A useful AI agent might need a market-data feed for fifteen seconds, a document-conversion API for one job and extra computing capacity for the next. The purchases could be small, frequent and difficult to predict in advance. Requiring a person to approve every transaction defeats much of the point of automation. Allowing the agent to spend freely creates a different problem.
That tension is the interesting part of an October 9 demonstration by HSBC and Ant Digital Technologies. In a joint announcement distributed by Ant Digital, the companies described a technical test in which an AI agent discovered a digital service, used it and completed a micropayment with real-time settlement through tokenized bank-deposit infrastructure.
It was a test, not a product launch. There were no published customer-adoption figures, production transaction volumes or independently established cost savings. But its architecture suggests where banks intend to compete as software agents begin making financial decisions.
The payment was one step in a longer workflow
Ant Digital's Anvita Flow network handled the service-discovery and invocation process. Jovay Testnet provided the blockchain testing environment. HSBC's Tokenised Deposit Service supplied settlement functionality and real-time risk checks through the bank's MCP interface.
Those functions solve different problems. A service-discovery network helps software find something worth buying. The payment infrastructure establishes whether the agent is permitted to pay for it, and how the transfer is recorded and settled. The merchant or service provider then needs a usable claim to the funds it receives.
The companies describe micropayments as amounts often below $2. That defines their intended use case; the announcement does not provide a complete distribution of test transaction values or a price comparison with conventional payment methods.
For production use, an organization would have to decide which agents can spend, on whose account, with which merchants, under what daily or per-transaction limits, and with what response when something goes wrong. Payment authorization, authentication and accounting are therefore part of the product—not optional administrative details around it.
Why a bank deposit matters in an AI payment story
Tokenized deposits are claims embedded in a banking relationship. Their transfer and redemption depend on the issuer bank and the relevant account framework. That differs from using an independently issued stablecoin, which involves a different issuer, reserve and redemption arrangement, and may circulate across a wider collection of networks.
A bank may have an advantage with enterprises that already need identity checks, treasury permissions and reconciled books. If an agent purchases dozens of tiny services on behalf of a company, the finance team eventually needs to know what was bought, why it was authorized and how those payments appear in the ledger. The ability to settle on a blockchain does not remove those requirements.
Stablecoins could compete effectively where open interoperability and broad merchant reach matter more. Tokenized deposits may suit customers that prefer to keep payment flows inside existing bank relationships. The most competitive architecture could also combine several payment instruments, depending on where the agent and its counterparties operate.
There is a question the demonstration leaves unanswered: why would a merchant or API provider prefer this arrangement to an existing subscription, card, prepaid balance or offchain account ledger？ A credible commercial case would need to compare total costs, failure handling, onboarding, settlement finality and the ease with which the recipient can use or redeem the money.
The pilot connects service discovery to bank-controlled settlement in a single transaction flow. That is useful progress. The next breakthrough will be less about an AI model's ability to press a virtual pay button and more about giving software limited, auditable financial authority across a network large enough to be useful.
