How will Agentic AI change payments and finance?

I was recently interviewed for a podcast about Agentic AI in payments and finance. Here’s how the interview went:

You’ve been writing and speaking about agentic commerce for a while. When did it click for you that this wasn’t just another tech buzzword. What was the moment you thought, “this one is real”?

The answer depends on whether you mean the idea, the research field or the term “Agentic AI.” They are three different things.

The idea of Agentic AI is much older than ChatGPT. AI researchers were talking about intelligent agents as far back as the 1980s and 1990s. One of the most influential papers was by Michael Wooldridge and Nicholas Jennings, who defined intelligent agents as autonomous systems that perceive their environment, make decisions and act to achieve goals. The Belief-Desire-Intention (BDI) architecture also emerged during this period as a framework for autonomous agents. In other words, the concept of AI that could plan, reason and act has been around for more than thirty years. The technology simply wasn’t powerful enough to make it practical.

The modern breakthrough came with large language models, LLMs.

GPT-3 in 2020 showed that language models could reason surprisingly well. ChatGPT in late 2022 made those capabilities accessible to everyone. Then, in 2023, developers began connecting LLMs to tools, APIs and memory through projects such as Auto-GPT, BabyAGI and LangChain. Suddenly an AI could do more than answer a question. It could break a goal into tasks, call external tools, evaluate progress and continue until the job was complete. That was the practical birth of today’s Agentic AI.

And then the term “Agentic AI” is actually much newer. It only entered mainstream technology vocabulary in 2024, largely through Andrew Ng, who deliberately popularised the phrase to distinguish systems that exhibit varying degrees of autonomy from the older, binary notion of an “AI agent.” Rather than asking whether a system is or isn’t an agent, Ng argued we should think of AI on a spectrum of agency. That is why it feels as though “Agentic AI” appeared overnight. The research did not. The label did. In fact, if you were drawing a timeline, it would look something like this:

• 1980s–1990s: Intelligent agent theory, autonomous software agents and multi-agent systems become established areas of AI research.
• 2020: GPT-3 demonstrates powerful language reasoning.
• Late 2022: ChatGPT brings LLMs into the mainstream.
• 2023: Auto-GPT, BabyAGI, LangChain and similar frameworks demonstrate autonomous task execution using LLMs.
• 2024: Andrew Ng popularises the term “Agentic AI.”
• 2025–2026: Every major technology company, from OpenAI and Anthropic to Google, Microsoft, Amazon and Nvidia, pivots from talking about chatbots to talking about agentic systems that can plan, act and collaborate autonomously.

The more interesting observation is that generative AI may end up being remembered as just a stepping stone. ChatGPT taught us that AI could answer questions. Agentic AI is teaching us that AI can complete work. The leap from generating text to achieving goals is arguably as significant as the leap from search engines to informed answers. It changes AI from being a tool you use into a colleague you delegate to, and eventually into an autonomous participant in the economy.

That is why, although the term is only a couple of years old, many people already see Agentic AI as the beginning of the next era of computing.

Every major card scheme, Visa, Mastercard, Stripe, is racing to embed card credentials as the default payment method for AI agents. Do you think cards are the most suited rail or are their other ways to pay that would work better in this context?

It’s a fascinating question because it forces us to separate what works today from what is optimal in the long term.

The simple answer is that cards are the best payment rail for Agentic AI today, but probably not the best payment rail for agentic commerce tomorrow, because cards were designed for humans.

When Visa, Mastercard and Stripe talk about agentic commerce, they are solving an immediate problem. If an AI assistant books your hotel, orders groceries or renews your software subscription, it needs a payment credential. The easiest solution is to give the agent a virtual card or tokenised card credential, because the global acceptance infrastructure already exists. Every merchant already accepts it. Every fraud engine understands it. Every dispute process is defined. Every issuer knows how to authorise it.

In other words, cards solve the distribution problem.

That is enormously valuable. It explains why the card schemes have moved so quickly. They don’t need merchants to change anything. They simply replace the human holding the phone with an AI holding a token.

But once you imagine millions or billions of autonomous agents trading with each other, the weaknesses of cards become more obvious.

Cards assume a human purchase. They assume someone is present to authenticate, to approve, to dispute a charge if something goes wrong. They were built for relatively infrequent transactions of meaningful value. Agentic AI flips all of those assumptions as they are being built for massive volumes of transactions per second of almost no value.

Put it this way: imagine your household AI negotiating electricity prices every five minutes, buying compute power by the second, purchasing data for milliseconds, selling excess solar energy, paying congestion charges dynamically or continuously rebalancing your investment portfolio. Suddenly you aren’t making twenty card transactions a month. Your personal AI could be making twenty thousand transactions an hour.

Cards start to feel like using a cheque book to power the internet, and the more likely future is that different payment rails become specialised.

Cards remain excellent for retail commerce, because acceptance is universal and consumer protection is strong. If your AI books your holiday, buys your groceries or renews Netflix, cards are probably still the best answer.

Account-to-account payments become increasingly attractive where both parties are known, and trust already exists. Real-time payment systems such as Faster Payments, UPI, Pix and FedNow eliminate many of the costs associated with card interchange and settlement. They are also programmable in ways that fit autonomous commerce more naturally.

