A payments company building an AI agent that settles invoices from a corporate wallet discovers that the agent's autonomy – its ability to initiate transfers without human confirmation on each transaction – may pull the business into the regulated perimeter for money transmission or payment services. The question is not whether the underlying rails are crypto or fiat; it is whether the operator, by deploying an agent with spending authority, has constructed an arrangement that resembles a payment service under the applicable regime. When the agent acts on instruction but without real-time consent, liability for error, for breach of spending limits and for regulatory classification all turn on how authorisation was evidenced at the outset and how control is maintained through the agent's lifecycle.
This analysis maps the authorisation perimeter for agentic payments, the points at which autonomous transaction initiation triggers a licensing obligation, and the legal architecture – mandate design, spending controls, audit trails – that operators use to stay outside the regulated boundary or to structure compliance when they cross it. It is written for the general counsel at a crypto business, a stablecoin issuer or a payments company deciding whether an autonomous-agent product can launch under existing authorisation or whether a payment licence, an e-money institution (EMI) registration or enhanced contractual safeguards are required first.
What brings agentic payments inside the regulated perimeter
An agentic payment is a transaction initiated by software acting on pre-set rules, market signals or natural-language instruction, without requiring the payer's manual approval for that specific transfer. The agent may operate within a crypto wallet, a stablecoin balance, a fiat account or across all three. Regulatory classification does not depend on the technology stack; it depends on whether the arrangement, viewed functionally, constitutes the acceptance and transmission of funds on behalf of another person – the definition that triggers money-transmission licensing in the United States, payment-institution authorisation under the EU Payment Services Directive (PSD2) and similar obligations in the United Kingdom, Singapore and the UAE.
The key legal question is: does the agent act as the customer (an extension of the customer's own control, analogous to standing instructions on a personal account) or does it act for the customer in a manner that introduces the operator as an intermediary in the transmission chain? If the agent operates within the customer's own wallet or account and the operator never holds, controls or transmits the funds, the arrangement typically stays outside payment-service definitions. If the operator provides the infrastructure, holds a pooled balance from which the agent draws or guarantees settlement to the recipient before the customer's funds clear, the operator has crossed into a regulated activity.
Under PSD2, payment initiation (instructing a transfer from the payer's account) and account information aggregation are carved out as lighter-touch regulated activities, but only when the service provider acts on explicit consent for a defined transaction or a defined series. An agent that initiates payments on standing authority – "pay any invoice under €5,000 from supplier X" – sits at the edge of that carve-out. Regulators in the major EU hubs increasingly treat broad standing mandates as requiring full payment-institution authorisation if the mandates allow the agent to determine timing, counterparty or amount within parameters set months earlier. The European Banking Authority (EBA) has signalled that AI-driven payment initiation must meet the strong-customer-authentication (SCA) framework under PSD2, meaning that the initial consent must be dynamic-linked to a specific amount and payee, or the agent must re-authenticate periodically. An operator that cannot show dynamic linking or periodic re-authentication will find that the regulator classifies the service as a payment account with continuous access – a Category-6 activity under PSD2, carrying full capital and safeguarding obligations.
In the United States, the question is whether the arrangement constitutes money transmission under the state-level money-transmitter statutes and the federal Bank Secrecy Act (administered by FinCEN). Most states define a money transmitter as a person who accepts currency or value for transmission to another location or person. If the agent operates within the customer's own non-custodial wallet and the operator never takes possession, there is no acceptance of funds and therefore no transmission. If the operator runs a hosted-wallet service where customer balances are pooled on-chain or off-chain and the agent initiates outbound transfers from that pool, the operator is a transmitter in every major state and must hold money-transmitter licences (MTLs) or a federal equivalent (the NYDFS BitLicense in New York, the Nationwide Multistate Licensing System registration in most others). The licensing analysis does not change if the agent is autonomous; the question is custody and transmission, not decision-making speed.
In Singapore, the Monetary Authority of Singapore (MAS) regulates payment services under the Payment Services Act. A digital payment token service (dealing in, exchanging or facilitating exchange of digital payment tokens) or an e-money issuance both fall within the standard payment institution or major payment institution licensing tracks. An agent that moves stablecoins on a customer's instruction, where the operator controls the private keys, is providing a DPT service; the autonomy of the instruction does not exclude the activity. For fiat payments, the MAS applies functional equivalence: if an agent initiates a bank transfer on behalf of a business customer under pre-authorised rules, the operator is a merchant acquisition service or account issuance service depending on the flow, and licensing follows.
The ADGM (Abu Dhabi Global Market) and the DIFC (Dubai International Financial Centre) both regulate payments and virtual-asset activities under their respective financial-services frameworks. In ADGM, an operator providing payment services (including operating a payment account or executing payment transactions) requires authorisation from the Financial Services Regulatory Authority (FSRA). Agentic initiation does not exempt the activity. In Dubai, VARA (the Virtual Assets Regulatory Authority) applies to virtual-asset transfers; an agent moving crypto on behalf of customers falls within VARA's transfer-and-settlement activity, which is a separately licensable function.
