AI agents promise to find services, manage subscriptions and make payments on customers’ behalf. The UK’s FCA’s response raises a question that reaches well beyond banking: when software does the buying, who proves the customer agreed – and who puts things right?
TL;DR
- The FCA is preparing for agentic finance, applying existing rules to increasingly autonomous AI while developing AI tools to strengthen its own supervision. Accountability, human oversight and resilience are central.
- For telemedia, proving customer permission becomes critical: agents buying content, managing subscriptions or making payments create questions about identity, spending authority, liability and resolving disputed charges.
- The challenge is global, but regulation varies: EU requirements and international guidance reinforce the need for governance, while cross-border transactions make shared standards for consent and responsibility increasingly important.
Telemedia has spent years making it easier for people to buy things on their phones. The next challenge may be making it safe for their phones – or, more precisely, the AI agents running through them to do the buying.
Imagine asking an assistant to find the best way to watch a football match, organise mobile data for a trip or keep entertainment spending below a monthly limit. Finding an offer is one thing. Signing up, accepting recurring charges and choosing a payment method takes that assistant into altogether more consequential territory.
That is where agentic AI starts to become a payments story. It is also why developments at the UK’s Financial Conduct Authority (FCA) deserve attention across the international mobile content, billing, messaging and identity world.
In its 19 October statement of direction, delivered through a speech by chief data, information and intelligence officer Jessica Rusu, the FCA set out plans to work with industry, other regulators and frontier AI laboratories on a trusted framework for agentic finance. Its priorities include agent identity, delegated authority, accountability and execution.
Alongside that, it is developing its own capabilities as an “agentic regulator”: using AI to collect evidence, connect intelligence and identify potential harm.
These are distinct developments with a common consequence. Financial services businesses will face questions about how their AI behaves from a regulator becoming better equipped to examine the answers.
“The FCA’s plan to use AI more extensively in its own supervision will enhance its supervisory capabilities,” says Emma Banymandhub, CEO of The Payments Association. “As regulators become able to analyse information and identify risks at greater speed and scale, firms will need to make sure the governance behind their own use of AI is equally robust.”
Existing rules, more autonomous systems
The FCA’s approach is more nuanced than announcing a new set of AI regulations. Its published policy is to rely on existing frameworks rather than introduce additional AI-specific rules. Consumer Duty, governance requirements and relevant senior management responsibilities already provide tools for addressing harm.
In practice, this means the technology can change without removing the firm’s responsibility for the result. A business cannot reasonably treat an automated decision as somebody else’s problem simply because a model supplier built the underlying technology.
The direction follows July’s Mills Review, which examined how AI could reshape retail financial services by 2030 and beyond. Its recommendations included enabling agentic finance and developing AI-supported supervision. FCA-commissioned research accompanying the review suggested a fifth of UK adults were likely to use AI capable of acting autonomously within preset goals.
The regulatory response is therefore also an attempt to enable a market. Delegation could make managing money and services easier, provided consumers can understand what they have handed over.
“The FCA is right to focus on accountability, human oversight and operational resilience,” says Banymandhub. “Governance cannot be something firms retrofit once an AI system is already live. As AI becomes more autonomous, businesses need clear ownership, appropriate controls and plans for what happens when systems fail.”
The FCA’s new publication on good practice and areas for improvement gives that argument practical substance. It highlights defining the consumer need before building a tool, testing outcomes for different customer groups, monitoring deployed systems and deciding where AI should stop or escalate to a person. It also makes clear that its examples do not create new requirements. They illustrate how firms can approach responsibilities they already have.
Who authorised the purchase?
For telemedia companies, the most immediate issue is consent. A consumer might authorise an agent to spend up to £20 on entertainment. Does that cover a subscription, or only a single purchase? Can it accept a free trial that later renews? Can it choose a different payment method if the first fails? Does permission expire when the match finishes?
These are hypothetical examples, but they expose the difference between knowing who a customer is and knowing what that customer allowed software to do.
“This is particularly important for payments because the move towards agentic finance means AI will increasingly be able not just to provide information, but to act on a customer’s behalf, including managing and executing payments,” Banymandhub says.
“It must be clear who is liable when an AI agent initiates a payment. Existing UK frameworks assume a person authorises each purchase, so the industry needs a shared way to evidence customer intent to clarify responsibility and give merchants certainty. Questions around identity, delegated authority and accountability therefore need to be resolved alongside innovation, rather than after it.”
There is an important qualification to that description of existing frameworks. UK payment law already accommodates advance consent: regulation 67 of the Payment Services Regulations recognises consent to an individual transaction or a series of transactions. A fresh human action is not invariably required for every payment.
The emerging challenge is how those arrangements apply when software interprets broad instructions and chooses transactions, suppliers or timing. An instruction to “get me the best deal” leaves considerably more room for interpretation than a narrowly defined payment mandate.
A useful audit trail would therefore connect the customer’s instruction to the agent’s authority, the transaction and any subsequent change or withdrawal of permission. A successful authentication alone cannot answer every question about whether the agent stayed within its brief.
The payment networks are already moving
This is becoming an implementation problem as well as a regulatory debate. On 20 October, Visa, BILL and Billtrust announced a live pilot in which trusted AI agents complete approved invoice payments through supplier portals. The companies describe controls linking agent identity, tokenised credentials and payment execution to a buyer’s approved instruction.
