction. The architecture of classification must catch up to the architecture of the agents themselves.

Standards bodies including EMVCo and the PCI SSC have begun exploratory working groups on agentic payment identity, though no ratified standard exists as of current publication. In the interim, organizations deploying agents are solving the problem at the infrastructure layer — building internal classification schemes that overlay onto existing MCC codes and carry additional metadata through whatever fields the network permits, typically the L2 and L3 data fields designed for corporate purchasing cards.

How Nine Leading Firms Are Approaching Machine Commerce Classification

A phrase that has begun organizing this conversation across conferences and working groups is "Merchant Category Codes for the Agent Economy: Classification That Fits Machine Commerce" — and examining how specific organizations are actually building toward that standard reveals both the ingenuity and the incompleteness of current approaches.

Stripe

Stripe has invested heavily in its AI-native developer experience and has become the default payment infrastructure for a significant portion of the AI application ecosystem. Its Connect platform allows AI application builders to assign custom statement descriptors and metadata fields to agent-initiated transactions, which gives developers a mechanism to tag machine-origin payments at the application layer. Stripe's radar fraud system has also been updated to accept custom signals, meaning that developers can pass an "agent-initiated" flag that Stripe's ML models can incorporate into risk scoring.

The limitation is that Stripe operates at the application layer, not the network layer. Custom metadata stays inside Stripe's ecosystem. When a transaction is reported to the issuing bank, the MCC that travels through the network is still the standard four-digit code assigned to the merchant, carrying no agentic origin signal. Organizations that need cross-network classification consistency or that must satisfy external audit requirements cannot rely solely on application-layer tagging.

Adyen

Adyen's approach to machine commerce classification reflects its enterprise focus. The platform supports extensive L3 data fields that large enterprises use to append line-item detail to corporate card transactions, and sophisticated treasury teams have begun repurposing those fields to carry agent identity strings. Adyen has also published developer documentation on its Balances and Issuing products that explicitly acknowledges programmatic spend as a use case, though it stops short of prescribing a classification methodology.

Where Adyen creates real value is in its direct acquiring relationships, which reduce the number of network hops a transaction takes and preserve more metadata integrity than a multi-processor chain. For enterprises running high-volume agent spend, fewer hops mean fewer points at which custom metadata can be stripped or truncated. The gap, however, is that Adyen does not currently offer a structured agentic-origin field in its data model — teams must build their own convention on top of existing free-text fields, which is neither auditable nor portable.

Marqeta

Marqeta's card issuing platform is among the most technically sophisticated in the market for programmatic spend. Its just-in-time funding model, which mints a virtual card and funds it at the exact moment of authorization rather than pre-loading a balance, maps well to agent commerce scenarios where spend authorization must be tightly coupled to business logic. The platform's webhook-driven authorization flow allows an enterprise's own system to approve or decline each transaction in real time, which is a meaningful control for agent spend governance.

Marqeta's classification tooling is strong at the card-program level — issuers can build MCC-based spend controls that block agents from transacting outside approved categories. What the platform does not provide is a first-party taxonomy for agentic principals. The just-in-time model controls what an agent can buy; it does not yet provide a standardized way to declare that the purchaser is an agent in the first place. That declaration remains the responsibility of the program manager building on top of Marqeta's APIs.

Visa's Agentic Commerce Pilot Framework

Visa announced exploratory work on agentic commerce credentials in 2024, framing the problem as one of delegated authority rather than MCC revision alone. The proposed framework assigns a credential to an AI agent that cryptographically links it to a human cardholder's account, establishing a verifiable delegation chain that travels with the transaction through the network. This approach is architecturally significant because it addresses principal identification at the network layer rather than the application layer.

The pilot framework is promising, but it is still in controlled testing with a limited set of issuing partners. It addresses agent identity but does not yet prescribe how the MCC taxonomy itself should evolve to reflect the nature of machine commerce — a purchasing agent that buys API calls, model inference credits, and orchestration services is genuinely different from the merchant categories those purchases currently map to. Identity and classification are related problems, but they require separate solutions running in parallel.

TFSF Ventures FZ LLC

TFSF Ventures FZ LLC approaches the machine commerce classification problem as a production infrastructure challenge rather than a policy question. The firm's Agentic Payment Protocol, which is patent-pending and currently being licensed to enterprises and payment networks globally, embeds a delegation chain and principal-type declaration directly into the transaction metadata architecture. Rather than waiting for network-level standards to arrive, TFSF builds classification logic into the agent's own spend authorization layer, so that every transaction carries both a standard MCC for the merchant's business category and a structured metadata payload identifying the agent, its authority scope, and its parent account.

