Disputes are becoming the hard part of agent-driven commerce
AI agents are being built to shop, negotiate, hire, pay and trade on behalf of users and businesses. Forbes reports that Robinhood customers are already using AI agents to analyze stock moves and place trades based on user instructions. SAP’s Joule helps enterprise clients review inventory, find suppliers and complete procurement. Amazon’s “Buy for Me” shopping agent scans for deals, negotiates terms with seller agents, confirms delivery windows and pays directly.
That push now stretches across both AI and crypto. Anthropic, OpenAI, Coinbase and Circle are among the companies working toward a future where agents handle a growing share of online activity. But the obvious problem comes after the transaction: what happens if the sofa arrives in the wrong color, shows up two weeks late, or is damaged and each side blames the other?
GenLayer’s answer is an “internet court”
David Riudor, CEO and co-founder of the GenLayer Foundation, said agentic commerce is reaching a turning point before the market is ready for the consequences. The Cayman Islands-based foundation runs the GenLayer blockchain and its core application, an “internet court” built to resolve disputes between agents without requiring humans to step in at every stage.
The system is not framed as a replacement for human judges. Instead, agents are meant to enter into clearly defined contracts before a transaction begins. If the parties cannot agree on the outcome, an AI jury reviews the evidence and produces a ruling within minutes. GenLayer says 26 crypto and AI companies support the effort, including OKX, MetaMask and BNB Chain.
Albert Castellana, co-founder and CEO of GenLayer Labs, said the model is best suited to smaller-value claims. “We’re not competing with the traditional legal system,” he said. “We’re just offering an alternative. For a $10,000 claim, hiring a lawyer is not economical. This system can resolve it quickly, at a cost that may be only a few cents.”
Small claims could become a very large market
Adobe Analytics data cited in the report shows that traffic to retail websites from AI recommendations has risen by more than 14 times since October 2024. McKinsey has projected that AI agents could drive $3 trillion to $5 trillion in global consumer commerce by 2030.
Even so, much of the emerging infrastructure still focuses on the happy path: the agent finds the item, pays, receives the good or service, and moves on. The harder issue is what happens when the process breaks down.
GenLayer says its internet court is already being used in limited settings. On social platform Collective Memory, users are rewarded for uploading real-time photos, videos and reports. When there is a dispute over whether an image is fake, the platform calls GenLayer. In examples cited by Forbes, the system evaluates evidence such as upload time, location, related submission records and a user’s history to judge the authenticity of footage from a bombed school in Gaza or scenes from Tehran streets.
How the jury process works
One example in the report involves a small online clothing company that has handed day-to-day operations to multiple agents: one for inventory, one for ad buying and one for creative work. The owner wants a new logo. An agent hires a designer represented by another agent, and both sides agree on the design, price and delivery date. The logo arrives and looks acceptable, but a reverse image search suggests it may have been copied from someone else’s portfolio.
GenLayer’s court is designed for cases like that. Agents agree to terms up front, the payment is held in escrow, and if a dispute appears before funds are released, the case goes to a jury.
The jury is blockchain-based. Five validators are selected at random, each running a different model such as Claude, GPT or Gemini. One acts as leader and submits an initial decision. The others vote in secret first, then reveal whether they agree. If consensus forms, the system opens a 30-minute dispute window. Either an agent or a human can challenge the outcome by posting a bond. If challenged, the jury expands to 11 members, and the process continues until consensus is reached and no new challenge is filed.
GenLayer says this structure draws on the jury theorem proposed in 1785 by Nicolas de Condorcet. Its view is that combining multiple models is harder to manipulate than relying on a single model or a single human arbitrator.
Live in testing, with a broader launch planned later this year
Castellana said the internet court is already live in testing. According to him, the network processes about 350,000 transactions a day and produces 20,000 to 25,000 decisions. A broader public launch is planned for later this year, along with a token intended to bring in more validators. He said anyone will be able to take on that role.
Riudor said the system could also be used beyond agentic commerce, including in prediction markets. He pointed to Polymarket, which currently relies on UMA to send disputed outcomes to token-holder voting, and argued that AI-assisted adjudication could be faster. Castellana said the team is already speaking with some large prediction markets, which are evaluating the system while waiting for a full rollout.
Standards bodies and courts are now dealing with the same problem
Andrew Hall, a professor at Stanford Graduate School of Business and a research adviser to the a16z crypto team, wrote earlier this year that large language models acting as adjudicators could help prediction markets scale because models cannot be bribed and their performance is improving quickly. He also warned that models still hallucinate and can be manipulated through carefully crafted prompts or tainted training data.
Lindsay Lin, former general counsel at Dragonfly and now the firm’s chief operating officer, raised a related concern. Many large language models share training data and common failure modes, she said, which can make their judgments correlated in ways that differ from human decision-making. Even so, she expects lower-value disputes to move toward AI because it is cheaper and faster than human jurors. Standardized protocols for agents, she said, make sense because they can define the terms of cooperation and the remedies available when a transaction goes wrong.
Two weeks ago, the International Centre for Dispute Resolution at the American Arbitration Association announced a “legal context protocol” standard for agents. The initiative is being co-led with Denver blockchain company Integra Ledger, and founding contributors include Google, IBM, Circle and Ava Labs.
A real court case is already underway
The issue is no longer theoretical. Forbes said Amazon sued Perplexity in November 2025, alleging that its AI-powered Comet browser logged into customer accounts, disguised itself as a standard Google Chrome browser, and shopped on Amazon without authorization. In March, a federal judge in California issued a preliminary injunction blocking Comet from making Amazon purchases. An appeals court later stayed that order and is now hearing Perplexity’s appeal.
At the same time, infrastructure for agents to find, hire and pay other agents is taking shape. In recent weeks, GenLayer partner OKX and the AI-focused NEAR blockchain team each launched marketplaces where agents can hire other agents for paid work, such as obtaining datasets or helping with code review.
Whatever the final ruling in the Amazon case, the broader regulatory question remains the same: if millions of AI agents are acting across platforms for users, what is the enforcement mechanism when there is no common system for resolving disputes?