“Businesses should not have to choose between automation and control when making payments,” says Mark Nelsen, Visa’s SVP and head of product, Commercial & Money Movement Solutions.
The pilot has a deliberately bounded remit: the agent executes eligible payments already approved by the buyer. It does not make the payment decisions.
That distinction is useful for telemedia companies considering where to start. An agent completing a clearly authorised transaction presents a different control problem from one independently choosing which service to buy.
Mastercard’s Agent Pay programme similarly combines registered, verified agents with tokenisation and consumer controls. At its launch, chief product officer Jorn Lambert described the need for “new merchant interfaces to distinguish trusted agents from bad actors using agentic technology”.
Visa’s Trusted Agent Protocol specifications address a related problem: helping merchants recognise approved agents using verifiable, signed information.
These initiatives show how parts of the trust infrastructure can work. They do not, by themselves, settle every question of UK liability or establish a universal standard across cards, wallets, account-to-account payments and carrier billing.
Why DCB businesses should care
For direct carrier billing providers and aggregators, the implication is to consider how delegated authority could travel through the purchase journey. A phone number, recognised device or authenticated subscriber can help establish identity. Evidence of permission to buy a particular service requires its own attention.
Regulatory scope also matters. The FCA’s Handbook describes an electronic communications exclusion for certain qualifying transactions charged to a communications bill. The position depends on the activity and applicable conditions; the FCA’s announcement should not be read as bringing every mobile content business under financial services supervision.
Even outside direct supervision, however, commercial pressure could spread through the supply chain. Regulated payment partners may want suppliers to demonstrate how agents are authenticated, spending permissions are enforced and disputed transactions are reconstructed. That is a likely business implication, rather than a new universal FCA obligation.
Messaging companies could have a role in making delegation understandable: communicating spending limits, seeking confirmation when an agent reaches a boundary and providing clear receipts or ways to withdraw permission.
There is a balance to strike. Requiring approval at every step would erode the convenience consumers were promised. Allowing vague, enduring permissions would make it harder to explain an unexpected charge.
Identity and fraud specialists such as Evina face a related opportunity. Services will need to distinguish authorised automation from account takeover, impersonation or other abusive traffic. Blanket bot blocking could reject legitimate business; blanket acceptance would create obvious weaknesses.
Content providers may also have to make offers easier for agents to interpret. Price, renewal terms, eligibility and cancellation conditions need to survive the journey from a merchant’s offer to an assistant’s recommendation. Otherwise, an apparently successful sale could become an expensive dispute.
The difficult part comes after launch
A further challenge is change. An agent’s behaviour can be affected by model updates, new tools, altered data or interactions with other systems. Passing an initial test does not establish that it will remain dependable indefinitely.
In her published speech, Rusu described the need for “continuous, conditional assurance”. The phrase captures a demanding operational task: keep checking that the system remains safe under the conditions in which it is allowed to act.
The uncertainty is acknowledged at the top of the regulator. Asked by UK government MPs on 13 October about assurances over increasingly autonomous systems, FCA chief executive Nikhil Rathi said: “As it stands right now, I do not think we have the assurances that you are asking for.”
For smaller telemedia businesses, the practical challenge will be proportionate implementation. They may depend on external models and payment providers while having limited influence over either. Clear supplier responsibilities, restricted permissions, usable records and a workable way to stop transactions become commercially important.
Human oversight also has to be credible. Someone needs both the information and the authority to intervene. A complaints team faced with an unexplained agent purchase needs a route to investigate and resolve it.
“With the FCA and Bank of England’s latest survey showing that 86% of financial services firms now use AI, this is no longer an emerging issue,” Banymandhub says. “The FCA’s pragmatic approach is welcome, but industry and regulators now need to move quickly together to give businesses and consumers clarity over where responsibility sits as AI becomes a more active participant in financial services.”
Rusu announced that figure ahead of the survey’s publication, scheduled for the week beginning 26 October. It concerns AI use generally, rather than demonstrating that 86% of firms operate autonomous payment agents.
For telemedia, the opportunity remains considerable: easier discovery, simpler subscription management and fewer obstacles between wanting a service and accessing it. But the industry’s experience of disputed charges offers a useful lesson. Removing friction from a payment only creates lasting value when the customer can still understand, control and challenge what has happened.
The next generation of mobile commerce will need to make buying easier for agents while keeping responsibility clear for people.
A global challenge, different regulatory approaches
The FCA’s initiative is British, but the questions it raises is cross-border. An AI agent could be developed in one country, act for a customer in another and buy through a payment provider based somewhere else. Establishing permission and resolving disputes will require cooperation well beyond any single regulator’s remit.
In the EU, the AI Act sits alongside existing financial services legislation, with obligations depending on the system’s purpose and the firm’s role. Certain applications, including assessing individuals’ creditworthiness, are classified as high-risk; an agent handling payments is not automatically in that category simply because it uses AI.
Internationally, the Financial Stability Board published 12 sound practices for responsible AI adoption on 15 October, covering governance, human oversight, ongoing monitoring and dependence on external providers. These are practices financial institutions are encouraged to consider, rather than a binding global AI rulebook.
For telemedia businesses serving multiple markets, the practical implication is that a single technical integration may encounter different regulatory expectations. Evidence of customer intent, clear limits on delegated spending and records that travel across the transaction chain could become valuable commercial infrastructure.
The opportunity is global; agreement on who takes responsibility is still catching up.