In our cross-border practice, operators launching agent-driven payment products typically face this decision tree: (1) deploy the agent within each customer's self-custody wallet, avoiding any operator control over funds, or (2) obtain the relevant payment or e-money licence and structure the agent as a feature of a regulated payment account. The first path limits functionality – no pooled liquidity, no settlement guarantees – but avoids licensing. The second path is the institutional route, used by businesses that already hold EMI or payment-institution authorisation and that treat the agent as an API-driven enhancement to their existing service.
How authorisation and consent are evidenced when an agent acts
A regulated payment service requires consent from the payer before each transaction or, under limited circumstances, before a series of transactions under defined parameters. When a human initiates a transfer by clicking "send" or by entering a one-time password, consent is straightforward. When an autonomous agent initiates the same transfer, the legal question is whether the original instruction to the agent – "pay my suppliers on these terms" – constitutes valid consent for every subsequent payment the agent makes under those terms, and whether that consent remains valid across weeks, months or contract renewals.
Under PSD2, consent must be given in a form that the payment service user can understand, and it must be explicit for the specific transaction. For standing orders or direct debits, PSD2 allows a single consent to cover a series, but the mandate must specify the payee, the recurrence pattern and, often, a maximum amount per transaction or per period. An agent instructed in natural language – "pay my energy bills promptly" – does not produce a PSD2-compliant mandate because the counterparty, amount and timing are open. A compliant agentic payment under PSD2 typically requires the customer to sign a digital mandate at onboarding that names each payee category, sets per-transaction and daily limits, and defines the conditions under which the agent may act. The operator then maintains an audit log showing that each payment the agent initiates falls within the signed parameters, and the customer must have a real-time method to revoke or adjust the mandate.
In the United States, consent standards for automated payments depend on whether the underlying rail is an ACH debit, a card transaction or a crypto transfer. For ACH debits under NACHA rules, the payer's authorisation may be written, oral or electronic, but it must identify the payee, authorise the payee (or its agent) to initiate debits, and specify the timing or triggering event. An agent acting under a customer's signed terms of service that meet these elements can initiate ACH debits on the customer's behalf, and liability for unauthorised debits then turns on whether the debit matched the authorisation. For card payments, the card networks (Visa, Mastercard) prohibit storing a customer's CVV and require the customer's explicit consent for each recurring charge. An agent initiating card payments must either use a stored payment token (which the networks allow under their recurring-billing frameworks) or must prompt the customer for authentication at the time of each payment. Most agentic-payment operators on card rails therefore use tokenisation and treat the agent's action as falling within a merchant-initiated transaction (MIT) under the card-network rules, which exempts the transaction from SCA but requires the original consent to specify that future agent-driven charges are authorised.
For crypto and stablecoin payments, there is no equivalent to PSD2's consent regime or to NACHA's authorisation rule, because there is no intermediary payment service provider in a pure peer-to-peer transfer. Consent becomes relevant only when the operator provides a hosted wallet or a custodial service, at which point the operator's terms of service must define the scope of the agent's authority. In a recent matter for a stablecoin-payments company, the customer's wallet allowed an agent to initiate transfers up to a specified daily limit, and the question was whether a transfer that exceeded the limit due to a miscalculation by the agent was an unauthorised transaction under the wallet agreement. The answer turned on whether the agreement treated the limit as a hard cap (making any excess unauthorised) or as a guideline with a breach-notification obligation. The agreement was drafted as a guideline, and the company therefore bore the risk of the excess and reimbursed the customer. The revision for the next contract version made the limit a hard cap enforced by the wallet's signing logic, eliminating ambiguity.
Across all rails, the enforceability of an agent's instruction depends on attribution: can the operator prove that the agent acted within the scope of authority granted by the customer? Attribution requires three evidentiary layers. First, a signed digital mandate or terms-of-service acceptance that describes the agent's authority in functional terms (what it will do, under what conditions, subject to what limits). Second, a cryptographic or session-based link between the customer's identity and the agent's actions – typically an API key, a wallet signature or a session token tied to the customer's authenticated login. Third, an immutable audit log recording each action the agent took, the rule or signal that triggered it, the timestamp and the resulting transaction hash or reference number. Operators we advise in the EU and Singapore use append-only audit trails on permissioned ledgers (Hyperledger, Quorum, or even a public chain for hash anchoring) to meet both the PSD2 record-keeping obligation and the evidential standard for disputes.
