HSBC and Ant Digital Put AI Agents on Tokenized Deposits — But Who Judges Each Payment?

On October 9, 2026, banking giant HSBC and Ant Digital Technologies unveiled a test that allows AI agents to access digital services and perform micropayments using tokenized deposits, with real-time blockchain settlement. The three components involved are HSBC's Tokenised Deposit Service, Ant Digital's Anvita Flow network, and the Jovay Testnet, a layer-2 blockchain testing environment.

The announcement did not mention that risk checks, which determine if a payment appears suspicious, are solely the bank's responsibility and do not evaluate the wisdom of the instruction behind the payment. Anvita enables agents to independently discover, select, and pay for services, with no scoring of the agent's choice of service in the announced test.

Risk assessment is the domain of the bank, while judgment regarding the agent's actions remains unaccounted for. This marks the first instance where discovery and payment operations are intertwined in a bank-agent system, building upon a series of bank-led developments in 2026, including Santander's AI-agent-initiated Mastercard Agent Pay, Sygnum's blockchain mainnet transactions requiring customer approval, and CaixaBank's Visa Intelligent Commerce card transactions initiated by AI agents.

HSBC's partnership with Ant Digital today signifies the progression of decision-making power toward the agent. A recent report by Citrini Research, titled "Breaking The Wall," suggests that autonomous AI agents could generate demand for programmable financial infrastructure, as banks must adapt to accommodate these evolving systems.

The bank-agent sequence has been tested against an internal decision-gate endpoint, with two instructions scoring as auto-approve: a payment of $0.40 for a service discovered autonomously through a service-finding network, subject to real-time bank risk checks, and a payment of $0.02 for a listed catalog service within a budget and allowlist.

Despite lacking human review, both payments were auto-approved at a score of 0.80. This outcome demonstrates that a simple heuristic can process well-structured instructions that lack verification of the agent's service choice. The six banks jointly published a September 2026 paper titled "Building Trust in Agentic Commerce," stressing the need for auditable records of instruction, authority, intent, and outcome.

The gate's ledger (instruction, score, band, and outcome) serves as this crucial record. Builders are advised to incorporate the scored gate for every agent-initiated payment, with thresholds of 0.80 for auto-approval, 0.50 to 0.79 for confirmation, and below 0.50 for escalation. Implementing a service allowlist for discovery-driven flows and maintaining a log of scored payments with detailed information (instruction, authority, score, band, outcome) in an append-only manner is recommended.

Escalation should be based on patterns rather than transaction amounts alone. Detailed receipts, including curl-shaped data, and comprehensive FAQs are available at https://scriptmasterlabs.com/hsbc-ant-digital-ai-agent-payments.