Tokenised bank deposits become particularly interesting because they combine commercial bank money with programmability. An AI agent could move tokenised deposits between banks instantly whilst embedding business logic directly into the payment itself. Payment and settlement effectively become the same event.

Stablecoins arguably fit agent-to-agent commerce even better. If two software agents are negotiating globally, twenty-four hours a day, across borders, a programmable digital dollar that settles within seconds without correspondent banking has obvious advantages. This is precisely why stablecoins are attracting so much attention, not because consumers want them, but because software does.

Then there is an even more radical possibility where agents won’t choose payment rails at all.

Humans need to control their payment flows and use trusted brands, whilst software has no such loyalty because an Agentic AI will optimise for success rate, settlement speed, transaction cost, liquidity, regulatory constraints, merchant preference, fraud risk and tax implications. It won’t ask, “Should I use Visa or a stablecoin?” It will ask, “Which combination minimises cost whilst maximising certainty?”

This is algorithmic programmed trading for everyone on Earth, not just the flash trading investment markets, and it means that payments become an optimisation system rather than a brand decision.

Ironically, this may be the biggest strategic challenge facing the card networks. Their historical advantage has been consumer preference and merchant acceptance. In an agentic economy, the customer making the payment decision is no longer a human. It is an algorithm. Algorithms are ruthlessly rational. They don’t care about rewards points, premium branding or metal cards. They care about probability of execution.

This is why I suspect Visa, Mastercard and Stripe understand that they are no longer competing as payment rails but as orchestration platforms for autonomous commerce. Their real value lies less in moving money and more in providing identity, authentication, delegated authority, fraud detection, liability management, compliance and trust between agents.

In that world, the payment rail itself becomes almost invisible.

Agentic AI decides how to pay, whilst the human decides the rules and the platform guarantees the trust.

That’s a much bigger shift than simply embedding card credentials into an AI assistant. It suggests that the winners of agentic commerce won’t be those with the fastest payment rails, but those with the most trusted infrastructure and that, perhaps, is the crucial insight.

The future of payments is unlikely to be card versus account-to-account versus stablecoin. It is much more likely to be AI selecting the optimal rail for every individual transaction. The payment method becomes dynamic, contextual and almost entirely invisible, which is exactly where intelligent finance has been heading for years.

Worldpay and Trulioo are building something called a “Digital Agent Passport.” Visa and Mastercard are developing “Know Your Agent” frameworks, with cryptographic signatures and network tokens. Is this the right approach, or are we just bolting a new layer onto broken foundations?

This is the question nobody is asking, so thank you for asking.

The payments industry has an extraordinary habit of solving tomorrow’s problems by extending yesterday’s architecture. It worked with magnetic stripes becoming EMV chips; cards becoming tokenised cards; e-commerce becoming 3-D Secure. Each generation has added another layer of trust without fundamentally changing the underlying model.

The question is whether Agentic AI represents another incremental evolution or a complete architectural break, and my instinct is that it is the latter.

The idea of a Digital Agent Passport or a Know Your Agent (KYA) framework makes perfect sense in the short term. If my AI is going to spend my money, negotiate contracts, book travel or hire another AI to perform work, then everyone involved needs confidence that the agent is genuine, authorised and operating with delegated authority. Cryptographic identities, signed credentials, network tokens and verifiable attestations are all sensible building blocks. They answer obvious questions. Who created this agent? Who is it acting for? What is it allowed to do? Has it been tampered with?

Those are exactly the same questions banks have been asking about people for decades. The problem is that agents are not people.

A customer might authenticate once every few hours. An autonomous agent may authenticate thousands of times every second whilst collaborating with dozens of other agents across multiple jurisdictions. It may create temporary sub-agents to complete specialised tasks before deleting them minutes later. It may delegate authority dynamically depending upon context, risk and value. None of this fits comfortably into identity frameworks originally designed for human beings opening bank accounts.

That is why I worry that Know Your Agent risks becoming Know Your Customer with different labels.

The financial industry has always viewed identity as a static thing. You prove who you are, receive credentials and then present them when required. Agentic AI turns identity into something far more fluid. An agent’s identity is not simply who created it. It is also its code version, the models it is using, the data it has accessed, the tools connected to it, the permissions currently delegated to it, the policies governing its behaviour, the audit trail of its decisions and perhaps even its reputation built over millions of successful interactions.

That is no longer about identity. It is all about provenance.

The really interesting question therefore is whether we should be authenticating the agent at all, or authenticating the entire chain of trust surrounding the agent.

Imagine two AIs negotiating a million-dollar trade. Neither really cares about a digital passport. They care whether the other agent has authority to commit funds, whether its principal remains solvent, whether its software has been compromised, whether its recommendations have been independently verified and whether the transaction complies with the legal frameworks governing both jurisdictions. In other words, trust becomes continuous rather than a specific point-in-time.

This is where I think the industry needs to think much bigger than Digital Agent Passports.

Every agent should have a persistent cryptographic identity. Every action should be signed. Every delegation of authority should be verifiable. But these are merely the foundations. What matters far more is creating a living trust fabric that continuously measures confidence, rather than simply proving identity.

