Home › Blog › UK payments

AI Agents and Payments Regulation 2026: What Fintechs Need to Know Before Launching Agentic Payments

Published 21 July 2026 · Last reviewed 17 October 2026 · Buckingham Capital Consulting
Contents
  1. What agentic payments are
  2. How agentic payments are regulated in the UK
  3. Which permission an agentic payment business needs
  4. Payment initiation and open banking
  5. Consent and mandates
  6. Strong customer authentication
  7. Liability when an agent gets it wrong
  8. Know your agent
  9. Consumer Duty for agentic products
  10. Business and treasury agents
  11. Safeguarding and financial crime
  12. Fraud when machines transact with machines
  13. Third-party AI providers and governance
  14. Operational resilience
  15. Using a regulated payments partner
  16. Before launch: a checklist
  17. Where the framework is heading
  18. How Buckingham Capital Consulting helps
  19. Frequently asked questions
Key takeaways
  • There is no separate UK licence for agentic payments; the Payment Services Regulations 2017 apply according to the activity the business performs.
  • An agent that initiates payments from a customer's bank account usually means payment initiation permission, with €50,000 initial capital and professional indemnity insurance.
  • Strong customer authentication still applies to AI-initiated payments; existing exemptions such as trusted beneficiaries remain the only routes around it.
  • PSRs liability rules, including the £35 customer cap and the 8-week refund right for unspecified amounts, were not written for agents with discretion.
  • HM Treasury's consultation, closed on 6 October 2026, asked how authentication, consent and liability should change for agentic payments.

What agentic payments are

An agentic payment is a payment made by an AI system acting with some autonomy for a consumer or business. The customer sets the objective, the permissions and the limits; the agent then chooses what to buy, from whom, when and how to pay, without a separate instruction for each payment.

  • A consumer asks an agent to book a flight to Madrid for under £400 next month, paid from a named account.
  • A retailer lets an agent reorder stock from approved suppliers when inventory falls below a set level, within a monthly budget.
  • A finance team lets a treasury agent move surplus cash between accounts or pay approved invoices under predefined rules.

Unlike a standing order or direct debit, the agent exercises judgement between alternatives. That is what makes the regulatory analysis different: the UK payments rules assume an identifiable customer gives each instruction to a regulated firm, and an agent sits between the two.

How agentic payments are regulated in the UK

There is no separate FCA licence or rulebook for agentic payments, and the FCA has said it does not plan to introduce AI-specific rules, relying instead on its existing framework. Regulation follows the activity. Using AI in the customer journey does not take a firm outside the regulatory perimeter, and describing software as an agent does not make the payment activity unregulated.

If a business receives customer funds, operates payment accounts, initiates payments from accounts held elsewhere or executes payment transactions, it is likely to need FCA authorisation or registration under the Payment Services Regulations 2017 (the PSRs). If customers hold a stored balance, the Electronic Money Regulations 2011 apply. The agent itself is software and holds no licence; the business that provides the service is the one that may need permission.

Policy is moving. HM Treasury's consultation on modernising payment services regulation, which ran from 14 July to 6 October 2026, asks what changes are needed to authentication, consent and liability for unauthorised transactions when AI agents make payments. The FCA's 2026 Regulatory Priorities report for payments says it is working with the Treasury on whether rules are needed for agentic AI payments.

Which permission an agentic payment business needs

The answer depends on who contracts with the customer for the payment function, who can initiate transactions, whether the business touches customer money and which authorised firm executes the payment. The table gives the usual starting points.

Business modelLikely positionKey test
Agent recommends; customer approves each payment at their own bankUsually outside the PSRsThe business never initiates or executes the payment
Agent initiates payments from the customer's bank account through open bankingPayment initiation services (authorised payment institution, €50,000 capital)The business initiates payments at the customer's request
Agent reads account data to make decisionsAccount information services (RAISP registration)The business accesses accounts held at other providers
Business receives customer funds and pays suppliersMoney remittance or execution (authorised payment institution)Customer money passes through the business
Customers hold a balance the agent spendsElectronic money (EMI, €350,000 capital)Stored value is issued
Technology supplied to a bank, PI or EMI that provides the serviceOften outside direct regulationThe regulated partner contracts with the customer and controls the payment

Firms providing payment initiation or account information services also need professional indemnity insurance or a comparable guarantee. Our guide to the UK payment institution licence covers capital, safeguarding and the application, and the UK EMI licence page covers stored-value models.

Worked example

An AI procurement platform recommends purchases and sends the user to their bank app to approve each one: it is a software provider. The same platform, once authorised by the customer to pay suppliers automatically whenever stock falls below a threshold, using open banking to move money from the customer's account, is initiating payments. It needs payment initiation permission itself or must restructure so that an authorised provider contracts with the customer for that service.