Where liability lands when an agent makes an error
An autonomous agent operating under pre-set rules will, over a sufficient number of transactions, execute a payment that the customer did not intend – a wrong counterparty, a wrong amount, a payment triggered by stale data or by a spoofed input. Liability for that error is a function of contract (the customer agreement), the applicable payment-services regulation, and, in cases involving fraud or negligence, tort. The question is whether the error is treated as an unauthorised transaction (for which the customer has a statutory right to a refund) or as a transaction authorised under the signed mandate but executed incorrectly (for which liability is contractual and may be excluded or capped).
Under PSD2 Article 73, a payment service user may request a refund of an unauthorised transaction, and the payment service provider must refund the amount unless it can prove that the user authorised the transaction or failed to keep personalised security credentials safe. If an agent acting under a valid mandate initiates a payment to the wrong recipient because of a coding error, the transaction is authorised under the mandate but mis-executed, and PSD2 treats this as an incorrectly executed transaction under Article 89. The provider must then prove that the payment reached the payee's payment service provider correctly, or it must refund the payer and pursue recovery from the recipient's provider. The practical result: if the error is in target selection (the agent sent funds to Supplier B instead of Supplier A, both of whom were within the authorised payee list), the operator bears the loss unless it can prove the customer's instruction was ambiguous. If the error is in amount calculation (the agent paid €10,000 instead of €1,000 due to a decimal error), the operator again bears the initial liability and must seek recovery from the recipient if the recipient does not voluntarily return the excess.
In the United States, the Electronic Fund Transfer Act (Regulation E) applies to consumer transactions but not to most business payments. For consumer agentic payments, an unauthorised transaction triggers the same liability caps as a stolen debit card: $50 if reported within two business days, $500 if reported within 60 days, and unlimited if reported later. For business payments, Regulation E does not apply, and the Uniform Commercial Code (UCC Article 4A, governing wire transfers) or the NACHA rules (for ACH) set the framework. Under UCC 4A, if a payment order is issued by an agent authorised by the customer, the customer is bound even if the agent made an error, unless the customer can prove that the error resulted from the bank's failure to follow a security procedure agreed with the customer. Most payment agreements for businesses therefore include a clause stating that the customer accepts the risk of agent errors and that the operator's liability is capped at direct damages or at a specified multiple of the monthly fee.
For crypto payments, there is no statutory refund right. Liability is purely contractual, and most hosted-wallet terms of service state that transactions, once broadcast, are final and that the operator has no ability to reverse them. The question then is whether the agent's error constitutes a breach of the wallet agreement (for example, because the agent exceeded a spending limit) and whether that breach gives rise to a damages claim. Operators in this space typically cap liability at the value of the erroneous transaction or at the customer's account balance, whichever is lower, and exclude liability for consequential or indirect loss. In a recovery matter involving a DeFi protocol where an agent interacting with a liquidity pool executed a swap at an unfavourable rate due to a front-running attack, the customer argued that the protocol operator owed a duty of care in designing the agent's execution logic. The protocol's terms of service disclaimed any agency relationship and stated that users interacted with the smart contracts directly; the protocol operator was not a party to the swap and owed no duty beyond providing access to the contracts as-is.
The distinction between agent error and customer error also matters for liability allocation. If the customer mis-configured the agent's parameters (for example, by setting a daily limit in dollars when the agent interprets the limit in euros), is that the operator's failure to provide clear instructions or the customer's failure to understand the product? Courts and regulators have not yet drawn a settled line, but the emerging approach in the EU and UK is that the operator must provide sufficient interface design and warnings to prevent foreseeable misconfiguration. A simple example: if a spending-limit field accepts numeric input without a currency selector and the operator assumes the input is in the customer's account currency, that design will be questioned if it leads to an unintended payment. Operators launching agentic products in 2025 are therefore building confirmation layers into the onboarding flow – a final screen that shows the customer what the agent will do under sample scenarios – and are requiring the customer to acknowledge each scenario before the agent is activated.
The stablecoin-payment and fiat-rails divide for agent authorisation
The question whether an agentic payment moves over crypto rails, fiat rails or both changes the licensing analysis but does not eliminate it. A stablecoin payment initiated by an agent falls under the VASP or crypto-asset-service-provider (CASP) licensing regimes in jurisdictions that regulate virtual assets, and under e-money or payment-institution licensing in jurisdictions that classify stablecoins as e-money. A fiat payment initiated by an agent falls under the payment-services framework of the jurisdiction where the payment account sits. Many operators run hybrid products – the agent holds a stablecoin balance for speed and low fees, and converts to fiat at the point of payment to the recipient – and those products must satisfy both regimes.
