Policy

AI Agents Need Liability Rules Before They Can Pay Alone

OpenAI's ChatGPT Wallet requires human approval for every purchase, but the real challenge emerges when agents make authorized payments that don't match user intent.

Omega Editorial · November 21, 2026 · 4 min read

Key takeaways

  • ChatGPT Wallet currently requires human approval for every purchase, but this safeguard will disappear as agents mature—shifting authorization from transactions to instructions.
  • Agents create a new failure mode: purchases that are properly authorized and delivered but don't match user intent, falling outside existing dispute mechanisms.
  • Scaling autonomous payments requires three missing components: mandates that capture intent beyond spending limits, clear liability frameworks, and enforceable agent loyalty standards.
  • Private industry standards excel at interoperability but structurally avoid liability allocation—the question most critical for consumer trust and regulatory stability.
  • Companies that build liability rules and loyalty obligations into agents now will earn trust and avoid regulations written in response to future crises.

The approval bottleneck

OpenAI recently launched ChatGPT Wallet, enabling AI agents to execute payments on behalf of users. Every transaction currently requires human approval before completion. This safeguard addresses immediate concerns about unauthorized spending, but it also reveals a fundamental limitation: an agent requiring approval for each purchase functions more as an accelerated checkout tool than as true autonomous software.

The constraint will not persist indefinitely. As agent capabilities mature, approval will shift from individual transactions to broader instructions—from "approve this $47 hotel booking" to "plan my vacation within budget." That transition breaks a core assumption underlying payment infrastructure: that a human evaluates and authorizes each specific purchase at the moment funds move.

When authorized payments go wrong

Consider an agent instructed to arrange a family vacation for under $5,000. It books non-refundable flights for the wrong dates, reserves a hotel far from the intended destination, and accepts terms from an airline's agent that eliminate checked baggage allowances. Every payment was authorized within spending limits. Every merchant delivered exactly what was ordered. The outcome still fails to match what the user requested.

OpenAI's current terms preserve existing card dispute protections and leave delivery issues with merchants. Yet agents introduce a distinct failure mode: purchases that were properly authorized and fulfilled but did not reflect user intent. Existing dispute mechanisms may provide some recourse, but they do not clearly establish liability when the breakdown occurs in the agent's interpretation or judgment rather than in fraud or non-delivery.

Why it matters

The infrastructure gap around agent liability will determine which companies can scale autonomous payments and which face endless disputes. Organizations that establish clear liability frameworks, verifiable mandates, and loyalty standards now will earn consumer trust and avoid regulatory intervention written in response to high-profile failures. The card payment system became global infrastructure because users believed someone would make things right when problems occurred—agents need the same credible promise.

Three missing layers

Google's Agent Payments Protocol and similar initiatives use cryptographic signatures to record user authorizations. This verifies what was approved but cannot confirm whether the agent correctly interpreted preferences, acted in the user's interest, or operated free from conflicting incentives. Closing this gap requires three components.

First, mandates must capture intent beyond spending limits. Meaningful permissions encode user priorities, acceptable trade-offs, and escalation conditions such as "nothing non-refundable without confirmation." These constraints need verification across the transaction by merchants, payment providers, and dispute arbiters without unnecessarily exposing private instructions.

Second, a liability framework must determine who absorbs losses when agents act within spending authority but outside their mandate. Potential responsible parties include users who gave ambiguous instructions, providers whose models misinterpreted them, merchants whose agents extracted unfavorable terms, platforms that ranked options, or payment providers that executed transactions. No clear rule currently exists. The card industry's EMV liability shift offers a model: assigning fraud losses to whichever party had not adopted chip technology created powerful upgrade incentives. Clear allocation also enables pricing and insurance markets for agent-related risk.

Third, agents need enforceable loyalty standards. OpenAI is not currently taking transaction fees from Wallet payments, but business models evolve. When platforms earn revenue from merchants through fees, placement, or advertising, agents face conflicts of interest users cannot observe. Financial regulation addressed similar issues for brokers through disclosure requirements and best-interest obligations. Consumer agents require comparable standards defining whose interests they serve and how mandate compliance can be independently verified.

Who decides

AI labs, payment companies, and commerce platforms are currently writing agentic protocols. Private standards work well for identity, credentials, and payment flows where all participants benefit from interoperability. They perform poorly on liability allocation because every party prefers losses to land elsewhere. Systems that settle liability questions in advance scale successfully. Those that leave the question open either stall or face regulatory answers imposed after visible failures.

The analysis was first reported by Fortune in a commentary piece examining the structural challenges facing autonomous agent payments.

  • #ai agents
  • #autonomous payments
  • #chatgpt wallet
  • #payment liability
  • #agent mandates
  • #openai

This is an original analysis by the Omega editorial team. Source reporting: AI Watch.

Want systems like this working for your business?

Book a Call

More in Policy