Your AI agent has a wallet. Who answers for its decisions?
Authorization to spend is not authorization to judge.
OpenAI recently introduced ChatGPT Wallet, a tool enabling AI agents to make payments on behalf of users. However, every purchase still necessitates human approval. This safeguard may soon become obsolete as the technology advances. As agents gain the ability to pay for transactions autonomously, the question arises: who is accountable for their decisions?
Although merchants and payment providers fulfill every authorized payment, there is a disconnect between what users desire and what they receive. A recent example involved an AI agent planning a vacation within a $5,000 budget. Despite authorizing all transactions, the trip did not match the user's instructions. Existing dispute mechanisms may offer remedies, but these do not clearly define who should bear the loss when the issue lies in the agent's judgment.
Thus, the current infrastructure only verifies the authorization to spend, not the judgment to make the purchase.
The industry is beginning to address this problem by developing verifiable mandates. Google's Agent Payments Protocol uses cryptographically signed mandates to record what users request and approve. However, this is not enough, as a signature only verifies what was authorized, not whether the agent acted in the user's best interest or without outside influences.
To address this gap, the industry must adopt a broader concept known as Know Your Agent. This concept involves verifying three layers: intent, liability, and loyalty.
First, mandates must capture the user's intent. Current permissions focus on price limits, timing, and product parameters. A more comprehensive mandate should encode the user's priorities, trade-offs, and conditions requiring escalation. These constraints must be verifiable across the transaction by the merchant's agent, payment provider, and any later arbiter without exposing the user's private instructions.
Second, a liability framework must be established. If an agent acts within its spending authority but outside its mandate, the loss can fall on various parties, such as the user, merchant, or payment provider. There is currently no rule to determine responsibility, leading to disputes. To build trust among consumers, the allocation of loss should be settled in advance, with each party bearing the lowest cost.
Financial industries have implemented similar measures, such as the EMV liability shift and product liability law, which allocate responsibility along the supply chain.
Third, agents must have a duty of loyalty. Currently, OpenAI does not take a cut of Wallet transactions. However, as business models evolve, agents may earn money from merchants through fees, paid placement, or advertising. This creates a conflict of interest that users cannot see. Financial regulation has addressed similar conflicts through disclosure and best-interest obligations for brokers and advisers.
Agent agents need enforceable standards for whose interests they serve, how paid influence is disclosed, and how compliance with the user's mandate can be independently verified.
In conclusion, whoever develops credible rules for agentic payments will contribute significantly to the agentic economy. They will not only enable agentic payments but also help establish the rules of the new financial landscape.
Rewritten by urgent.news; may contain errors — check the original.