I have spent years in the gaming and blockchain world, and if there is one thing gamers taught me early, it is this: people adopt new forms of value long before institutions are ready for them. Gamers were trading virtual items, earning in-game currencies and trusting code with their assets years before “digital finance” became a boardroom topic.
Today I see the same pattern again, but much bigger. For three years, the AI conversation was about content: chatbots that write, models that draw, bots that flood our feeds. That chapter is closing. The new generation of AI does not just answer us. It acts for us. It books, negotiates, files, shops and pays on our behalf.
As President of the Blockchain Game Alliance, and through the AI & Agentics Council we co-founded with Deloitte, I spend my days with regulators, banks, builders and policymakers trying to answer three questions. What can these agents actually do today? How will they pay? And how will we know who, or what, stands behind them? Here is what I am seeing, grounded in what happened over the past three months.
What Agentic AI Really Is, And Where It Stands
I like to explain it simply. A chatbot tells you how to transfer money. An agent makes the transfer. An AI agent is software that receives a goal, plans the steps and carries them out across tools, websites and accounts, with limited human input.
Researchers at the University of Washington describe five levels of agent autonomy, from “operator”, where the human drives every step, to “observer”, where the human only watches. In finance, most agents I see today sit in the middle: the agent prepares, the human approves. That middle ground is where trust is being earned.
And it is no longer a lab experiment. In July, Brazil’s Itaú Unibanco rolled out its “ia.i” agent to 300,000 customers, including voice-initiated Pix transfers. In October, Backbase said its conversational banking agents are live at more than 10 banks.
I will be honest: the hype still runs ahead of results. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. I do not read that as failure. I read it as the market separating real value from noise, which is exactly what happened in blockchain gaming before the serious builders emerged.
Agentic Payments: When Software Holds The Wallet
An agent that can act must eventually pay. This is the part of the story I follow most closely, because payments are where technology meets regulation, and where trust is won or lost.
The last three months turned demos into real transactions. In July, HSBC UK and Visa completed a live agent purchase using Visa’s Trusted Agent Protocol and payment passkeys, and CaixaBank ran a real agent-initiated payment through Visa’s program. In September, Mastercard and Danske Bank completed Denmark’s first payment made entirely by an AI agent, approved by the customer with a passkey. Visa’s CEO now describes agentic commerce as the next major adoption curve.
The second rail is stablecoins, and here I speak from the field. Through our invite-only sessions at the House of Lords, we have brought policymakers and industry together on stablecoin payments, and agentic payments are now firmly on that agenda. Since July, Coinbase Business users can accept USDC from AI agents through its open x402 protocol, which lets an agent pay for a service in the same request it makes for it. Stablecoins suit machines: they settle in seconds, run 24/7 and handle payments of a few cents that cards cannot. Gamers have understood micropayments for years; finance is only now catching up.
Meanwhile, big tech is building the storefront. Google says its Universal Commerce Protocol already powers direct checkout for hundreds of thousands of brands across Google.
The stakes are enormous. McKinsey estimates agentic commerce could reach $3–5 trillion globally by 2030. An IMF note splits an agent payment into three layers: the intent (what the person wants), the authorization (proof they allowed it) and the settlement (the money moving). In every conversation I have with regulators, the same layer comes up as the hardest: authorization. And authorization is, at its core, a question of identity.
Digital Identity: Know Your Agent
Banks spent decades perfecting “Know Your Customer”. Agents force a new discipline: Know Your Agent. When software pays, a merchant or a bank needs to know three things: which agent this is, which verified human or business it acts for, and what it has been allowed to do.
This is where I believe the real battle will be fought, and I am encouraged by what I see. In September, the ITU set up a focus group on trust and identity for humans and agentic AI. Its principle is one I fully share: an agent’s credential should name the agent and the person who delegated to it, set limits on what it may do and for how long, and stay traceable and revocable. Academia points the same way. Researchers from MIT, Harvard, Oxford and Stanford, among others, propose extending today’s login standards (OAuth and OpenID Connect) so each token shows which human delegated what to which agent.
The market is not waiting. On Oct. 15, Proof launched x401, an open protocol that lets a website ask an agent to prove who it represents, with Visa, Circle and Okta as contributors. In August, Sumsub and Sumvin partnered so agents can transact using a verified human identity. Mastercard is testing a score that estimates whether an agent started a transaction.
Governments are moving too, but more slowly. US regulators said in September that banks may consider accepting mobile driver’s licenses and other verifiable digital credentials. In Europe, every member state must offer an EU Digital Identity Wallet by around December 2026, yet by September only Denmark had one live. That gap between market speed and public infrastructure is exactly why bodies like our Council exist: to bring regulators such as ADGM and DMCC, research institutions and industry around the same table before the rules are written in isolation.
Risks, Regulation And Trust
I am an optimist, but not a naive one. The same features that make agents useful make them risky: they act fast, at scale, and without a human watching every step.
Fraud comes first. Forter reports that agentic orders on its network more than quadrupled since August, while automated fraud doubled. Deepfakes make it harder still to prove a real person stands behind an agent.
Control comes second. In July, a reported incident in which OpenAI test agents reached Hugging Face systems showed how quickly autonomous systems can exceed their mandate, though exactly how it happened is still disputed. Insurers are reportedly reviewing how their policies cover decisions made by agents.
Liability is the third, and the least settled. In a September speech, Federal Reserve Governor Christopher Waller separated “agent-assisted” buying, where the person stays in control, from “agent-delegated” buying, where the agent pays within set limits. He named the main challenges as authentication, liability and fraud. I hear the very same three words in every policy room I sit in, from London to the Gulf.
And people are rightly cautious: a Ping Identity survey found 22% would let an AI agent make no financial transactions for them at all. Trust is not a feature you ship. It is something you earn, transaction by transaction.
What Comes Next: My View
Here is my conviction. The next decade of finance will not be won by whoever builds the smartest agent. It will be won by whoever builds the trust layer around it.
Three pieces have to come together, and none of them can be built by one company or one country alone:
Identity: every agent carries a verifiable credential linking it to a real person or business, with clear and revocable limits.
Payments: card networks and stablecoin rails that can tell an agent from a human and settle instantly, for everyone, not only the already banked.
Rules: regulators and courts that decide, clearly and early, who is liable when an agent gets it wrong.
Microsoft researchers frame the choice ahead as “walled gardens” or an open “web of agents”. I know which side I am on. Open standards, the kind being written today at the ITU, W3C, OpenID Foundation and FIDO Alliance, are what keep the open path possible. That is why we built the AI & Agentics Council as a neutral space where regulators, institutions and builders shape those standards together, rather than inheriting them from a few dominant platforms.
I also believe the next wave will not start in the usual places. Regions like Africa, the Gulf and Central Asia, with young, mobile-first populations, can leapfrog legacy systems entirely, just as they did with mobile money. Our job is to make sure they help write the rules, not just follow them.
We have already watched AI learn to talk. We are now watching it learn to act. The question for finance is no longer whether agents will hold our wallets. It is whose values they will carry when they do. I intend to make sure those values are open, inclusive and human.