Payment initiation and open banking

Agentic payments connect most closely with payment initiation services and open banking. In a conventional journey the customer instructs each payment and authenticates it with their bank. An agent may instead act under a broader mandate, initiating later payments when conditions are met and the customer is not present.

Three questions are open under the current rules: whether a broad mandate is sufficient consent for each later payment, when authentication must take place, and how much discretion an agent can have before the character of the service changes. Variable recurring payments, where a customer authorises a series of future account-to-account payments within agreed limits, are the closest existing model. HM Treasury's consultation proposes a new access right for payment initiation providers to lodge variable recurring payment mandates with account providers, and the open banking statutory instrument under the Data (Use and Access) Act 2025 is due by the end of 2026.

Using an authorised provider underneath does not by itself place a fintech outside the perimeter. What matters is the role the fintech actually plays in the customer relationship and the payment.

The PSRs treat a payment as authorised only where the customer has given consent in the form and procedure agreed. An agent acts on parameters rather than individual instructions, so the firm needs to show that the consent the customer gave covers each payment the agent makes.

  • Spending limits per payment, per day and per month.
  • Permitted counterparties, such as named suppliers or merchant categories.
  • Transaction types and periods the mandate covers.
  • Triggers for customer approval, such as a new payee, an unusual amount or a price above a set level.
  • Change and withdrawal: the customer can see, amend and cancel the mandate at any time, and withdrawal takes effect for future payments.

Worked example: a travel booking mandate

A customer authorises a travel agent to book a return flight to Madrid. A well-designed mandate records: the purpose (one return flight, London to Madrid); the dates (departing between 10 and 14 March); the maximum price (£400 including fees); the payment source (a named debit card or bank account); permitted sellers (airlines and named booking sites only); an approval trigger (any fare above £350 or any non-refundable fare goes to the customer first); and an expiry (the mandate lapses after one booking or on 1 March). Each payment the agent makes can then be checked against the mandate, which gives the firm the evidence of consent it needs and gives the customer a clear basis for any dispute.

An agent that stays within its mandate but acts in a way the customer could not reasonably anticipate can still create conduct and liability concerns. Mandates should be specific enough that the customer understands what they have authorised.

Strong customer authentication

Strong customer authentication (two independent checks, such as a device and a fingerprint) was designed for journeys where the customer interacts directly with their provider. An agent may pay when the customer is not present. The current requirements still apply, and AI-initiated payments are not exempt. Treatment depends on the journey and on whether an existing exemption or out-of-scope category applies, such as payments to trusted beneficiaries, recurring payments of the same amount to the same payee, or payments initiated by a merchant under an agreed mandate.

The likely direction is authentication of the mandate itself, with controls that set which payments can then proceed without repeated customer intervention. The Government has committed to replacing the technical standards in the PSRs with outcomes-based FCA rules, which is the route through which agent-specific authentication could be introduced. Products should be designed so authentication logic can be changed without rebuilding the service.

Liability when an agent gets it wrong

Liability is the hardest open question. An agent might order ten times the quantity needed after misreading inventory data, select a fraudulent merchant, act outside its mandate, or be manipulated by malicious data planted on a website. Potentially responsible parties include the customer, the AI provider, the payment firm, the bank, the merchant and the model provider.

What the current rules say

SituationPSRs positionAgentic question
Unauthorised paymentRefund by the end of the next business day (regulation 76); customer liability capped at £35 unless fraud or gross negligence (regulation 77)Is a payment outside the agent's mandate unauthorised?
Payment where the exact amount was not agreed in advanceRefund right if the amount exceeded what the customer could reasonably expect, claimed within 8 weeks (regulations 79 and 80)How does this apply to agents with price discretion?
Customer tricked into authorising a paymentAuthorised push payment fraud reimbursement for Faster Payments, up to £85,000Is a manipulated agent's payment within the scheme?
Late notice by the customerUnauthorised or incorrect payments notified within 13 months (regulation 74)Does the firm's monitoring alert the customer in time?

A payment can be validly mandated yet produce an outcome the customer never intended, which blurs the line the PSRs draw between authorised and unauthorised payments. Until the law is clarified, firms should set out responsibility in contracts with customers, AI providers and payment partners, and build dispute handling alongside the technology.

Know your agent

Payment firms and merchants increasingly need to know not only who the customer is but what is acting on the customer's behalf. "Know your agent" describes the controls that establish whether a payment came directly from the customer, from deterministic software the customer configured, or from an autonomous agent, and whether that agent holds genuine authority from the customer.

It is not a regulatory requirement today. It is likely to become part of payment infrastructure, through agent credentials, signed mandates and scheme rules, as agentic commerce grows. Identity and permission architecture may matter as much to a regulated firm as the AI model itself, because it is how the firm evidences consent and detects an agent that has been compromised.