Under MiCA (the EU Markets in Crypto-Assets Regulation), a stablecoin classified as an e-money token (EMT) requires the issuer to be an electronic money institution or a credit institution. If the operator of the agentic-payment product is also the issuer, the operator must hold EMI authorisation. If the operator uses a third-party stablecoin, the operator must still determine whether its service – accepting the stablecoin from customers, holding it in custody and initiating transfers on the customer's instruction – constitutes custody and administration of crypto-assets on behalf of clients, which is a separately licensable CASP activity under MiCA Article 3(1)(5). The answer, again, depends on custody: if the customer retains control of the private keys and the operator's agent simply constructs and broadcasts signed transactions on the customer's behalf, the operator is not providing custody and may avoid CASP authorisation. If the operator holds the keys, CASP authorisation is required, and the capital and organisational requirements of MiCA Title III apply.
In the UAE, VARA (Virtual Assets Regulatory Authority, governing mainland Dubai) applies to transfer and settlement services for virtual assets. An agent that moves stablecoins on behalf of customers is providing a transfer service under VARA's framework, and the operator must hold a VARA licence for that activity. The licensing application assesses the operator's AML procedures, its technology safeguards (key management, multi-signature arrangements, cold storage for client funds) and its insurance coverage. VARA does not distinguish between human-initiated and agent-initiated transfers; both fall within the same regulated activity. In ADGM, the FSRA applies a similar functional test: if the service involves holding or transferring virtual assets for customers, it is a regulated activity. ADGM does, however, offer a FinTech regulatory laboratory that allows operators to test agentic-payment products under restricted conditions (limited customer numbers, capped transaction volumes) before applying for full authorisation, and several agent-payment startups launched in ADGM used that pathway in 2024.
For fiat payments, the EU/EEA, UK and Singapore all treat payment initiation as a regulated activity when provided as a service to third parties. An agent that initiates SEPA transfers on behalf of a corporate customer, even if the agent operates under a pre-signed mandate, is providing a payment-initiation service under PSD2 or an equivalent. If the operator also holds customer funds (for example, by collecting payments into a pooled account before the agent disburses them), the operator is providing a payment account service and must hold full payment-institution authorisation. The capital requirement for a payment institution in the EU ranges from €20,000 (for payment initiation only) to €125,000 (for operating payment accounts), and the operator must meet safeguarding obligations – client funds held in a segregated account at a credit institution or covered by insurance or a guarantee.
The cross-border dimension is that many agentic-payment operators want to serve customers across multiple jurisdictions, using stablecoins to settle cross-border in minutes and fiat rails for local on-ramps and off-ramps. That model requires either (1) obtaining licences in each jurisdiction where customers sit, (2) partnering with licensed entities in each jurisdiction and white-labelling the agent layer, or (3) structuring the product so that the operator never holds, controls or transmits funds – the agent operates purely as software provided to customers who retain full control of their own accounts or wallets. In our licensing practice, the third model is common for early-stage operators who want market validation before investing in multi-jurisdiction licensing, and the second model is the scale path once product-market fit is proven. The first model – obtaining licences directly in 20+ jurisdictions – is typically viable only for well-capitalised operators or for those entering via a jurisdiction that allows passporting (an EMI authorised in Lithuania, for instance, may passport across the EU/EEA under PSD2).
How banking access and crypto-banking interaction affect agent deployment
An operator launching an agentic-payment product needs banking – a fiat account to settle on-ramps and off-ramps, to pay suppliers and to hold safeguarded client funds if the product includes a payment account. Access to banking for crypto-related businesses remains uneven: some banks in the major hubs serve licensed VASPs and payment institutions, but most retail and mid-tier banks decline crypto accounts as a category. The result is that an operator's choice of licensing jurisdiction is often driven as much by where banking is available as by the speed or cost of the licence itself.
In the EU, banks are more willing to open accounts for licensed EMIs and payment institutions than for unlicensed crypto businesses. An operator that holds a Lithuanian EMI licence or a Malta CASP authorisation under MiCA has a credible case for banking in those jurisdictions and, increasingly, in other EU states via passporting. In practice, however, banking access often requires substance – a local office, local staff and a credible explanation of why the business needs an account in that bank's jurisdiction rather than in the operator's home state. For a remote-first operator with developers in three countries and no physical office, that threshold is difficult to meet, and the path is to partner with a licensed EMI that already has banking and that can white-label a payment account to the operator's customers.
In Singapore, banks apply a risk-based approach to crypto businesses. An operator that holds a MAS payment-institution licence and that can demonstrate strong AML controls, a clear business model and acceptable transaction volumes may obtain an account with DBS, OCBC or UOB, though onboarding often takes 6–12 months and requires regular dialogue with the bank's crypto-risk desk. An unlicensed operator will struggle, and many choose to partner with a licensed Singaporean payment-services provider rather than applying directly for banking.
