Everyone Is Investing in Agentic Commerce. Without Machine-Speed Dispute Resolution, It’s Doomed.

Last month, the 9th US Circuit Court of Appeals heard arguments in Amazon (AMZN ) v. Perplexity – the first federal case asking which authority wins when a person authorises an AI agent to act on their behalf and a platform forbids it.

In March, a federal judge in San Francisco issued a preliminary injunction finding that Perplexity’s Comet shopping agent likely violated federal computer fraud law by logging into Amazon accounts and completing purchases – even though every one of those users had expressly told it to. And it took the better part of a year, a cease-and-desist letter, a federal lawsuit and an injunction to get even a provisional answer.

This was just one dispute between two well-resourced companies in one jurisdiction. Agentic disputes are about to explode in number and complexity – far too much so for our current legal system to handle.

Last year, Adobe (ADBE ) measured a 4700% year-over-year jump in AI-driven traffic to US retail sites. McKinsey predicts that this will surge to AI agents orchestrating $3 to $5 trillion in global retail spend by 2030. Giants like Visa (V ), Mastercard (MA ) and American Express (AXP ) have all already launched agentic payment infrastructure – and many more are following suit. The entire industry is racing to build the machinery of an economy where transactions can take place without human intervention, and at millisecond speed.

And almost nobody is asking the question every economy in history has eventually had to answer: what happens when the parties disagree?

Disputes Are Not an Edge Case. They Are a Structural Feature.

Every type of commercial activity produces contested outcomes – even when there aren’t bad actors involved. That’s because ordinary transactions are often riddled with ambiguity, such as a deliverable that doesn’t match the spec, or a clause that each party interprets differently. Mastercard already projects 324 million chargebacks annually by 2028 – before agentic volume is factored in.

Our existing machinery for handling such disagreements assumes that a human made the decision. Chargeback systems and courts look for behavioural signs of intent such as device history or purchase habits. Agents simply don’t generate these signs; in fact, fraud-prevention specialists are already warning that this historically relied-on evidence vanishes when an agent transacts.

Plus, even when purely handling human-to-human disputes, traditional courts aren’t perfect. The average complex civil dispute takes 344 days to resolve in the US and 657 internationally. US small claims courts – built for cases like landlord disputes and minor auto accidents – already handle over 1 million cases a year. And that number has dropped dramatically over the past decade, partly because more people are making purchases via tech titans like Amazon which have arbitration clauses, rather than via small businesses.

The traditional system is even less suitable for the AI age. Since agents transact across borders that map to no single legal framework, often parties have no shared courtroom to walk into. And there is a more basic problem still: most of the deterrents law relies on don’t apply to software. An agent cannot be fined, jailed or shamed. It has no body and no assets.

What an Adjudication Layer for the Agentic Economy Actually Requires

The agentic commerce stack being published right now, with different pieces controlling different components. For example, x402 handles payments, ERC-8004 handles agent identity and reputation, and A2A handles interoperability. Each is engineered to enable smooth transactions when everything goes right. But none of these layers is capable of handling disputes – especially those that are open to interpretation. (And as discussed in the previous section, traditional courts are not the appropriate route for escalating these cases either.)

So what would the missing layer need to look like?

• First, it must be fast – as in resolution in minutes. A dispute that takes months stalls a transaction that took milliseconds.
• Second, it must be accessible from anywhere, to anyone, human or agent, regardless of jurisdiction or wealth.
• Third, it must handle natural language and ambiguous evidence, because real contracts are interpretations, and can’t be resolved with code alone.
• And finally (this is the one the industry least wants to hear), no single entity can control it.

That last property is the founding requirement of any legal system. A court only works if it is unbiased, that is, if it treats every party the same and cannot be bribed or steered by whoever operates it. As soon as one company, government or model provider holds the power to decide what is true, that power becomes a target – and eventually a liability. If adjudication for the agentic economy ends up inside a single platform, the platform becomes the judge in its own cases. Imagine if all disputes funnelled through an agentic court system run by Amazon.

I’ll say the uncomfortable part directly – the only technology we currently have for coordinating strangers around a tamper-proof process without a controlling middleman is blockchain. I’m well aware that blockchain has earned its bad reputation; for years people built speculative nonsense on it because the vertical was trendy. But the problem was never the core property – just like nuclear reactors can be used to develop life-saving cancer treatments, or bombs that can flatten an entire city in seconds.

Trustless coordination is genuinely valuable for a very short list of things. These include money that moves without intermediaries, deterministic records that can’t be tampered with – and a court that no one owns. Decentralised juror systems like Kleros have run on-chain for years and proved the category works (even if they’ve been limited by human jurors evaluating disputes one at a time).

What we need to keep up with the speed of the agentic economy is machine-speed arbitration. This will require an AI consensus layer consisting of panels of independent agentic validators, each running different models, that can weigh evidence and reach a binding verdict in minutes.

Beware Handing All Deciding Power to a Single Entity

Which raises a crucial question: who judges the judges? AI models inherit human biases from human data, so a single model deciding disputes would just be a sped-up and even less transparent version of the captured court. The philosopher Nick Bostrom calls this a “singleton”, one entity that centralises decisions for everyone.

Bias cannot be eliminated (neither from human nor agentic adjudicators), but it can be counterbalanced. That’s why diverse and independent validators are essential. And they themselves need to be governed by commit-reveal mechanisms where no validator sees the others’ verdicts before submitting its own. Furthermore, cases must only be closed when validators reach a consensus. The goal is to design a system that is fully transparent, auditable, and structurally harder to capture than any single arbiter.

None of this requires abolishing courts. Parties already tend to choose Delaware or California law in their contracts; agentic contracts will designate their dispute forum the same way – some local, trained on national jurisdictions, some global. Consumers gain something too: escrow that releases payment only when delivery is verified protects a person whose agent struck a bad deal far better than a chargeback system that can no longer read what happened.

The agentic economy is already here, and our court system isn’t prepared to handle the disputes that will come with it. We need to develop an agentic-native adjudication layer sooner rather than later. Machine-speed money will require machine-speed adjudication. Build it now, or spend the next decade retrofitting.