Consumer Duty for agentic products

Where retail customers are involved, the Consumer Duty applies to the whole service, including how authority is obtained and how it is exercised, not only to whether the AI performs accurately. Customers relying on an agent to make decisions they would otherwise make themselves need particular protection.

  • Customers understand the authority they have given and how to change it.
  • Controls prevent foreseeable harm, such as overspending or paying unfamiliar merchants.
  • Alerts, spending summaries and explanations of the agent's decisions are clear and timely.
  • Cancellation is as easy as set-up.
  • Vulnerable customers and those with less digital confidence are considered in design and testing.

Outcome monitoring should include complaints about agent decisions, disputed payments, mandate changes and cancellations, so the firm can see where the agent is producing results customers did not expect.

Business and treasury agents

Many early agentic products serve businesses rather than consumers: procurement agents, accounts payable automation and treasury agents. The Consumer Duty does not apply to these customers, and the PSRs allow a payment firm and a business customer that is not a consumer, micro-enterprise or charity to agree that some of the information and liability provisions do not apply, known as the corporate opt-out.

That gives more contractual flexibility over mandates, refunds and liability, but the core questions remain. The firm still needs a clear record of the authority the business granted, controls that keep the agent within it, and a dispute process. Treasury agents that move funds between a group's accounts, or sweep surplus cash, should be checked against the permissions analysis in the same way as consumer products, because moving money for a client remains a payment service whatever the size of the client.

Safeguarding and financial crime

When the business holds customer money

Some agentic models hold funds before the agent spends them, for example a budget wallet the agent draws on. If the business receives customer funds for payment transactions, it is a safeguarding institution under the PSRs or EMRs and the FCA's CASS 15 rules apply: segregated safeguarding accounts, reconciliations at least once each reconciliation day, the monthly REP027 return within 15 business days and, above £100,000 safeguarded, an annual safeguarding audit. A stored balance the customer can redeem is likely to be e-money.

Anti-money laundering and sanctions

The Money Laundering Regulations 2017 apply to the regulated firm in the usual way: customer due diligence on the person or business that grants the mandate, sanctions screening of every counterparty the agent pays, and transaction monitoring. Monitoring should look at the agent's behaviour as well as the customer's, because an agent paying many new counterparties, or splitting payments, may indicate that it has been compromised or misused. Suspicious activity is reported through the money laundering reporting officer to the National Crime Agency.

Fraud when machines transact with machines

Agents can apply controls more consistently than people, but they create new attack routes. Criminals may manipulate an agent through planted data, compromise its credentials, alter its instructions or set up fraudulent counterparties for agents to find. Because agents can transact without human review and at machine speed, losses can mount before conventional controls react. Slowing the growth of fraud, including authorised push payment fraud, is part of the FCA's 2026 integrity priority.

  • Transaction and velocity limits set per agent, not only per customer.
  • Allow-lists of counterparties, with step-up authentication for any new payee.
  • Behavioural monitoring that compares the agent's activity with its normal pattern.
  • The ability to suspend an agent instantly, across every channel.
  • Use of the existing power to delay an outbound payment by up to 72 hours where authorised push payment fraud is suspected.

Third-party AI providers and governance

Agentic products depend on a chain of providers: model developers, cloud infrastructure, agent frameworks, payment application programming interfaces (APIs) and data sources. Outsourcing the model does not outsource accountability for the regulated service or for customer outcomes.

  • Identify which providers are critical to the payment service and plan for each one being unavailable.
  • Control model changes: test a new model version before it can make payments, and monitor for changes in behaviour after release.
  • Apply data security, access control and incident management to the agent as to any other part of the payment system.
  • Name who approves AI use, who monitors its performance and who can intervene.

From 18 March 2027, under FCA PS26/2, authorised payment institutions and EMIs notify the FCA of material third-party arrangements and submit an annual register, which will include material AI and cloud dependencies. Our article on AI governance for financial services covers the wider control framework, and the firm's AI governance page describes its support.

Operational resilience

Where an agent is part of how customers make payments, it is likely to be part of an important business service under the FCA's operational resilience rules (SYSC 15A), which apply to payment institutions and EMIs. The firm sets an impact tolerance for the service, maps the model, data and payment dependencies behind it and tests severe but plausible scenarios, such as the model provider becoming unavailable, an agent behaving erratically after an update or a payment API outage.

Two controls matter most. The first is a way to switch agent payments off instantly while leaving customers able to pay manually. The second is a defined process for reversing or resolving the payments an agent made while it was behaving incorrectly, with customer communication ready in advance.

Using a regulated payments partner

An AI fintech can often work with a regulated bank, payment institution, EMI or open banking provider that performs the regulated activity, while the fintech supplies the technology and customer experience. That can be efficient, particularly where the fintech never controls customer funds.