In the UAE, banking access for crypto businesses has improved since 2023. Several banks in ADGM and in Dubai now serve VARA-licensed entities and ADGM-licensed virtual-asset firms, and account opening is part of the licensing-application process – the regulator expects the applicant to demonstrate a banking relationship or a letter of intent before licence issuance. For a crypto-payment operator launching an agent-driven product, the ADGM or VARA route offers both licensing and banking in a single jurisdiction, which simplifies compliance and reduces the need for correspondent banking. That said, the operator must still meet the bank's KYC and transaction-monitoring expectations, and banks in the UAE typically require the operator to use a recognised blockchain-analytics provider (Chainalysis, Elliptic, TRM Labs) and to provide quarterly AML reports.
The stablecoin-fiat bridge is a specific banking pain point. An agent that converts stablecoins to fiat to pay a supplier needs an off-ramp – a fiat account at a bank or EMI that will accept incoming funds from a crypto exchange or liquidity provider. Many banks treat these flows as high-risk and either decline them or impose enhanced due diligence on each transaction. Operators solve this in three ways: (1) use a crypto-friendly EMI (such as those in Lithuania or the UK) as the off-ramp, then transfer fiat onward to a mainstream bank for final settlement; (2) use an over-the-counter (OTC) desk that settles fiat directly to the supplier's bank, skipping the operator's own account; or (3) structure the product so that the customer, not the operator, is the party converting stablecoins to fiat, and the operator simply provides the software that triggers the conversion – avoiding any money-transmission or payment-account characterisation.
Mandate and contract design for compliant agentic payment
A compliant agentic-payment product rests on a digital mandate that defines the agent's authority, the limits on that authority, and the customer's rights to monitor, adjust or revoke the mandate. The mandate must meet the consent requirements of the applicable payment-services regulation, and it must allocate liability in a manner that the regulator and, if disputed, a court will enforce.
The mandate structure we recommend in cross-border work includes six elements. First, a definition of the agent's function in plain language, supplemented by a technical description in an appendix. For example: "The Agent will monitor your designated invoices and, when an invoice is due, will initiate a payment from your Wallet to the invoice issuer in the amount stated on the invoice, provided that the payment does not exceed your Daily Limit and that the payee is on your Approved Payee List." Second, a clear identification of the triggering events or conditions – time-based (every Monday at 09:00 UTC), event-based (when an invoice is marked "due"), market-based (when the price of asset X reaches threshold Y), or instruction-based (when you send the Agent a command via the API). Third, hard caps on transaction size, daily and monthly volume, and the number of transactions per period. Fourth, an audit and notification protocol – the Agent logs every action, and the customer receives a summary (real-time, daily or weekly, depending on the product tier). Fifth, a revocation and adjustment mechanism – the customer may pause the Agent, adjust the limits or remove a payee at any time through the platform interface or via API call, and changes take effect within a defined window (immediately for a pause, within one business day for limit adjustments). Sixth, a liability allocation – the operator is liable for unauthorised transactions (those outside the signed mandate) and for errors caused by the operator's systems; the customer is liable for errors caused by mis-configuration of the Agent's parameters or by failure to monitor the account and report a problem within the timeframe specified in the agreement (commonly 60 days under PSD2, 120 days under NACHA rules).
The mandate must be presented in a manner that the customer can review and understand before signing. For PSD2 compliance, that means the mandate cannot be buried in 30 pages of terms and conditions; it must be a standalone document or a clearly delineated section with a separate acceptance tick-box. For crypto-native products, operators increasingly use a simulation interface – the customer inputs sample parameters (daily limit $5,000, approved payees A, B and C, trigger condition "invoice due"), and the interface shows a mock timeline of payments the Agent would have made over the past 30 days under those parameters. The customer then confirms that the simulated behaviour matches intent, and that confirmation forms part of the signed mandate.
The enforceability of liability caps in the mandate depends on jurisdiction and customer type. Under PSD2, an operator cannot exclude or limit liability for unauthorised transactions or for gross negligence. Liability for negligence (not gross) may be limited if the customer is a microenterprise (under 10 employees, turnover under €2 million) and only if the term is individually negotiated or if the customer had a genuine opportunity to use a different provider. For larger businesses, PSD2's consumer-protection provisions do not apply, and liability may be capped or excluded by contract, but the term must still meet the general fairness standards of the national contract law (such as the Unfair Contract Terms Act in the UK or the comparable provisions in civil-law jurisdictions). In the United States, liability caps in commercial agreements are generally enforceable for business customers, provided that they are not unconscionable and that they do not exclude liability for wilful misconduct or fraud. The cap is typically expressed as a multiple of fees paid (for example, "total liability in any 12-month period shall not exceed three times the fees paid by Customer during that period") or as a fixed amount per incident (e.g., $10,000 per unauthorised transaction).
The audit trail and the regulator's compliance expectations
An operator providing agentic payments must demonstrate to the regulator that every transaction the agent initiated was within the scope of the customer's mandate, that the customer had the ability to monitor and control the agent, and that the operator has the records to reconstruct any disputed transaction. Those requirements translate into a comprehensive audit trail that meets both payment-services record-keeping rules and the evidential standards for disputes.
