Legal Context Protocol Logs the Terms AI Agents Accept

When a consumer instructs an artificial intelligence (AI) agent to book a vacation for less than $3,000, they grant a budget, not permission for the agent to make any choices within that limit. The agent handles decisions like selecting the airline, accepting non-refundable fares, adding travel insurance, and dividing the trip across multiple charges without further input.

However, when issues arise, there is often no record of what the shopper permitted and what the agent decided. The American Arbitration Association and Integra Ledger unveiled the Legal Context Protocol in June to create such a record. Founding contributors include Google, IBM, Circle, Wayfair, and UiPath, along with numerous blockchain and identity firms.

The protocol functions as an open standard which permits a merchant to publish its terms on a static web address, allows an agent to confirm which version it viewed, and enables both parties to sign the agreement. Bridget McCormack, president and CEO of the American Arbitration Association, stated that the agentic economy requires "the same capacity delivered at machine speed."

The protocol does not necessitate blockchain or any intermediary; any organization with a web server can integrate it. The protocol divides a transaction into two components. Authorization systems, such as Google's Agent Payments Protocol, record the consumer's side, while this protocol captures the merchant's side: the offered terms, the obligations accepted, and the recourse available.

The protocol establishes four levels of proof. The simplest involves the agent uncovering the terms, which is considered consent. The most stringent requires a digital signature binding a specific party to a particular document, linking to arbitration, escrow, and compliance systems. A hash of the terms accompanies the payment record, allowing both sides to verify what the document contained at the moment of purchase.

The consumer's restrictions reside in what the protocol refers to as a buyer policy. This policy outlines spending limits, permissible regions, acceptable dispute bodies, and the threshold beyond which a person must review before the agent signs. Triggers for this review may include charges exceeding a certain amount, first-time merchants, or terms from unfamiliar jurisdictions.

When the policy escalates, the agent's signing process pauses until a human or supervising agent approves. This design aligns with consumers' expressed desires. Data from PYMNTS Intelligence and Visa Acceptance Solutions, surveying 5,241 consumers, 1,185 merchants, and 150 acquirers across the U.S., Brazil, and the UAE, showed that only 23% of U.S. consumers trust generative AI to manage payment transactions.

Retailers have begun incorporating their own conditions into their fine print. Target revised its terms of service in March, indicating that once a customer authorizes an agent, any choices made by the agent are treated as if they were made by the customer. An airline might hold the shopper responsible for the selected carrier or fare.

Bloomberg Intelligence anticipates that about 20% of online commerce, or $500 billion, will pass through agents acting on a single prompt by 2030, up from less than 5% this year, as reported by Bloomberg Law. An additional 6.2% will be fully autonomous, with agents choosing what to purchase and when.