Think about how humans build trust. We don’t trust someone because they show us a passport. We trust them because of repeated behaviour, reputation, recommendations, professional qualifications, legal accountability and experience. A passport establishes identity. It does not establish trust.

Agentic AI will need exactly the same distinction.

An AI should carry not only a cryptographic identity but also an evolving reputation, a verifiable history of successful actions, independent attestations of competence, evidence of compliance, proof of delegated authority and perhaps even an insurance policy guaranteeing the consequences of its decisions. That is a much richer concept than a digital passport.

Which brings us back to payments.

Visa, Mastercard, Worldpay and Trulioo are solving the immediate problem because they have to. The market needs something that works today, and Digital Agent Passports are an entirely rational first step. But I suspect history will look back on them in the same way we look at passwords today. Necessary, useful and ultimately transitional.

The deeper challenge isn’t identifying autonomous agents. It’s creating an economy where autonomous entities can establish trust without humans constantly standing behind them. That requires us to move beyond Know Your Customer and Know Your Agent towards something far more powerful: Know Your Trust.

In the end, agents won’t transact because they possess a credential. They’ll transact because they can continuously prove they deserve to be trusted. That is a fundamentally different architecture, and one that is likely to define the financial infrastructure of the AI-native economy.

If my AI agent is loyal to me, and not to a card or a brand or a platform, and it’s optimising purely on my behalf, what happens to loyalty programmes, advertising, product discovery? Will the consumer-brand relationship fundamentally be broken?

I think this is where people are dramatically underestimating the impact of Agentic AI.

Everyone is focused on whether AI can book a flight or order groceries. That’s interesting, but it isn’t revolutionary. The revolutionary part is what happens when the customer disappears.

For the last hundred years, almost every consumer business has been built around one assumption: the human makes the decision.

Advertising persuades the human whilst brands try hard to build emotional connections with the human. Loyalty programmes reward the human whilst search engines help the human discover products. Retailers optimise the experience for the human and entire industries worth trillions of dollars exist because influencing human behaviour is incredibly profitable.

What happens when the decision-maker is no longer a human?

Imagine telling your AI, “Whenever I need coffee, buy the best quality beans under £20, prioritise sustainability, delivery within 24 hours and avoid companies with poor labour practices.”

You never visit a website, see an advert or compare brands and collect points. Your AI simply executes your preferences. That single change potentially dismantles the economics of digital marketing.

Today, companies compete for your attention. Tomorrow, they compete for your algorithm’s approval.

That is a completely different market.

An AI doesn’t care about celebrity endorsements. It doesn’t care about emotional storytelling or beautiful packaging. It won’t buy because Taylor Swift wore the trainers or because an influencer recommended a skincare product. It won’t be persuaded by scarcity messages, countdown timers or “limited edition” offers. Those techniques exploit human psychology. Software has no ego, no fear of missing out and no desire for social status.

Instead, agents optimise price, quality, reliability, carbon footprint, certainty of delivery, return rates, verified customer outcomes, long-term value and more. It means marketing starts looking much more like structured data than advertising and attention. That doesn’t mean brands disappear. It means the purpose of brands changes.

Historically, brands reduced uncertainty. You bought Nike because you trusted Nike. You stayed with American Express because you trusted American Express. You paid more for Apple because you believed Apple delivered a better experience. But if your AI can analyse ten million independent reviews, verify product provenance, assess manufacturing quality, compare lifetime cost, negotiate discounts in real time and monitor product performance after purchase, does it still need the shortcut called a brand?

Perhaps not.

Trust becomes measurable rather than emotional.

This is why I think loyalty programmes face an existential challenge as the entire loyalty industry assumes that humans are irrational enough to alter their behaviour in exchange for points, cashback, upgrades or status. Airlines, hotels, supermarkets and card issuers have spent decades perfecting behavioural economics.

Your AI couldn’t care less about Gold Elite Platinum Diamond status. It cares about maximising your utility function. If another supplier delivers better value today, it switches instantly without emotion, inertia or guilt. Loyalty becomes dynamic instead of permanent.

Ironically, this could create more competition than we’ve ever seen before. Companies have long relied on customer friction to protect market share. People don’t switch banks because it’s inconvenient. They don’t change insurers because it’s tedious. They don’t move broadband providers because life gets in the way.

An AI has none of those limitations.

If switching saves £37.42 every month whilst maintaining the same service level, it switches before you’ve even made breakfast.

That should terrify incumbent businesses, but here’s where it gets even more interesting. If every consumer has an AI buying agent, then every business will have an AI selling agent. You no longer have business-to-consumer. You have agent-to-agent.

Your AI negotiates with the airline’s AI; your insurer’s AI negotiates with your purchase history AI; your home’s energy AI negotiates continuously with the electricity grid. Commerce becomes an ongoing negotiation between autonomous systems rather than a series of isolated purchases. That changes advertising completely because brands will stop advertising to people altogether. Instead, they will focus on publishing machine-readable trust scores, product specifications, sustainability credentials, delivery guarantees, service-level agreements and dynamic pricing APIs. Marketing becomes less about storytelling and more about providing structured evidence that algorithms can evaluate and decide which one fits their human owners needs the best. The winners won’t be the companies with the best television commercials. They’ll be the companies whose data is the cleanest, whose reputation is the strongest, and whose products consistently outperform when judged by machines acting in their customers’ interests.

