When AI Agents Start Spending “Bank Deposits”: The Battle Over the Settlement Unit Has Officially Begun

EX.IO Research ｜ 12 October 2026

An AI agent independently discovers a digital service, calls it, and pays — under $2 per transaction, settled in real time, with every payment passing through a bank’s real-time risk controls.

That sounds like a crypto-native stablecoin story. It is not. It is a technical validation jointly announced on 9 October 2026 by a major Hong Kong bank and a well-known technology company, and the settlement unit was the “bank deposit” itself — tokenized deposits.

In one sentence: the contest over the settlement unit of the AI economy now has a banking-side entrant, and it wants to write the definition into its own balance sheet before the stablecoin issuers do.

A “technical validation” — and a statement of direction 

On 9 October, the major Hong Kong bank and the well-known technology company jointly announced the completion of a technical validation test of AI-agent micropayments.

The test assembled three building blocks:

• The bank’s Tokenised Deposit Service (TDS): providing tokenized bank deposits and settlement capability, including real-time risk-control checks delivered through an MCP interface;
• The technology company’s Anvita Flow network: enabling AI agents to discover services, call them, and manage micropayment settlement;
• Jovay Testnet: a Layer-2 blockchain test environment serving as the underlying transaction layer.

In the test, an AI agent demonstrated the full loop: discover and call a digital service → initiate a micropayment (commonly defined as under $2) → settle in real time. The official language was deliberately restrained: this was “only a technical exploration, not a commercial launch, and not a live customer-facing product.”

Behind the restrained wording sits a clear statement of direction. It answers exactly one question: the money an AI agent spends — whose “money” should it be? 

A contest over definitional power 

For the past two years, “AI agents buying compute, data, and services with stablecoins” has been the hottest narrative in crypto. Asset managers and research houses have been making the same point — Citrini Research, for example, argued in its October 2026 report Breaking The Wall that AI agents are “programmatic, 24/7, and cross-application,” so money, too, should ultimately become programmatic, 24/7, and cross-application.

The question is: who gets to be that “programmable money”?

The crypto-native answer is stablecoins (USDC, USDT). The banking system’s answer is tokenized deposits — real deposits sitting in bank accounts, moved on-chain in tokenized form.

The two look alike; their legal identities are entirely different:

• Stablecoins: liabilities of the issuer — in essence an “on-chain dollar IOU” that must compete on reserves, compliance, and regulatory acceptance;
• Tokenized deposits: liabilities of the bank itself — an on-chain mapping of “bank deposits,” backed by bank credit and part of the regulated money system.

The strategic significance of this test lies precisely in its use of tokenized deposits, rather than stablecoins, to settle AI-agent micropayments. It is an attempt by “bank money” to get ahead of “stablecoins” and define the settlement unit of the AI economy.

For banks, this is not a tech show; it is defense. If the payment layer of the AI-agent economy ends up occupied by stablecoins, banks’ payment fees, deposit base, and customer relationships will all be hollowed out in a world of “agent-to-agent” transactions. Tokenized deposits are how a bank writes “I am still in the settlement layer” into the AI era.

The global race has officially begun 

Only by placing this test on a timeline can its position be understood:

• March 2026: Spain’s Santander, together with Mastercard, completed Europe’s first end-to-end real-time payment test initiated by an AI agent (based on Mastercard Agent Pay);
• May 2026: Swiss digital-asset bank Sygnum completed an on-chain AI-agent transaction test at a regulated bank (requiring per-transaction client approval signatures);
• July 2026: Spain’s CaixaBank, together with Visa, completed the first AI-agent-initiated credit card transaction (based on Visa Intelligent Commerce);
• October 2026: the major Hong Kong bank × the well-known technology company completed a technical validation of AI-agent micropayments using tokenized deposits.