Under PSD2 Article 42, a payment institution must keep records of all transactions for at least five years. The records must include the transaction amount, currency, date, payer and payee identifiers, and the type of payment service. For agentic payments, the regulator expects the audit trail to also record the trigger – what caused the agent to initiate the transaction – and the authorisation basis – which element of the customer's mandate permitted that transaction. In practice, that means each log entry contains the transaction hash or reference, the timestamp, the payer and payee identifiers, the amount, the agent's version number, the rule or signal that triggered the transaction, and a pointer to the signed mandate document that granted authority for that type of transaction.
In our licensing advisory work, we have seen national competent authorities (NCAs) in the EU examine operators' audit trails during initial-authorisation assessments and during ongoing supervision. The common findings: (1) logs are incomplete – the operator records the transaction but not the trigger or the mandate version; (2) logs are mutable – they are stored in a SQL database where entries can be edited or deleted after the fact; (3) there is no link between the transaction log and the customer's signed mandate – the operator can show what the agent did but cannot prove that the customer authorised it. All three findings delay or prevent authorisation. The solution is an append-only, tamper-evident log, typically implemented as a permissioned blockchain (Hyperledger Fabric, Quorum, or Polygon for lower cost), a hash chain anchored to a public blockchain (Bitcoin, Ethereum), or a cloud-based immutable log service (AWS QLDB, Azure Confidential Ledger). The log records each agent action as a signed, timestamped entry, and the hash of each block of entries is published on-chain or shared with the regulator, making post-facto alteration detectable.
For AML purposes, the operator must monitor the agent's transactions for patterns consistent with money laundering or terrorist financing. Under the EU's AML directives (AMLD5 and AMLD6), a payment institution must apply customer due diligence, monitor transactions and report suspicious activity to the financial intelligence unit. An agent that initiates hundreds or thousands of transactions per day on behalf of multiple customers creates a monitoring challenge: the operator cannot rely on human review of every transaction, and must instead implement automated transaction monitoring that flags high-risk patterns (structuring, rapid movement of funds through multiple accounts, payments to high-risk jurisdictions). The monitoring system must distinguish between legitimate autonomous behaviour (for example, an agent paying 50 small suppliers in one day under a valid procurement mandate) and suspicious activity (an agent that suddenly shifts from small supplier payments to large transfers to newly added payees in a non-cooperative jurisdiction). In a recent matter, an agent-payment operator in the EU was flagged by its bank for unusual transaction patterns; the investigation revealed that a customer had mis-configured the agent to pay invoices from a spoofed email, and the funds had moved to an account in a high-risk jurisdiction. The operator's transaction-monitoring system had not flagged the pattern because the payee was on the customer's approved list. The operator revised the monitoring rules to flag any new payee added within 24 hours of a large payment, and to require manual review of payments over a certain threshold to payees added in the past 30 days.
A common assumption – that AI autonomy exempts the operator from payment-services regulation
A recurring assumption among operators building agentic-payment products is that if the agent is "autonomous" – if it makes decisions based on AI, market data or natural-language input without direct human triggering – the operator is not providing a payment service because the operator is not executing the transaction, the agent is. That assumption misreads the regulatory test. Payment-services regulation applies to any person that provides a service consisting of accepting, transmitting or making funds available, not to the mechanism by which the service is provided. Autonomy, from a regulatory perspective, is a feature of the service, not a defence against classification.
The legal analogy is to a standing order or direct debit, which have existed for decades. A customer authorises a bank to make recurring payments on set dates or when triggered by the payee. The fact that no human at the bank manually executes each payment does not mean the bank is not providing a payment service; it means the bank is providing a payment service efficiently. Regulatory frameworks account for automation; PSD2 explicitly defines payment initiation and recurring payments, and MAS's Payment Services Act includes automated fund transfers within the definition of payment services. The question is not whether a human clicked "send", but whether the arrangement constitutes the provision of a payment service under the functional definitions in the statute.
A second variant of the objection is that if the agent operates on-chain and the operator never holds fiat, the operator is outside the payment-services perimeter. That is half correct: an operator that provides software enabling customers to interact with smart contracts on a public blockchain, where the customers retain control of their private keys and the operator never has access to funds, is not providing a payment service or a custody service – the customers are transacting peer-to-peer using the operator's software as a tool. But if the operator runs a hosted-wallet service, holds pooled customer balances (even on-chain in a multisig contract), or offers any settlement guarantee to the payee, the operator has introduced itself as a functional intermediary, and licensing analysis follows. The fact that the underlying infrastructure is a blockchain does not exempt the operator; MiCA and VARA both regulate custody and transfer services provided via blockchain infrastructure.
