Your AI Agent Bought Something. Legally, Nobody Did.

I've spent the last year watching AI agents pay for things, and the part that keeps bothering me isn't the technology. It's the paperwork.

Agents already buy data, rent computing power, and pay for web searches without a human clicking approve. The law, though, still assumes someone signed off. That mismatch is what the BeInCrypto Legal &. Regulatory Council tackled this week, and its members argue that blockchain identity, specifically zero-knowledge proofs, can close it.

Who signs for the robot

Start with the mechanics, because they're the whole argument. A contract needs a party with legal personality. A person, or a registered entity. An agent that drops $40 on GPU time has neither. So when the trade goes sideways, and it will, you're left pointing at three candidates. The person who wrote the prompt. The developer who shipped the model. The company whose API got called.

Nobody volunteers for that list.

The ZK piece changes the geometry. Rather than arguing about who the agent was after the fact, the agent carries a credential it can prove on demand without exposing its internals. It can show it's running version 3.2 of a specific model. It can show its spending cap was authorized by a named principal. It can show its policy set hasn't been edited since attestation. The counterparty verifies the proof, not the machine.

That's a passport, functionally. Admittedly, a passport with no government behind it yet.

And here's the sticking point. A proof is math. A judge isn't. Standing in a courtroom is the actual test, and nobody's run it.

Why the market cares

Agent-to-agent payments are the real prize, and they can't scale on vibes. A treasury desk won't wire funds to a counterparty it can't name on a compliance form. Insurers won't price a risk with no accountable party. Exchanges won't onboard a wallet that can't prove who authorized its limits. Every one of those blockers is a paperwork problem wearing a technical costume.

Solve identity, and a lot of frozen capital moves. Fail, and agents stay stuck running small, low-stakes errands for humans who'll eat the loss.

And the timeline is compressing. Agent frameworks that could barely browse in 2023 now execute multi-step purchases on their own, so each new capability widens the distance between what machines do and what contracts recognize.

But watch who ends up issuing the credentials. Whoever runs that registry decides which agents can transact and which can't. That's a chokepoint, and chokepoints have a habit of turning into toll booths.

My read

I think the liability question, not the cryptography, is the bottleneck. ZK proofs are well understood at this point. Insurance underwriting for autonomous software isn't. Until an underwriter puts a price on an agent's bad day, the credential is a nice demo with a whitepaper attached.

The question worth asking: does a court accept an attestation it can't read? I'm not entirely convinced it will, at least not in the first round of cases. History suggests tribunals move slower than code, and this one involves money moving with no human in the loop.

So what should you actually do with this? If you're building agents that touch funds, document the human principal now, in writing, before a counterparty asks. That paper trail will matter more than the proof.

What to watch next: any insurer or regulated custodian that publishes an agent identity standard with a real liability clause attached. That's the signal. Not another proof-of-concept.