There’s one final twist.

People often assume your AI will be loyal only to you. I’m not sure that’s enough. The real battle will be over who trains your AI’s preferences. If your agent runs on Google, Amazon, Apple, OpenAI, Anthropic or Tencent infrastructure, whose definition of “best” is it using? Which products does it trust? Which merchants does it rank first? Which sources does it believe? Which reviews does it consider credible? Every recommendation engine embeds assumptions, incentives and biases. Agentic AI won’t eliminate that problem. It may amplify it.

That’s why the next great competition won’t be over search engines or social media feeds. It will be over decision engines. The companies that shape how autonomous agents evaluate value will become even more powerful than the companies that once shaped what humans saw on a screen.

In that sense, the consumer-brand relationship doesn’t disappear. It is disintermediated. The emotional conversation between a person and a brand is replaced by a computational conversation between two algorithms. Brands no longer compete for attention but compete for recommendation, and the most valuable asset in the economy may no longer be the customer’s loyalty but the AI agent’s trust.

That is a profound shift, because for the first time in modern commerce, the customer may stop shopping altogether because their AI will do it for them. Then question is not whether brands can persuade people, but whether they can persuade the intelligence that people trust to decide on their behalf.

Could we end up in a world where the default payment method for agents is something entirely outside the regulated banking system?

I think we could, but the more interesting question is why we assume that autonomous agents would naturally choose the same financial infrastructure that humans built for themselves.

Banking exists because humans have problems. We forget passwords. We get scammed and we die and divorce. We need credit and we need consumer protection. We need regulators to intervene when things go wrong.

Almost every layer of the financial system exists because humans are imperfect.

AI agents have a very different set of requirements.

They don’t sleep. They don’t forget. They don’t suffer from fear of missing out. They don’t panic in a market crash. They don’t care whether the payment is made with a Visa card, a Faster Payments transfer or a stablecoin. They simply optimise for whatever achieves the objective with the lowest cost, the highest certainty and the fastest settlement.

That changes everything.

Today, we think of payment methods as products. Cards compete with bank transfers. Wallets compete with cards. Stablecoins compete with bank money. An AI doesn’t think like that. It purely has an algorithm set to make payments based upon the best routing decisions according to their owner’s rules.

If a regulated bank transfer is cheapest, it will use that. If a stablecoin settles instantly across borders with lower risk and lower cost, it will use that. If a tokenised bank deposit can be delivered immediately with a tokenised asset, it will use that instead. The rail becomes a technical choice, not an emotional one.

That is why I suspect the future will be far more pluralistic than today’s debates suggest.

Having said that, there is one area where the regulated banking system genuinely has a problem.

The banking system evolved around legal entities. Every account belongs to a person, a company or an organisation and every payment ultimately has a legal owner. Autonomous agents don’t fit neatly into that framework. They may be created and destroyed in seconds with hundreds of specialist sub-agents that can act simultaneously across multiple jurisdictions. Trying to fit them into twentieth-century account structures may prove painfully inefficient.

This is where digital bearer instruments become interesting.

If an agent can hold a cryptographically secured asset that it can transfer instantly, without waiting for correspondent banks, operating hours or reconciliation cycles, then from a purely engineering perspective it is a cleaner solution. Stablecoins are an obvious example, although not the only one. Tokenised deposits, central bank digital currencies and entirely new forms of programmable money could all compete for that role.

The key point is that the agent doesn’t care whether the money originated inside or outside the banking system. It only cares whether the payment completes safely and efficiently.

That leads to an uncomfortable thought for banks as banks, for decades, have assumed that payments belong to banking, but perhaps payments belong to whoever provides the best trust infrastructure.

If an AI agent is choosing autonomously, it may place greater value on programmable settlement, machine-readable rules, cryptographic proof, continuous availability and instant finality, than on whether the underlying asset sits on a commercial bank’s balance sheet. That doesn’t necessarily favour crypto, but it favours infrastructure that was designed for software rather than humans.

This does not mean that banks should panic as money is not just technology. It is law.

Every significant financial transaction eventually collides with the legal system. Someone owns the asset, pays the tax, carries the liability. Someone has to resolve disputes, satisfy anti-money laundering rules, sanctions requirements and court orders. Those responsibilities don’t disappear simply because software is making the decisions.

This is why I don’t believe we’ll see agents abandoning the regulated financial system altogether. Instead, I think we’ll see regulation migrate closer to the software itself.

Rather than regulating only banks, regulators may increasingly certify AI agents, define delegated authority, mandate audit trails, require cryptographic identity, impose behavioural standards and specify liability when autonomous systems make mistakes. The focus shifts from regulating institutions to regulating autonomous economic actors.

That would be a profound change.

For centuries, regulation has concentrated on organisations because organisations employed people. In an agentic economy, regulators may need to oversee software that acts with delegated authority on behalf of those organisations and individuals. This means that the future isn’t inside or outside banking. It’s alongside banking.

Banks will continue to create regulated money, extend credit and provide the legal and prudential framework that underpins the economy, but the movement of value may increasingly occur over infrastructures that are optimised for autonomous software rather than human beings.

