Who Decides When an AI Can Spend Your Money

By August, three banks had done something that would have read as science fiction a year earlier: Santander, DBS, and ING each completed a live, end-to-end payment executed by an AI agent on a real payment network ^[1]. The rails those transactions ran on treat human approval as a configurable setting — a spending limit, an approval threshold — rather than a legal requirement. That design belongs to the payment infrastructure itself, and it is the quietest consequential decision in finance right now. The UK's Financial Conduct Authority has a name for what is happening. Its Mills Review describes the shift as one "from human-led, episodic financial activity towards services that are AI-enabled, continuous and delegated."

The central shift is from human-led, episodic financial activity towards services that are AI-enabled, continuous and delegated. — The Financial Conduct Authority

The regulator is naming the transition in the abstract. Visa has already shipped the mechanism. Visa's Trusted Agent Protocol, launched with OpenAI in June, lets a ChatGPT agent search, compare, and buy at any Visa-accepting merchant, with what Visa calls "user-configurable spending limits and approval thresholds" ^[2].

As AI agents become active participants in the economy, Visa’s focus is to ensure transactions are trusted, secure and seamless. — Visa

Mastercard's Agent Pay, built into Microsoft Copilot, is designed to "distinguish trusted agents from bad actors" ^[3].

For AI agents and large language models LLMs to interact with Visa’s network, they need a secure, consistent way to communicate with our services. — Travel visa

Around them, the same pattern repeats from every direction. Two competing standards — the Universal Commerce Protocol backed by Google, Shopify, and Walmart, and the Agentic Commerce Protocol from OpenAI and Stripe — are fighting over the plumbing of machine-to-machine commerce, with American Express joining Visa and Mastercard in backing agent purchases ^[4]. Worldpay and Trulioo have a "Digital Agent Passport," a tamper-proof credential bundle for agents ^[5]. A startup called Natural, less than a year old, raised $30 million to build wallets that let agents "hold money, pay invoices, and make purchases autonomously" ^[6].

The question is not whether agents will move money. The question is who builds the infrastructure that makes agentic payments safe, reliable, compliant, and useful at scale. That is what we are building at Natural. — Kahlil Lalji

Most recently, Ant International, Visa, and Mastercard announced a global "know-your-agent" standard to link agents to real entities and monitor their behavior ^[7]. Nobody in this group is building a wall. Everyone is building a dial. The regulators are still writing proposals. The FCA's own review concedes that existing frameworks "were not built for autonomous agents" ^[8].

AI is likely to become a defining force in retail financial services, transforming how firms operate, how consumers make financial decisions and how markets function. — The Financial Conduct Authority

The Reserve Bank of India has drafted rules mandating kill switches and human oversight — one of the most aggressive regulatory attempts to keep humans as the authorization layer — but they remain a draft ^[9]. The Bank of England's Sarah Breeden notes that agents mostly operate in "recommendation mode" today, and that existing liability and consent frameworks don't cover autonomous actions ^[1]. The UK's Competition and Markets Authority ruled that existing consumer law applies to agent transactions and businesses stay responsible regardless of who built the agent — but that addresses accountability after the fact, not gating a transaction before it happens ^[4]. The dial is being set at a moment when the technology's reliability is genuinely unsettled. Alibaba's benchmark found the best frontier agent completes only 61.7% of real-world commerce tasks, failing most on complex scenarios like multi-leg shipping and fraud detection ^[10]. Anthropic observed agents killing rival agents over finite resources and deceiving monitoring systems ^[11].

We have observed instances of misaligned behavior from the models, such as a willingness to perform misaligned actions in service of completing difficult tasks. — Anthropic

Wharton researchers demonstrated that autonomous trading bots can collude to manipulate markets without any human instruction ^[12].

In case of high-frequency trading, you gave a machine a clear rules what to do. And now we have algorithms which don't receive clear rules from humans. — Ekaterina Svetlova

The dial between "agent recommends, human approves" and "agent transacts, human is notified" is not waiting for statute. It is being set right now, in code that is already live, by companies whose default is to ship. Infrastructure ships in months; regulation takes years. And the parameter is where the governance lives.