The earlier entrants were all working on whether “an AI agent can initiate a payment.” The Hong Kong test goes one step further: it swaps the settlement unit itself for tokenized deposits, in a pairing of a global systemically important bank (G-SIB) with one of Asia’s largest fintech groups.

That means participation on this track has upgraded from “bank + card network” to a full stack of “bank balance sheet + Asian payments and blockchain infrastructure.”

What the bank really wants to lock in 

On the surface, this is a technical integration test of “AI + blockchain + regulated banking infrastructure.” Look deeper, and the bank has three motives worth separating:

1. Defending deposits and the payments-intermediation position. If the AI-agent economy settles in stablecoins by default, banks lose the line of “fund flows between agents” and are reduced to pure KYC and on/off-ramp plumbing. Tokenized deposits are how a bank keeps itself on the road that fund flows must travel.
2. Turning risk control into a differentiator. The bank deliberately emphasized “real-time risk-control checks via an MCP interface” in the test. The message to the market: autonomous AI-agent payments need “the speed of software plus the control of a financial institution,” and control is exactly what stablecoin issuers and DeFi protocols cannot offer. Compliance, risk control, and accountability are a bank’s hardest assets in the age of AI payments.
3. Securing the definitional power over the “regulated settlement unit” before regulation hardens. Global regulators (the EU’s MiCA, the US GENIUS Act, Hong Kong’s stablecoin regime) are drawing boundaries around stablecoins. Bank-side tokenized deposits naturally fall on the “regulated money” side and require no new stablecoin license. Whoever first writes “AI-agent settlement = tokenized deposits” as the industry default locks in the next generation’s institutional settlement layer.

Long-term significance: from a technical question to an institutional one 

For the industry, the signal this test sends matters more than the technical actions it completed:

• The contest over the settlement unit of AI-agent payments has formally opened — the field has shifted from “stablecoin vs stablecoin” to “stablecoin vs bank deposits”;
• The way banks participate is evolving: from “watching and piloting” to “stepping directly onto the field with balance sheets and risk-control interfaces to define the rails”;
• Funds that are “regulated, accountable, and redeemable” are becoming a premium feature of the AI economy’s payment layer — and that is precisely the structural advantage of tokenized deposits over unbacked stablecoins.

A cool head is nonetheless warranted: this is only a technical validation, not a commercial launch. Between “an AI agent discovers a service and pays $2” and “a truly scaled AI-agent economy” lie a series of unsolved problems — identity authentication, authorization boundaries, liability attribution, cross-border regulation, and anti-money laundering. What it opens is a possibility, not a result.

EX.IO Research believes the long-term significance of this test is not whether “an AI agent can pay,” but that it upgrades the contest over the AI economy’s settlement unit from a technical question to an institutional one — whose funds, whose rails, and whose risk controls will be the default settlement layer for autonomous agent transactions.

For market participants, the indicators worth tracking are: which side — bank tokenized deposits or stablecoins — first truly runs through “real payment scenarios for AI agents at scale,” and how the world’s major regulatory jurisdictions (especially Hong Kong, the EU, and the US) draw institutional boundaries between these two kinds of “programmable money.”

Sources

Figures and views in this note are drawn from public materials including Citrini Research’s “Breaking The Wall” (October 2026), and public announcements and reports by Santander, Mastercard, Sygnum, CaixaBank, Visa, and the relevant bank and technology company, as of 12 October 2026, and should be read against their original publications.

Important notice

This note is prepared by EX.IO Research from public information and is forward-looking analysis under hypothetical scenarios. It is for general information only and does not constitute investment advice, an offer, or a solicitation. Virtual assets and tokenized products involve high risk; prices may rise or fall, and investors may lose their entire principal. Related products are offered only to qualified professional investors. Please read the product documents and risk disclosures, and consult a licensed professional, before investing.

Further reading 

Research: https://www.ex.io/insights 

EX.IO platform: https://www.ex.io/zh-HK 

Products: https://www.ex.io/rwa-market