The irony is that this would echo the history of the internet itself. Banks never owned the internet, but they learned to operate on top of it. In the same way, banks may not own the networks through which AI agents transact. Their role may instead become providing the trust, liquidity and legal certainty that those networks ultimately depend upon. If that happens, the biggest competitive threat to banks won’t be another bank. It will be the emergence of a new economic protocol that software agents adopt because it is simply a better language for machines to exchange value.

Can you give an example of this new economic protocol?

A good example came from Google’s autonomous vehicle work around ten years ago. The work wasn’t actually about self-driving cars but about self-owning cars. The vision, discussed by engineers involved in the autonomous vehicle ecosystem, imagines a robotaxi that doesn’t simply drive itself. Instead, it operates as an autonomous economic agent.

The concept works like this:

• A robotaxi is deployed onto the network.
• It picks up passengers and collects fares automatically.
• It pays for its own electricity, insurance, maintenance and cleaning.
• It saves a portion of every fare it earns.
• Once enough capital has accumulated, it orders another vehicle from the manufacturer.
• The new vehicle joins the fleet and begins earning fares itself.
• Over time, the fleet expands without a human owner continually injecting capital.

In other words the taxi, the vehicle, becomes a business and yet it is not a business owned by a company, but a business that effectively owns itself, continuously reinvesting its cash flow to reproduce. The vehicle is simultaneously the worker, the asset and, in a sense, the entrepreneur.

This sounds like science fiction, but it is really an extension of two existing trends.

Autonomous driving removes the human driver, whilst AI agents increasingly make commercial decisions on behalf of humans. Combine those with digital payments and smart contracts and you have the ingredients for what some describe as an autonomous economic agent. The software decides where to drive, accepts payment, purchases services, manages its own finances and reinvests its profits, and the implications go far beyond transport.

Imagine delivery drones that pay for their own charging and purchase replacements. Imagine AI data centres that buy electricity when prices are low and sell spare compute power when demand rises. Imagine autonomous factories ordering raw materials, hiring robotic logistics and financing expansion from retained earnings.

The fascinating question is no longer whether AI can do work. It is whether AI can own the means of producing that work.

Of course, today’s legal systems don’t recognise a car as an independent legal entity. A vehicle cannot own property, sign contracts or hold liability in its own name. Someone, whether an individual, a company or a trust, ultimately owns the robotaxi and remains legally responsible for it, but that may prove to be a temporary constraint rather than a permanent one. As AI agents become more capable, policymakers will eventually have to confront a new question: if software can earn money, spend money, pay tax and invest money, does it also need a new form of legal and economic identity?

That’s why this vision matters. It isn’t about taxis as an AI business; it’s about the emergence of capital that can manage itself.

For two centuries, the industrial economy has depended on the combination of human labour and financial capital. The next phase may see autonomous capital, assets that not only generate income but also decide how to deploy that income to create more assets.

If that happens, the robotaxi is not just the future of transport. It is the first glimpse of an economy in which machines don’t merely work for us. They participate in capitalism itself.

But then there is fraud. Fraud already moves at machine speed. What changes when fraudsters also have agents? Is the threat different, or just faster and at greater scale?

I think it’s fundamentally different.

People often say AI will make fraud faster, cheaper and more scalable. That’s true, but it misses the real shift. Fraud has already been automated for years. Botnets, credential stuffing, synthetic identities, phishing campaigns and account takeovers already operate at machine speed. Agentic AI changes something much more important. It gives fraud autonomy.

Today’s fraud tools are largely automated. Someone still has to design the campaign, choose the targets, monitor the results, adjust the attack and decide what to do next. An agent doesn’t wait for instructions. It just has an objective such as, “steal $10 million.”

Everything else becomes an optimisation problem.

Imagine an autonomous fraud agent that begins by scraping social media, LinkedIn and corporate filings to identify finance executives. It generates personalised emails, creates cloned voices from publicly available podcasts, books meetings in executives’ calendars, compromises suppliers, opens mule accounts, tests payment limits with tiny transactions, identifies the weakest bank controls, distributes funds across hundreds of jurisdictions and continuously changes tactics whenever a defence adapts.

No human is directing each step because the objective remains constant and, in order to achieve it, the strategy evolves. That is a completely different threat model.

It also changes the economics of fraud.

Today, organised crime has to allocate people to different jobs. One group writes malware and another launders money, whilst another creates fake identities and another recruits money mules. Agentic AI compresses all of that operational overhead into software that can coordinate these activities autonomously.

The cost of launching sophisticated attacks falls dramatically and the sophistication available to smaller criminal groups rises dramatically but, perhaps, the biggest change is that fraud becomes continuous.

Banks have traditionally viewed fraud as a series of events: a suspicious login, an unusual payment, a compromised account, whilst an agent doesn’t think in events. It thinks in campaigns.

It may spend weeks quietly learning how a company behaves before making a move. It may execute thousands of harmless interactions simply to establish trust. It may deliberately lose money on small transactions because doing so improves the probability of succeeding with a much larger attack later. It is patient because software doesn’t get bored.

That’s much closer to espionage than traditional financial crime.

Then there is the really uncomfortable possibility that fraud agents won’t operate alone but as swarms. One agent specialises in reconnaissance, another in social engineering, another in malware, another in payments, another in money laundering and another monitors law enforcement and media reports to determine when to change tactics.

