Argument architecture

How this editorial can be challenged

Core question

How should a bank tell the difference between a customer’s instruction and an agent’s plausible improvisation?

Proposed mechanism

An authenticated agent can execute a transaction efficiently while exceeding the customer’s intended mandate. Enforcing limits on payee, price, purpose and expiry before money moves makes the distinction testable.

Strongest counterargument

Requiring a fresh click for every ordinary variation destroys the convenience people delegate for, and can exclude customers who cannot respond immediately.

Our response

Offer a genuine choice of narrow, understandable tolerances in advance. A pre-agreed range is useful delegation; inventing that range after the cheaper option disappears is not.

Evidence limits

HSBC announced research plans and the HKMA an exploratory project. Neither record demonstrates an operational payment system, a settled liability rule or better customer outcomes.

What would change our mind

We would favor broader discretion if independent trials showed customers understood it beforehand, could revoke it promptly, and got faster resolutions without higher losses or false declines.

The fare changes at checkout

Imagine I tell an agent to buy the eight o’clock train ticket for $70. By checkout, the ticket is gone. There is one seat left at $90. The agent knows I want to get home, knows the journey matters, and has a valid way to pay. I want it to leave that seat unsold until it knows something else: did I give it the extra twenty dollars of discretion?

This is not a story about fraud in the usual sense. There need be no impersonator, stolen password or malicious instruction. A system can be helpful, authenticated and wrong about the most important fact in the transaction.

A name is not a mandate

Hong Kong’s Monetary Authority put a useful distinction into Project SAGE on November 2. The initiative will explore how to identify financial agents and verify their delegated authority. Its agenda also includes liability and consumer protection. These are questions for investigation, not rules already in force.

The distinction changes what a bank has to test. Proving that the software is mine does not prove that every choice it makes is mine. A bank that can produce a login record but cannot recover the instruction behind a disputed payment has answered a technical question and left its customer’s question unanswered.

Give the fraud desk the awkward cases

HSBC’s proposed institute with the Hong Kong University of Science and Technology gives researchers a place to work on safety and fraud. Before celebrating a detection rate, I would hand the team a small set of mundane surprises. The merchant changes, the fee rises, a single order becomes recurring, or the customer cancels while the agent is still comparing prices.

Keep the agent identity unchanged across those tests. The change in its permitted behavior should come from the instruction, not the confidence with which the software explains its decision. A refusal that arrives after the transfer is much less useful than one that arrives a moment before.

Don't solve this by making everyone click

The strongest objection is practical. If my agent asks again every time a fare moves by one dollar, I will either give up on it or approve every interruption without reading. And someone working a shift may not be able to answer in time. Friction has a cost, too.

Let me choose the range before I hand over the job: this route, today, up to $90, no subscription. Then let the agent do its work within that range. The important part is not the exact form of the control. It is whether a customer can describe what they agreed to without reading a transcript after the fact.

Who funds the right to object?

A transaction takes seconds. Reconstructing one can consume an afternoon. In the IANS and Artico survey, some organizations paid for AI security outside the security department. That survey does not tell us how HSBC funds its work. It does suggest a practical question for any bank: if a product team gets credit for completed payments, who has money and authority to resolve the ones customers dispute?

Keep a narrow record of the grant, the action and any change to permission. Preserve enough for a challenge, without hoarding every unrelated conversation a customer had with an assistant. Then give someone who can reverse or remedy a mistake access to that record.

What twenty dollars buys

There are days when I would gladly spend the extra twenty dollars to get home. There are others when it buys groceries. The purpose of a banking agent is not to guess which version of me deserves to win. It is to make the range I chose effective, including when that means missing a sale.

Give me speed inside that range and an honest interruption outside it. I would accept the occasional missed seat in exchange for knowing which decisions remain mine. Someone else might choose a wider range. A bank earns trust when both choices work, not when every transaction goes through.

Evidence behind the argument

Read the reporting

Opinion is ours. The factual record is linked below.

Hong Kong Monetary Authority — Project SAGE: agent identity, authority and consumer protection HSBC Hong Kong — planned AI and Emerging Technology Institute IANS / Artico Search — where AI security budgets sit