Decision matrix – which structure for which operator profile
An operator planning an agentic-payment launch faces a choice among four structural models, each with different licensing, banking and liability consequences. Profile A: the software-only provider. The operator builds and licences agent software that customers install in their own self-custody wallets or that interacts with customers' existing bank accounts via open-banking APIs. The operator never holds, controls or transmits funds. Licensing requirement: none for payment services (the customer is the payment service user interacting directly with their own account or wallet); possibly a software or data licence depending on the jurisdiction. Banking requirement: the operator needs a corporate account for fee collection but not a safeguarded client-funds account. Liability: contractual, typically capped; the operator is liable for software defects but not for transaction errors caused by the customer's configuration. This model suits early-stage operators, particularly those building for a technical audience (crypto traders, DeFi protocols, treasury teams at Web3 companies) who are comfortable with self-custody.
Profile B: the licensed payment institution or EMI. The operator holds a payment-institution or e-money-institution licence and provides a payment account (fiat, stablecoin, or both) from which the agent operates. The operator controls customer funds, subject to safeguarding obligations. Licensing requirement: full payment-institution or EMI authorisation in the EU, MAS payment-institution licensing in Singapore, state MTLs in the United States, or an equivalent in the target jurisdiction. Capital: €125,000 minimum in the EU for a payment account provider, higher in practice; S$100,000–S$250,000 in Singapore depending on volume; per-state requirements in the US (often $100,000–$500,000 net worth per state). Banking requirement: a safeguarded client-funds account at a credit institution; many EMIs use a large EU bank (often in Germany, France or the Netherlands) or a Singaporean bank. Liability: PSD2 refund rights apply; operator liable for unauthorised transactions and for execution errors. Timeline: 6–18 months from application to authorisation in the EU/Singapore; US MTL aggregation can take 12–24 months across the major states. This model suits operators with capital, regulatory experience and a path to banking, and it allows the operator to offer settlement guarantees, pooled liquidity and a seamless user experience (the customer's funds and the agent both sit within the operator's platform).
Profile C: the white-label partner. The operator partners with a licensed EMI or payment institution; the EMI provides the regulated payment account and the operator provides the agent software as a feature within the EMI's platform. Licensing requirement: the EMI holds the licence; the operator is typically the EMI's technology vendor, not a regulated entity itself (though some arrangements require the operator to register as an agent or representative of the EMI). Banking: the EMI's banking relationship is used. Liability: split contractually between the EMI (which is liable to the customer under PSD2) and the operator (which is liable to the EMI for defects in the agent software under the white-label agreement). This model suits operators without capital or regulatory capacity who want to launch quickly; the trade-off is that revenue is shared with the EMI and the operator's brand is subordinate to (or hidden behind) the EMI's brand.
Profile D: the VASP or CASP (crypto-only). The operator provides a hosted stablecoin wallet with agentic payment features; the agent moves only stablecoins, never fiat. Licensing requirement: VASP registration or licensing in the target jurisdiction (VARA in Dubai, ADGM in Abu Dhabi, BitLicense in New York, MAS in Singapore, MiCA CASP in the EU). Banking requirement: limited – the operator needs fiat banking for its own operations and possibly for an on-ramp/off-ramp service, but customer funds are held in stablecoins on-chain, not in a bank account. Liability: contractual; no PSD2 refund rights because stablecoins are not e-money under most current frameworks (though this changes if the stablecoin is classified as an EMT under MiCA, at which point EMI regulation applies). This model suits operators serving a crypto-native customer base (exchanges, token issuers, DAOs, cross-border contractors paid in stablecoins) and offers faster licensing and less banking friction than the fiat-payment models, at the cost of a narrower addressable market.
When counsel enters the design process
The legal architecture of an agentic-payment product – the mandate structure, the audit trail, the liability allocation, the licensing path – is not a compliance layer added after the product is built; it is a design constraint from the concept stage. A product designed without input on authorisation, consent and regulatory classification often cannot be retrofitted to meet licensing requirements, because the core assumptions (who holds the keys, who controls the funds, how consent is evidenced) are embedded in the technology stack. Waiting until the product is in beta to engage counsel typically means either a full rebuild or a decision to operate in a grey area with regulatory risk.
The right point to bring legal into the process is at the product-definition stage, when the founding team is deciding whether the agent will operate within customers' self-custody wallets or within a hosted-wallet service, whether the operator will offer settlement guarantees, and which jurisdictions the product will serve. Counsel at that stage pressure-tests the licensing path, maps the banking requirements, drafts the initial mandate structure, and often participates in the technical-architecture review to ensure that the logging, key-management and signing flows align with the regulatory and evidential requirements. For a cross-border agentic-payment product serving EU, UAE and Singapore customers, the licensing path might be: (1) obtain EMI authorisation in Lithuania, (2) passport to other EU states under PSD2, (3) partner with a licensed payment-services provider in Singapore rather than applying for a MAS licence directly, and (4) structure the UAE offering as a VASP-licensed stablecoin wallet with a white-label fiat on-ramp through a UAE-licensed EMI. Each of those decisions depends on capital availability, timeline and the operator's appetite for compliance overhead, and each decision affects product design.