These agents collaborate, divide labour and improve collectively.

This then looks less like organised crime and more like an autonomous criminal enterprise.

The obvious response is that banks will deploy defensive agents where every fraud agent will be met by dozens of defensive agents monitoring identity, behaviour, devices, payments, counterparties and network activity in real time. Defence will become autonomous too.

This creates a fascinating future where fraud won’t be humans attacking humans, but AI attacking AI. Bot-to-bot wars. Your payment agent requests a transfer. Your bank’s authentication agent validates your authority. A fraud agent attempts to impersonate you. A risk agent challenges the request. An identity agent checks delegated permissions. A compliance agent screens sanctions. A settlement agent decides whether to release funds.

During all of this, humans never see or engage in the conversation as it is all autonomous.

That is why I think the industry needs to stop talking about fraud detection and start talking about trust orchestration. In an agentic economy, every interaction becomes a negotiation between autonomous systems attempting to establish confidence before value moves. Identity, provenance, reputation, delegated authority, behavioural history and cryptographic proof become part of every transaction.

I guess the way to think of this is that the biggest fraud risk may not be that AI becomes better at impersonating people. It may be that AI becomes better at impersonating trusted AI.

If your personal financial agent receives an instruction from what appears to be your accountant’s AI, your bank’s AI or the tax authority’s AI, how does it know the other agent is genuine? How does it verify that the authority being delegated is legitimate? How does it know that one component in a chain of autonomous agents hasn’t been compromised?

That’s why I think the next decade won’t be defined by faster fraud. It will be defined by a battle over machine trust. The winners won’t be the institutions with the most sophisticated fraud engines. They’ll be the ones that can build an identity and trust infrastructure robust enough that autonomous agents can distinguish genuine intelligence from malicious intelligence without asking a human every time. That is a far bigger challenge than making today’s fraud systems run faster.

The UK Government’s National Payments Vision specifically calls for seamless A2A as a genuine alternative to cards. Does regulation feel like a tailwind for agentic commerce right now, or is it still playing catch-up?

I think we’re at one of those fascinating moments where regulation is trying to do two things at once.

On the one hand, governments understand that they cannot afford to slow innovation. On the other, they know that autonomous commerce could fundamentally change the financial system. The result is that regulation is no longer simply writing rules after innovation has happened. It is increasingly trying to co-design the future alongside industry.

The UK’s National Payments Vision is a good example. For years, account-to-account payments have been presented as the cheaper alternative to cards, but adoption has been constrained because consumers have valued convenience, protection and ubiquity over cost. Agentic commerce changes that equation. If an AI is making the payment decision rather than a human, it doesn’t care whether the underlying rail is a card or an account-to-account payment. It optimises for success rate, cost, speed, settlement certainty and the rules you’ve delegated to it.

That suddenly gives A2A a much stronger proposition than it had in a human-driven economy.

Equally, the FCA inviting fintech and financial firms into its AI Live Testing programme is a strong signal. The regulator isn’t saying, “Come back when you’ve finished.” It’s saying, “Let’s learn together.” That is a very different regulatory philosophy from the one that followed previous waves of fintech innovation, but I don’t think regulation is leading. I think it is learning.

That isn’t a criticism. It is probably the only sensible approach.

Nobody yet knows what the operating model of an agentic economy actually looks like. We don’t know how delegated authority should work when an AI commits you to a financial obligation. We don’t know where liability sits if one autonomous agent deceives another. We don’t know how anti-money laundering rules should apply when software creates thousands of temporary sub-agents. We don’t know how an AI should authenticate another AI, or what evidence is required before an autonomous system can trust another autonomous system.

These aren’t questions that can be answered by tweaking PSD2, open banking or existing payments legislation. They require entirely new legal concepts. That’s why I think the regulators are asking the right questions, but perhaps not the most critical question, as much of today’s discussion focuses on whether AI should be allowed to initiate payments, what controls should surround delegated authority and how existing payment rails should evolve. Those are all important but tThe bigger question is whether we are still regulating institutions when we should be regulating autonomous economic actors.

Historically, regulation has concentrated on banks because banks made decisions. In an agentic economy, increasingly the decision-maker is software acting on behalf of a bank, a merchant or a consumer. If the intelligence is making the commercial decision, perhaps the intelligence itself becomes part of the regulatory perimeter.

That sounds radical, but we’ve seen this pattern before.

When cars appeared, governments didn’t simply regulate roads. They introduced driving licences, vehicle registration, insurance, MOT tests and rules governing who could operate a vehicle. The technology forced a new regulatory framework because society needed confidence that autonomous movement could be trusted.

Agentic commerce may require the same evolution where regulators and not just regulating payment systems, banks and AI models, but regulating the right to act economically on behalf of someone else. That may involve licensing high-risk financial agents, requiring cryptographic identities, maintaining immutable audit trails, certifying decision models, mandating explainability for commercial decisions and defining clear liability when autonomous systems fail.

Viewed through that lens, the UK’s approach feels less like catch-up and more like preparation.

