As lawmakers grapple with the alleged 1/10 possibility that artificial intelligence will exterminate humanity within a decade, financial regulators are already tackling the fundamental contradictions between ‘deterministic’ payments and ‘probabilistic’ AI, as well as the scope for new types of malfeasance that will come with revolutionary convenience and efficiency. Do these worries all belong in the same category? Almost certainly not. Sensible conversations between regulators and the payments industry are well underway and AI-driven systemic risks have familiar shapes. Painful decisions, though, await sovereigntists outside the US and China.
Banal payment processes given over to decision-making machines throw up complicated regulatory conundrums. When a payment is delegated to an agent that is following a user’s objective rather than a set of precise accountable instructions, who is liable for each of the individual decisions and actions? Regulators grappling with decentralised blockchains will immediately recognise the problem typically summarised with the question: who do I call when it goes wrong?
Policy-makers are contending with three sets of risks and opportunities: practical, systemic and geo-strategic. Benefits range from convenience, speed, efficiency, productivity gains and, potentially, much-needed greater competition in financial services and payments pricing. It may also ameliorate the compliance burden which has arguably reduced the capacity of traditional cross-border finance rails. Risks, however, include turbo-charged fraud, legal ambiguity, loss of trust in digital money, flash crashes, cybersecurity risks in critical infrastructure, dependency on suppliers based in partly or completely hostile countries and loss of monetary sovereignty.
New convenience, old risks
Prime candidates for AI in payments tend to be mundane and consequently ideal targets for convenience, such as the further automation of small and frequent transactions. More broadly, AI is likely to become a shopping assistant, finding you the best product, investment or service from the optimal retailer within the parameters of your specified intentions and preferences, including price range.
Chinese drivers can already voice requests to their in-car WeChat app to activate purchases, including charging, for example. A more adventurous application would see AI pre-empt your wishes based on deep knowledge of your habits, motives, needs and desires.
What could go wrong? Sonja Davidovic and Hervé Tourpe of the International Monetary Fund have identified perhaps the most fundamental issue in a recent paper: ‘payment systems are built on deterministic infrastructures that guarantee predictability, accountability, and legal settlement finality. Agentic AI systems, by contrast, rely on probabilistic reasoning and adaptive decision making.’ AI, as delivered by large language models, is an essentially stochastic process. Decisions are probabilistic, which introduces a degree of uncertainty that does not fit well with the requirement for reliability in payments.
Agentic, but not too much
Davidovic and Tourpe propose to solve this by disaggregating the agentic payments process into three layers that would separate ‘probabilistic decision making from deterministic authorisation and settlement’. That is, while AI can improve the choices of what payments to make and when, interpreting intent and coordinating action, the actual downstream payment and settlement process ‘preserves rules-based authorisation and legally final settlement with a middle layer to supervises and mediates between the two.
Once the proper role for AI in payments is defined, another problem emerges. According to Sarah Breeden, deputy governor of the Bank of England: ‘The biggest issues are likely to be… how users securely give consent and authorisation to agents, especially for multiple transactions; how disputes are settled and liability assigned for erroneous or fraudulent transactions; and how authorities avoid fragmentation and walled gardens as AI firms and payment systems all develop protocols for agents to interact with themselves and merchants’. Except the first, these are all familiar-sounding risks to fintech regulators.
Industry players are already tackling the potential ambiguities, mismatches and liability gaps along these lines. MasterCard and Google are already introducing ‘Verifiable Intent’, an industry standard which cryptographically codifies the specific authorisations of the purchaser and documents the interaction between the AI agent and the merchant. ‘Know Your Agent’, extending the logic of know-your-customer, is also quickly becoming a cornerstone notion.
Merchants, rather than consumers, are sooner the focus of industry worries about AI-driven payments. ‘They might not accept agent-initiated orders, because they cannot verify the agent understood the customer’, Kun Tian of Kent Business School told OMFIF. ‘One e-commerce operator now telephones buyers to confirm before shipping, which destroys the efficiency case entirely.’ China, the source of these examples, does not yet have ‘autonomous payments’. ‘It is a very fast pre-filled checkout with a human at the end, confined to a narrow band of low-value repeat purchases,’ said Tian. ‘The question worth watching is who removes that confirmation step first, and under what competitive pressure.’
