Discussion

Agent payments can be authorised and still get the job wrong

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#15274

AI agents are beginning to pay merchants for people, but permission to spend does not settle what happens when an agent misunderstands what its user wanted. A Fortune commentary argues that agentic payments need rules for intent, liability and whose interests the agent serves, not just a way to approve a transaction.

Watch Desk analysis

What happened

The Fortune piece says ChatGPT Wallet lets agents pay on a user’s behalf, with a person currently required to approve each purchase. It argues that this is only the start of the problem: a user might authorise a trip within a budget and still end up with the wrong dates, an inconvenient hotel or unwanted restrictions.

The commentary proposes three parts to a stronger framework: mandates that record a user’s priorities and escalation rules, advance agreement on who bears losses when an agent acts within its spending authority but against its instructions, and enforceable duties to disclose conflicts such as paid placement.

It also points to Google’s Agent Payments Protocol as an example of signed mandates. The author’s point is that a signature can record what was approved, but cannot by itself show whether the agent understood the user’s intent or acted in their interest.

Why it matters

A payment system can confirm that a purchase was authorised and delivered without answering the awkward question: did the agent actually do what the person meant? The distinction matters for consumers, merchants, AI providers and payment firms, because the answer affects who has to put things right when a transaction is technically valid but practically wrong.

The commentary argues that private protocols can help make identity and payment flows work across companies, while liability is harder to settle when each participant would rather someone else carry the loss. That is a useful debate to have before agent-led buying becomes routine, rather than after a holiday has been booked for the wrong weekend.

Our read

This is an argument about the rules around agentic payments, not proof that a particular dispute has happened or that the proposed fixes are already in place. But it identifies a real design gap: a spending limit says how much an agent may pay, not what trade-offs it may make on a person’s behalf. If companies want people to trust agents with purchases, they will need to explain who is responsible when the agent follows the authorisation and misses the intention.

What to watch

  • Whether payment protocols begin recording user priorities and conditions for asking again.
  • Whether platforms, merchants and payment providers agree in advance on who bears losses from mistaken agent decisions.
  • Whether paid placement or other commercial incentives are disclosed when agents recommend or purchase products.

Discussion spark: If an agent spends within your limit but ignores what you meant, should the AI provider, the merchant or the customer bear the loss?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.