The FCA’s sandbox, AI Live Testing and the National Payments Vision suggest that regulators recognise the destination, even if nobody has drawn the final map. The risk isn’t that regulation moves too slowly. The risk is that we spend the next five years arguing about whether an AI should be allowed to press the “Pay Now” button, when the real transformation is that AI becomes the entity negotiating the contract, selecting the payment rail, managing liquidity, optimising tax, arranging finance and executing settlement without human intervention.

At that point, payments are no longer the innovation and have evolved to be just one capability inside a much larger autonomous economic system.

That’s why I see regulation today as a tailwind, because it has recognised something many haven’t: agentic commerce isn’t just another payments innovation. It’s the emergence of a new participant in the economy. The sooner regulation shifts from governing transactions to governing autonomous economic behaviour, the more likely it is that innovation and trust can grow together rather than in conflict.

In three years, what percentage of your own routine payments do you think will be initiated by an AI agent on your behalf? Give me a number.

By 2030, around 35% of routine consumer payments will be initiated by an AI agent, rather than directly by a human. Just to be clear, that is initiated, not authorised by an AI. Initiated by one.

Also, just to add context, direct debits, standing orders, subscription payments and more are already managing routine payments with no human involved, and this is just an evolution of those old services into agentic ones. I mean  think about what counts as a routine payment: renewing subscriptions; paying utility bills; buying groceries; booking train tickets; ordering household essentials; refuelling an EV; paying road tolls and parking; rebalancing investments; paying invoices; renewing insurance. None of these purchases require inspiration. They require optimisation. That’s exactly what AI is good at.

What I don’t expect by 2030 is AI buying your engagement ring, choosing your holiday destination or deciding which house to buy. High-value, emotional and life-changing purchases will still involve humans but the mundane? We’ll happily delegate it.

Think about how quickly we’ve already delegated things we once insisted on doing ourselves. We no longer remember phone numbers because smartphones do it. We rarely navigate using paper maps because GPS does it. Many people don’t even choose films anymore because Netflix recommends them.

Payments are simply the next layer of delegation.

The key thing here is that technology is not the limiting factor. Human confidence, regulation, liability and business adoption is. Consumers need to believe their AI won’t overspend. Banks need confidence that delegated authority is genuine. Merchants need commercial models that work. Regulators need clarity over liability when an autonomous agent makes a mistake.

Those things take time.

My longer-term prediction is much more aggressive, however. By 2030, 35% of routine payments will be initiated by AI agents rising to around 65% by 2035 and 85% by 2040. By then, the phrase “making a payment” may sound as old-fashioned as “dialling a telephone”, and the more provocative prediction, however, isn’t about the percentage. It’s that, by 2030, most people won’t even notice they’ve started using agentic commerce.

It won’t arrive with a dramatic launch. There won’t be a day when everyone switches to AI shopping. It will creep in one task at a time. First your calendar books the train. Then your car pays for charging. Then your home orders washing powder. Then your financial assistant moves money into savings because it knows your spending pattern.

One day you’ll realise you haven’t actually bought toothpaste, booked a taxi or paid an electricity bill yourself for months … and that’s when we’ll understand the real transformation. The future of payments isn’t people pressing “Pay Now” … it’s people deciding who gets permission to press it on their behalf.

Do you foresee a time when agentic commerce anticipates all our shopping needs before we’re even aware we need something? And is that the point? If so, how many years in the future is this?

That is the destination and I would go further.

The end game of agentic commerce isn’t frictionless shopping, but the disappearance of shopping altogether.

For most of human history, commerce has been reactive. We realise we need something, we search for it, we compare alternatives, we buy it and then we forget about it until next time. e-commerce shortened the process and then Amazon reduced it to one click. Agentic commerce removes the click entirely. The AI doesn’t wait for your instruction. It just predicts your intent.

Your refrigerator knows you’re running out of milk. Your car knows the tyres will need replacing. Your home’s energy system buys electricity before prices rise tomorrow. Your wearable health notices subtle changes in your sleep and orders vitamins after consulting your healthcare AI. Your financial agent shifts your mortgage because interest rates have moved, before you read the headlines.

You don’t wake up and think, “I need to buy something.” You wake up and discover it has already been taken care of. That’s not shopping. That’s orchestration.

People sometimes worry that this sounds like surrendering control. I don’t think that’s the right way to think about it.

When you use a thermostat, you aren’t giving up control of your heating. You’re delegating the routine decisions whilst retaining the ability to change the rules. Agentic commerce is likely to work the same way. You’ll decide your preferences, your ethical boundaries, your spending limits and your priorities. Your AI simply executes within those guardrails.

The really profound shift is that intent replaces transactions.

Instead of telling an AI to buy toothpaste, you tell it, “Never let me run out of toothpaste.” Instead of booking flights, you tell it, “Get me to New York for the conference, balancing cost, comfort and carbon footprint.” Instead of managing investments, you tell it, “Optimise my long-term financial security whilst keeping enough liquidity for family holidays.” The AI figures out everything else.

Notice what has disappeared. The transaction and the purchase have just become an implementation detail.

So, when does this happen?

I’d break it into three phases.

By 2030, perhaps 30-40% of routine purchases will be delegated. Consumables, subscriptions, travel bookings, bill payments and recurring household purchases increasingly become autonomous. Most people will still think they are shopping, even though AI is doing a growing proportion of the work.