In our cross-border practice, we see three recurring patterns of engagement. First, the operator has already launched an MVP and is now facing questions from a bank or a prospective investor about licensing – this is the "we need a licence yesterday" scenario. If the product design is fundamentally unlicensable under the target regime (for example, the operator holds customer funds but has structured the service as a software licence, not a payment account), the options are limited: pivot the product, wind down and relaunch under a new structure, or operate in a jurisdiction with lighter licensing requirements and accept the trade-off of narrower banking access. Second, the operator is pre-launch and wants a licensing roadmap before committing capital to development. This is the optimal engagement: the firm maps the regulated perimeter across the target jurisdictions, identifies the path requiring the least capital and time (often a white-label partnership or a single-jurisdiction licence with international customer access), and provides the mandate and contract templates that meet the requirements of that path. Development then proceeds with those constraints in mind. Third, the operator is mid-scale (thousands of users, meaningful transaction volume) and is now encountering banking friction – the current account provider has issued an exit notice, or a prospective banking partner has declined to onboard without a payment-institution licence. This scenario requires a fast-track licensing application (where available, such as Lithuania or Malta under MiCA) or a quick pivot to a white-label arrangement with a licensed partner, combined with a managed transition of customer balances and contractual novation of the customer agreements.
Related at OBOLUS
- Banking, Payments & EMI Onboarding for Digital-Asset Businesses – licence path mapping, application management, banking relationship establishment across the major hubs.
- Institutional Digital-Asset Custody: Legal and Regulatory Design – structuring custody arrangements, key-management frameworks and safeguarding for client assets in multi-jurisdiction operations.
- Corporate Bank Account Opening in Abu Dhabi Global Market (ADGM) – practical guidance on banking access in ADGM, the documentation timeline, and how digital-asset businesses meet the substance and KYC expectations.
FAQ
Can a crypto firm access banking under this regime?
Banking access for crypto firms depends on licensing status and jurisdiction. In the EU, an operator holding EMI or payment-institution authorisation has a credible path to banking with mainstream banks, though substance (local office, local staff) often matters as much as the licence. In Singapore, licensed payment institutions may bank with DBS, OCBC or UOB after a lengthy onboarding process. In the UAE, VARA-licensed and ADGM-licensed operators have dedicated banking channels with several UAE banks. Unlicensed operators typically must partner with a licensed EMI for banking access or accept limited service from crypto-friendly challenger banks.
What safeguarding applies to client funds?
Under PSD2, a payment institution holding client funds must segregate those funds in a separate account at a credit institution or cover them with insurance or a comparable guarantee. The safeguarded funds are protected in the event of the payment institution's insolvency and cannot be used for the institution's own business. E-money institutions under the EU E-Money Directive face similar safeguarding rules. For stablecoins held in a hosted wallet, safeguarding depends on whether the stablecoin is classified as e-money (in which case EMI safeguarding rules apply) or as a crypto-asset (in which case MiCA's custody rules apply, requiring segregation and a clear record of which assets belong to which client).
Which payment licence is required?
The licence requirement depends on the service provided and the jurisdiction. If the operator accepts and transmits fiat on behalf of customers, or operates a payment account, a payment-institution licence (EU, UK, Singapore) or money-transmitter licences (US) are required. If the operator issues or administers e-money, an EMI licence (EU, UK) is required. If the operator provides custody or transfer services for crypto-assets, a VASP licence (VARA, ADGM, BVI, Cayman) or a CASP authorisation (EU under MiCA) is required. If the operator provides only software and never holds or transmits customer funds, no payment-services licence is required, though local business licensing may apply.
About OBOLUS
OBOLUS is an independent digital-asset law boutique acting only for businesses. We advise exchanges, custodians, token issuers and funds on licensing across 70+ jurisdictions, on disputes and on-chain asset recovery across 25+ forums, and on the tax, banking and compliance that sit around them. Digital assets are the whole of our practice. Our clients include operators launching agentic-payment products who need the legal architecture built into the product from the design stage, and who need the licensing path mapped before capital is committed to development. We work at the intersection of payment-services regulation, digital-asset licensing, and banking access – the three constraints that determine whether an autonomous-payment product can launch and scale. To discuss your situation, contact info@oboluslaw.com.
By Victor Olsen, Regulatory & Compliance Analyst – advising on payment-services licensing, agent-software regulatory classification, and the interaction between PSD2, MiCA and multi-jurisdictional VASP frameworks for cross-border operators.
This publication is general information about the law and does not constitute legal advice. It is not a substitute for advice tailored to your circumstances. OBOLUS accepts no liability for action taken or not taken on the basis of this material. For advice on your situation, contact info@oboluslaw.com.