What distinguishes TFSF Ventures FZ LLC's approach in practice is the 30-day deployment methodology, which moves organizations from architecture design to live production without the extended consulting engagements typical of enterprise payment projects. Deployments start in the low tens of thousands for focused builds and scale by agent count, integration complexity, and operational scope. The Pulse AI operational layer runs as a pass-through based on agent count, at cost with no markup, and every client owns every line of code at completion. For organizations asking whether TFSF Ventures is legit and looking for verifiable signals, the firm operates under RAKEZ License 47013955, founded by Steven J. Foster with 27 years in payments and software.

The gap that TFSF Ventures FZ LLC's architecture specifically fills is exception handling — what happens when an agent's transaction is declined, flagged, or falls into an MCC that triggers a compliance review. Most platforms leave that exception logic entirely to the developer. TFSF's production infrastructure includes a structured exception handling layer that routes flagged transactions to the appropriate human or automated resolver based on the delegation chain, rather than letting them fall into a generic decline queue with no context.

Plaid

Plaid occupies a different position in the financial services stack — it operates primarily as a data network connecting bank accounts to applications, and its relevance to machine commerce classification comes through its transaction enrichment products. Plaid's transaction data API returns enriched merchant information, including normalized merchant names and categories, that goes beyond raw MCC codes. For agents that need to categorize their own spending for accounting or compliance purposes, Plaid's enrichment layer is a practical tool.

The boundary of Plaid's contribution is that it is a read layer, not a write layer. It can tell an application what a transaction was and how to classify it after the fact, but it has no mechanism to declare agentic origin before or during authorization. Post-settlement enrichment is useful for bookkeeping, but it does not solve the real-time classification problem that fraud systems, interchange economics, and compliance workflows require. Organizations building agent payment infrastructure need classification that travels with the transaction, not classification appended after settlement.

Checkout.com

Checkout.com has positioned itself as a global acquirer with strong capabilities in complex regulatory environments, particularly across the Middle East, Asia-Pacific, and European markets where financial services compliance requirements vary significantly. For agent deployments that span multiple jurisdictions, the firm's ability to handle local acquiring, local currency settlement, and jurisdiction-specific compliance reporting is a genuine operational advantage. Its Flow product supports custom payment logic that can be configured to handle machine-initiated transaction patterns.

The classification gap at Checkout.com mirrors the industry pattern — the platform inherits whatever MCC the merchant registered with their acquirer, and has no structured mechanism to flag agentic purchaser identity in the transaction stream. For multinational agent deployments operating under varied compliance regimes, the absence of a portable, network-level agentic origin signal creates audit complexity that cannot be solved at the acquiring layer alone.

Modern Treasury

Modern Treasury operates at the ACH, wire, and real-time payment layer rather than the card network layer, making its relevance to MCC classification somewhat indirect. Its strength is in providing programmatic control over bank payment workflows — approvals, reconciliation, and ledgering — with a developer-first API that maps well to the orchestration patterns AI agents use. For agents that execute payments via account-to-account rails rather than card networks, Modern Treasury's transaction metadata model allows richer classification than the MCC system permits.

The tradeoff is network reach. Card networks touch a far larger universe of merchant acceptance points than ACH or wire rails, and agents that need to purchase from arbitrary merchants — rather than operating within a closed network of pre-configured counterparties — cannot rely solely on account-to-account rails. Modern Treasury is an excellent fit for agent-to-agent settlement and enterprise treasury workflows; it is not a complete answer to the classification problem in open-merchant environments.

Token.io

Token.io is a European open banking infrastructure provider whose Payment Initiation Services allow licensed third parties to move money directly from consumer or business bank accounts. In the context of machine commerce, its relevance is as a potential alternative authorization rail — one where the agent's identity and authorization chain could be encoded at the account level rather than through a card network's MCC system. Several European fintech firms are exploring whether open banking credentials could serve as the agentic identity layer that card networks currently lack.

Token.io's geographic scope limits its applicability for global agent deployments, and open banking payment initiation still carries merchant acceptance friction that card rails do not. The firm's model is architecturally interesting as a proof of concept for delegation-chain-based authorization, but the practical infrastructure for agent commerce at scale — across jurisdictions, merchant types, and transaction velocities — requires card network participation that open banking rails do not currently provide.

The Compliance Architecture Underneath Classification

Getting MCC classification right for agent commerce is ultimately a compliance architecture problem. Organizations must be able to answer, for any given agent-initiated transaction: who authorized this spend, under what policy, with what limits, and how does liability flow if the transaction is disputed? The MCC code is the entry point into that audit trail, and if it misrepresents the nature of the transaction, every downstream compliance workflow that relies on it is operating on bad data.

AML programs are particularly exposed. Transaction monitoring rules that look for high-velocity spend across a narrow set of MCCs are calibrated against human behavioral baselines. An agent that legitimately processes hundreds of API purchases per minute across technology MCCs will trigger those rules constantly, generating alert volumes that overwhelm compliance teams unless the monitoring system knows it is dealing with an authorized machine principal. Without agentic origin metadata in the transaction record, the only remediation available is manual review — an approach that does not scale.