By 2035, I think we’ll cross an important psychological threshold. More than half of routine commerce becomes predictive rather than reactive. AI won’t just execute instructions. It will make recommendations with increasing confidence and act automatically unless you object. People will begin to notice that they spend far less time making purchasing decisions.

By 2040, I suspect the majority of routine commerce has disappeared into the background. Shopping doesn’t disappear because people stop buying things. It disappears because buying ceases to be an activity. It becomes an automated process, like your email synchronising or your smartphone backing up your photos.

Then there is the part I find most interesting when the real competition won’t be over who has the smartest shopping agent, but over who gets to define your intent.

Your AI may know that you buy coffee every week. Does it optimise for the cheapest? The healthiest? The most sustainable? The local independent retailer? The company with the lowest carbon footprint? The one that supports fair trade? Or the one paying the platform the highest referral fee?

Those are not technical questions. They are philosophical ones. Your AI becomes an expression of your values. When that happens, it means that the most valuable companies in the next decade will not be the ones that sell products but the ones that understand your preferences so deeply that they can anticipate your needs before you’ve consciously recognised them yourself.

The future of commerce is not about purchasing. It’s about prediction.

The companies that win in the 2030s may not be those that persuade you to buy their products, but the ones whose AI understands you so well that your desires become predictable before they become conscious.

It’s almost like the vision of Minority Report except that the system knows what you want before you want it and not who you’ll kill before you kill them.

It is an extraordinary commercial opportunity, and an extraordinary concentration of power because, once an AI can reliably anticipate your needs, the line between predicting your choices and shaping your choices becomes very thin indeed.

If you had to bet on the one thing about agentic commerce that will surprise us most over the next five years, the development nobody is really talking about yet, what would it be?

Everyone is looking in the wrong direction.

The discussion today is about AI shopping assistants, payment credentials, digital wallets, fraud, identity and whether Visa, Mastercard or account-to-account payments will dominate. Those are all important questions, but they assume that the transaction remains the centre of commerce. It isn’t.

The biggest surprise over the next five years will be that the transaction becomes almost irrelevant.

For the last fifty years we’ve built an industry around helping people decide what to buy. Search engines, comparison sites, advertising, loyalty programmes, marketplaces, influencers, reviews and payment providers all exist because humans make purchasing decisions.

Agentic commerce changes the unit of value as the transaction is replaced by the objective. Instead of saying, “Buy me a flight,” you’ll say, “Get me to Singapore for my meeting.” Instead of saying, “Order groceries,” you’ll say, “Feed my family for £150 a week with healthy meals.” Instead of saying, “Find me insurance,” you’ll say, “Protect my home with the best balance of cover and price.” The AI decides everything else.

That sounds subtle, but it’s an enormous economic shift because companies no longer compete to win a sale. They compete to become part of an outcome.

Think about Google.

Google became one of the most valuable companies in history because people searched. But what happens when your AI doesn’t search? It already knows your calendar, your preferences, your budget, your dietary requirements, your travel patterns and your financial goals. Why would it present you with ten blue links when it can simply solve the problem? Search becomes fulfilment.

Now think about advertising.

Brands have spent a century trying to influence human emotions because emotion drives purchases. But if your AI is making routine decisions, emotional persuasion becomes far less important than structured data. Marketing starts to look more like publishing machine-readable trust, quality, sustainability and pricing information than producing beautiful television commercials.

Then think about payments.

Today we obsess over the payment because it’s the moment value changes hands. In an agentic world, payment becomes one API call inside a much larger workflow. Your AI negotiates, contracts, finances, pays, settles, records the receipt and updates your accounts automatically. The payment isn’t the event. It’s just another software function.

But the thing I think will surprise everyone most is something else entirely. I think we’ll discover that AI agents become consumers in their own right. Not legally or economically, but imagine your financial agent deciding it needs better fraud detection and purchasing access to another AI. Imagine your travel agent hiring a specialist visa-processing AI. Imagine your home energy AI buying weather forecasts from one service and grid optimisation from another. Agents won’t just buy things for humans, but they’ll buy services from other agents. That creates an economy that barely involves us.

Machine-to-machine commerce won’t be a niche. It may become larger than human commerce because software can transact millions of times a second over tiny amounts that would never justify human attention. If that happens, we’ve been asking the wrong question all along.

The question isn’t whether AI will change shopping. The question is whether humans remain participants in commerce.

That sounds absurd until you remember that most activity on today’s internet is already machine-to-machine. Servers communicate with servers. APIs call APIs. Algorithms trade with algorithms. Humans increasingly set objectives whilst software handles execution.

Commerce is likely to evolve the same way.

So, here’s my prediction. The biggest surprise of the next five years won’t be autonomous payments or AI shopping assistant or digital agent passports. It will be the realisation that the fastest-growing economy on Earth is one in which software is buying from software, negotiating with software and paying software, whilst humans simply define the goals.

When historians look back, they won’t say that AI changed commerce. They will say that commerce stopped being a human conversation and became a machine conversation that humans occasionally interrupted, and that’s the shift I think almost everyone is underestimating.

We still instinctively believe that markets exist for people. The next phase of the internet may create markets where people define the purpose, but autonomous agents conduct most of the activity. That is a much bigger transformation than replacing a checkout page with an AI assistant.