The structure must match reality. The fintech should not present itself as the payment provider, control the regulated activity or initiate payments independently outside its technology role. Contracts, customer journeys and technical architecture should all describe the same division. If customers contract with the fintech for the payment service and the licensed provider is largely invisible, the analysis may differ from a simple software arrangement. Where the fintech becomes the agent of the regulated firm, the agent must be registered with the FCA before it acts.

Before launch: a checklist

  1. Map the authority. Record what the customer authorises and what the agent can do without further human involvement, across the whole transaction.
  2. Assign each regulated activity. Identify which entity performs each one and check that customer contracts reflect it.
  3. Complete the permissions analysis before commercial agreements and integrations make changes difficult. An existing PI or EMI adding agentic features may need a variation of permission; our guide to FCA variation of permission applications explains the process.
  4. Design consent and controls into the product. Limits, counterparties, approval triggers, monitoring and withdrawal.
  5. Set the liability and disputes model in contracts and in the complaints process.
  6. Govern the AI as part of the regulated service, with owners for model changes, third parties, security, fraud, resilience and customer outcomes.

Where the framework is heading

The PSRs and existing FCA requirements apply today. There is no standalone agentic payments authorisation, no current know-your-agent requirement and no general exemption for AI agents from the rules on payment initiation, consent or authentication.

DevelopmentStatus
HM Treasury consultation questions on agentic payments (authentication, consent, liability)Consultation closed 6 October 2026; response awaited
FCA work with HM Treasury on whether rules are needed for agentic AI paymentsUnder way (2026 priorities)
FCA AI Live Testing and Supercharged SandboxRunning; evaluation report due by end of 2026
Outcomes-based FCA authentication rules replacing the technical standardsGovernment commitment; rules to follow
Open banking framework and variable recurring payments access rightStatutory instrument due by end of 2026

Products can be built under the current framework, but they should be designed to adapt as authentication, liability and agent identity rules develop. Our guide to the proposed UK payment services reforms covers the wider consultation.

How Buckingham Capital Consulting helps

The firm assesses agentic payment products from the customer journey up: perimeter analysis, whether direct FCA authorisation is needed, and how responsibility is divided with regulated infrastructure providers. It prepares authorised payment institution, payment initiation and EMI applications and variations of permission, reviews payment flows, consent and liability models, and supports FCA engagement. The analysis is most valuable before payment architecture and partnerships are fixed.

Frequently asked questions

Are agentic payments regulated by the FCA?

There is no separate licence. Existing rules apply by activity, so payment initiation, money remittance, payment accounts or e-money may need FCA permission even where an AI agent performs part of the service.

Does an AI agent need a payment initiation licence?

The agent is software and holds no licence. The business providing the service may need payment initiation permission, depending on who contracts with the customer, who initiates the payment and how the service connects to the customer's account.

Can an AI agent make payments without strong customer authentication?

Current authentication requirements apply and AI-initiated payments are not exempt. Treatment depends on the journey and on whether an existing exemption, such as trusted beneficiaries or recurring payments, applies.

Who is liable if an AI agent makes a wrong payment?

The PSRs set refund and liability rules for unauthorised payments, but they were not written for agents with discretion. Until the law is clarified, firms should allocate responsibility in contracts and build a clear disputes process.

Can a fintech use an authorised payment provider instead of becoming FCA authorised?

Often, if the regulated partner genuinely contracts with the customer and controls the payment. Licensed infrastructure does not place the fintech outside regulation if it independently performs a regulated activity.

Is know your agent a legal requirement?

No. It describes emerging controls to verify that a payment comes from an agent with genuine authority from the customer, and is likely to develop through industry standards and scheme rules.

What is the Government doing about agentic payments?

HM Treasury's payment services consultation, which closed on 6 October 2026, asked what changes are needed to authentication, consent and liability. The FCA is considering with the Treasury whether rules are needed.

References

  1. HM Treasury: Modernising payment services regulation (consultation, July 2026)
  2. FCA: Regulatory priorities, payments (March 2026)
  3. The Payment Services Regulations 2017 (legislation.gov.uk)
  4. FCA: Payment institution applicants
  5. FCA PS26/2: Operational incident and third party reporting
  6. Data (Use and Access) Act 2025 (legislation.gov.uk)

About Buckingham Capital Consulting

Buckingham Capital Consulting has obtained FCA payment institution licences, including open banking permissions for Oval, and e-money licences for firms such as Dzing Finance since 2013. It advises AI and fintech businesses on the permissions, consent and liability models agentic payment products need before launch.

Speak to an expert: WhatsApp +44 7563 115980, Telegram @buckinghamcapital or info@buckinghamcapitalconsulting.co.uk.

This article is general information, reviewed against the regulator’s published requirements on the date shown. It is not legal advice; the regulator’s current publications take precedence.