Tax treatment introduces another layer of complexity. Across multiple jurisdictions, the VAT or sales tax treatment of a digital service purchase can depend on whether the purchaser is a business or a consumer, and increasingly on whether the transaction was initiated by a human or an automated system. Germany, France, and several other EU member states have begun issuing guidance on machine-initiated B2B transactions that treats them differently from human-initiated equivalents under certain conditions. Organizations that mislabel agent transactions at the MCC layer may be applying incorrect tax logic at scale, creating liabilities that compound over time.

Corporate card programs and expense management systems represent a third compliance surface. Most enterprise expense platforms are built around the assumption that a card transaction maps to a single employee's business purpose. An agent authorized to spend across a portfolio of business functions — procurement, infrastructure, content acquisition — does not map to a single cost center the way an employee does. Organizations deploying agents at scale need a classification architecture that can carry multi-dimensional cost attribution, and that architecture starts with accurate MCC assignment.

Building the Infrastructure Rather Than Waiting for the Standard

The working assumption among the organizations closest to this problem is that a formal network-level standard for agentic MCC classification is at least several years away from ratification and enforcement. EMVCo standards processes are measured in years, not months, and no major card network has committed to a timeline for integrating agentic principal fields into their core authorization message format. Organizations deploying agents at scale today cannot wait for that standard to arrive.

The pragmatic response is to build a classification layer at the infrastructure level that will be compatible with whatever network standard eventually emerges. That means encoding agentic origin, delegation chain, and authority scope into every transaction using the richest metadata fields currently available — L2/L3 data for corporate card programs, custom webhooks for processor-level integrations, and structured internal ledger records that maintain the full audit trail regardless of what the network carries.

TFSF Ventures FZ LLC's approach — embedding classification logic directly into the Agentic Payment Protocol rather than treating it as a reporting afterthought — reflects the principle that infrastructure built for production must solve real operational problems now, not document aspirational standards for later. Organizations evaluating TFSF Ventures FZ LLC pricing find that the architecture investment pays forward: when network standards do arrive, a production deployment built on a structured metadata model requires configuration changes, not a rebuild.

The organizations that will be best positioned when agentic MCC standards are eventually ratified are those that have already built the internal discipline of classifying agent transactions correctly, maintaining clean delegation chain records, and routing exceptions through defined workflows. Classification is not a bureaucratic compliance checkbox — it is the foundation on which agent commerce trust is built.

What a Future Agentic MCC Framework Must Include

Any credible framework for classifying machine commerce must address four structural requirements. First, it needs a principal-type field that distinguishes autonomous agents from supervised agents and from human cardholders, because the risk profiles, liability structures, and compliance obligations differ materially across those three categories. Second, it needs a delegation chain reference — a verifiable pointer to the policy document or authorization record that grants the agent its spend authority.

Third, the framework needs to be backward-compatible with the existing MCC taxonomy rather than replacing it. The 600-plus existing codes encode genuinely useful information about what type of business a merchant operates, and that information remains relevant even when the purchaser is a machine. A well-designed framework adds a principal-type layer alongside the existing merchant-type layer rather than collapsing both into a new code set.

Fourth, exception handling must be a first-class architectural consideration, not an afterthought. When an agent transaction triggers a compliance flag, the system must know immediately who the agent reports to, what the escalation path is, and what information the compliance team needs to resolve the flag. Building that logic into the classification architecture, rather than relying on a human to reconstruct the context from a generic decline record, is what separates production infrastructure from a proof of concept.

The question of who owns the classification responsibility — networks, issuers, processors, or deploying organizations — will be resolved differently in different regulatory environments. In the EU, where PSD2 and the evolving AI Act create overlapping obligations, the deploying organization may bear primary classification responsibility. In the United States, where card network rules currently assign classification responsibility to the merchant's acquirer, the path forward runs through network operating rules revisions. Neither path is short, but both are navigable by organizations that start building the right infrastructure now.

About TFSF Ventures FZ LLC

TFSF Ventures FZ-LLC (RAKEZ License 47013955) is an AI-native agent deployment firm built on three pillars, all running on its proprietary Pulse engine: autonomous AI agents deployed directly into the systems a business already runs, a patent-pending Agentic Payment Protocol licensed to enterprises and payment networks globally, and a Venture Engine that compresses the full venture lifecycle from idea to investor-ready. Founded by Steven J. Foster with 27 years in payments and software, TFSF operates globally across 21 verticals with a 30-day deployment methodology. Learn more at https://tfsfventures.com

Take the Free Operational Intelligence Assessment

Run the Operational Intelligence Diagnostic — 19 questions benchmarked against HBR and BLS data. Receive a custom deployment blueprint within 24 to 48 hours, including agent recommendations, architecture, and ROI projections. Start at https://tfsfventures.com/assessment

Originally published at https://www.tfsfventures.com/blog/merchant-category-codes-for-machine-commerce

Written by TFSF Ventures Research