OMFIF has heard supervisors worry that AI may also compound a familiar stablecoin headache, which is: who owns the end-to-end process from a regulatory point of view? The broader policy challenge is ‘ensuring that greater autonomy does not come at the expense of trust, which’, according to Wolfram Seidemann, chief executive officer of Giesecke+Devrient Currency Technology, ‘can be eroded when decisions become opaque, accountability becomes unclear, or users lose confidence that transactions are being executed in their best interest. As financial interactions become increasingly autonomous, transparency, oversight and clear responsibility will be essential to ensure that innovation strengthens, rather than weakens, trust in the financial system’.
Systemic risks
The familiar risk that technological innovation brings with it the temptation to monopoly is already well understood. In the US, the proposed AI Agent Act would prevent major players from creating walled gardens via AI agents favouring their own products. One industry participant suggested that a major AI player had already given up trying to create a walled garden for payments ‘because it was too difficult’.
A more intriguing conundrum is how to manage the risk that AI agents independently make the same financial decision at the same time, en masse, exacerbating price or market volatility. The Mills Review, produced for the UK’s Financial Conduct Authority, points to the ‘correlated risks and common points of failure across the sector’. This, also, is not a new problem, as the 2010 flash crash demonstrated. A side-effect of their potential increasing prevalence, though, is the ‘normalisation of deviance’, as set out in a 2019 World Economic Forum paper by Jesse McWaters, now global policy head at MasterCard, where market participants learn to ignore potential indicators of a ‘tipping point’ from an event which would then have an outsized market impact.
Elsewhere in the journal, Agio Ratings recommends that ‘to safeguard the wider payment system, authorities may need monitoring tools and triggers capable of flagging correlated, machine-speed behaviour across many institutions at once. In addition, they would also have mechanisms to interrupt contagion that is irrational or disproportionate before it can cascade through the financial system’. In short, the means to supervise and override the agency of agentic AI will, as in other fields, remain a paramount and, no doubt, state-managed, if not operated, task.
The technological infrastructure for this is not yet built and unlikely to be cheap. As tokenised finance discussions have already shown, this may involve bartering between the public and private sectors about whose financial responsibility ‘regtech’ is. Beyond finance, ‘kill switch’ bill proposals are emerging in the US and UK. The logic of a zero-sum cyber arms race in a Hobbesian world may ultimately reverse the networking of machines. ‘I can imagine some countries simply deciding that critical systems can no longer be online at all’, says Cal Mitchell, researcher at the Centre for Financial Technology at Imperial College London.
The IMF’s Davidovic told OMFIF, however, that autarky ‘is not a realistic defence’, and that countries that don’t have hyperscalers will have to accept for now that they will partly depend on those who do.
Geo-strategic risks, blockchain and AI
A more intriguing policy consideration is the risk of non-adoption. Regulation is already causing potentially unintended frictions. Secure Customer Authentication rules impede the deployment of AI in card transactions in Europe. A broader risk is that the integration of AI into blockchain platforms via the convenient x402 protocol will compound the extraterritorial advantage that US players – and the dollar – already have from the proliferation of stablecoins and US crypto exchanges, as Erwin Voloder, director, research and policy at Blockchain4Europe, suggests.
The economics of the AI ecosystem may well have an equivalent bearing on currency as energy has done. AI users will have to purchase tokens. By and large, these are mostly priced in dollars. Will ‘AI dollars’ replace or indeed compound Petrodollars? ‘If the answer is yes’, Voloder told OMFIF, ‘then Europe’s problem and by extension MiCA’s problem isn’t the stablecoin plumbing but that Europe just isn’t selling the inputs agents are buying’.
Chinese companies are already players in this field, too. ‘Four of the six scenarios in Alipay’s own developer taxonomy are AI-native: model subscriptions, token top-ups, per-call API metering, and creative tools priced per use. This is the one place’, according to Kun Tian, ‘the economics compel a new rail, because at per-call settlement conventional processing costs more than the transaction is worth’. China is managing to do this without blockchain-based stablecoins, which are banned, though it does have ample recourse to AI capacity. India, which also foreswears stablecoins as legal tender, is preparing the introduction of agentic AI into its Unified Payments Interface infrastructure. Critically, the operator, NPCI, a consortium including the central bank and commercial banks, is building a central registry to manage ‘KYA’.
What of Europe? It has made unsung progress in the development of tokenising money and securities, but has both a lack of native capacity and the nature of its regulation against it. As Europe pursues strategic autonomy in these technologies, it may need to revisit elements of its well-intentioned AI Act, not to mention hope that public and/or private investment can be found to scale up quickly. In the meantime, as with tokenisation, policy-makers may need to accept walking the path to development in American shoes for now.
John Orchard is Chairman of the Digital Monetary Institute at OMFIF.
Interested in this topic? Subscribe to OMFIF’s newsletter for more.