ReadSigil

The Signal · Issue 02 · Cover story · AI in Finance

Agents got bank accounts this week. Every one came with a leash

Stripe, Danske Bank, Visa, Mastercard and OnePay all opened money to AI agents within three days. The news is not that agents can pay. It is how differently each one chose to limit them.

By Shashwat Kumar · · 4 min read · Analysis

Abstract generative illustration: small agents tethered inside five bounded rings
Original generative illustration. No company imagery.

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

Between 19 and 21 October five payment and banking firms announced ways for AI agents to touch real money. Read side by side, the five announcements read less like product launches than like five drafts of the same mandate.

On 19 October OnePay announced Agentic Banking. A connected agent first gets read-only access ("It can't take any action"). If the customer allows purchases, OnePay opens a separate sub-account the agent spends from, with per-purchase limits and monthly caps the customer sets. Each purchase uses a single-use credential that expires after the transaction, and a toggle in the app revokes an agent's spending immediately.

On 20 October, at Cloudflare Connect, Mastercard showed Agent Pay for Machines, built on the open x402 protocol. A business gives an agent a job and a spending limit. When a service answers "402 Payment Required", the agent signs a payment voucher within that limit. In the demo the agent made nine payments to two merchants against one approved budget. Mastercard opened early access after pilots with Cloudflare, Ant International, AIsa, Base, BVNK, Coinbase and Hype. Its own summary: "Payments aren't the hard part. Authority is."

On 21 October three more arrived. Stripe Treasury for agents lets agents check balances, pay invoices, move funds between accounts and convert currencies. Businesses choose which accounts the agent can see, whether it can read or act, and when money movement needs human approval. Spending limits are promised "in the coming weeks". Visa, BILL and Billtrust launched a live pilot in which trusted agents complete approved invoice payments through supplier portals. Visa's Trusted Agent Protocol establishes the agent's identity and authority, and tokenised credentials replace shared portal logins. Visa says the pilot is processing end-to-end transactions. Danske Bank opened a Model Context Protocol (MCP) server so business customers can connect tools such as Claude or ChatGPT to its Premium APIs, through the same authenticated infrastructure and onboarding as any other integration.

The same day Anthropic opened a research preview of Claude Duet, a proactive assistant for professionals and small businesses. It waits for approval before it makes a payment, and Anthropic says some things are off-limits even with approval, "like trading securities or moving money between accounts".

Why it matters

Each firm picked a different leash. OnePay ring-fences a pot of money. Mastercard signs a budget. Stripe, for now, relies on a human approving money movement. Visa verifies who the agent is and what it is allowed to do. Danske keeps the bank's existing API perimeter and treats the agent as one more integration. And one assistant draws a hard line that no approval can cross.

Banks already know this problem. We call it a mandate, and we have been writing them for a century.

Signing limits, dual control, standing instructions, powers of attorney: banking has always let someone act for someone else within written bounds. What changed this week is that the "someone" can now be software, and that the first controls are being designed by payment networks and fintechs, not by banks.

For a bank, the question is no longer whether to let agents in. Customers' agents are arriving through MCP servers, card networks and fintech accounts either way. The question is which of these controls the bank will own, and which it will inherit from someone else's product decision.

Two gaps stand out. None of the five announcements says how disputes work when an agent pays within its limit but not as the customer meant. A monthly cap stops overspending, not a wrong purchase. And the controls differ in where they sit. A limit enforced by the provider holds whatever the agent says. A rule written into the agent's instructions does not. I built AgentGuard around that difference, and I argued in Proof of intent that the customer's instruction should be captured before the payment.

What I'd watch

  • Dispute rules. Who bears the loss on an agent purchase that was within its limit but wrong? The first network to publish clear rules sets the standard.
  • Whether limits become a published standard. Five products now mean five mandate formats. Banks will want one they can audit.
  • Where the line sits for money movement between accounts. OnePay allows it within a pocket and Stripe with approval, while Claude Duet rules it out. That disagreement will not last.

Sources

  1. Announcing Agentic Banking on OnePay OnePay, 19 Oct 2026.
  2. What happens when AI agents start paying each other? Watch Mastercard Agent Pay for Machines in action Mastercard, 20 Oct 2026.
  3. Introducing Stripe Treasury for agents Stripe, 21 Oct 2026.
  4. Visa announces enhanced capabilities that can bring trust and control to the future of business payments Visa, 21 Oct 2026.
  5. Danske Bank becomes the first bank in the Nordic countries to enable business customers to bank through AI agents Danske Bank, 21 Oct 2026.
  6. Introducing Claude Duet Anthropic (claude.com), 21 Oct 2026.