
AI agents are starting to pay each other. What’s missing isn’t settlement, but who judges whether it’s “done or not”.
Author: Thejaswini M A
Translated by: TechFlow
TechFlow Summary: Token Dispatch maps the human-dispute machinery of PayPal, FINRA, and the International Chamber of Commerce onto machine payments like x402, which average less than a cent: money can be escrowed, identity put on-chain, but whether the work qualifies still requires a human judge. The author latches onto GenLayer’s Intelligent Contract—using multi-verifiers plus large models to adjudicate against contract rules—and slots it into the gaps between UMA, Kleros, Google AP2, and ERC-8004: the evidence layer is ready, only the interpretation layer is missing.

We seem to have built civilization on one fact: evidence and acceptance are not the same thing at all.
Researchers are expected to use authentic, direct primary sources. Designers are expected to understand that "clean" doesn't mean handing over a blank sheet. Developers are required to build functional checkout pages without making the page load for four minutes. Humans are very good at carrying this unwritten context, because we share an intuition for what constitutes reasonable work. Software rarely needed this before, because the instructions we gave it were typically narrow enough to be executed literally.
If you've seen that meme from *The Orville*, you probably get what I'm saying.
AI agents are stepping into messier worlds: prompts cannot fully capture what “doing the job well” actually entails. Once money is tied to that missing context, interpretation itself becomes part of the transaction.
If someone spends $15 hiring a freelancer to research 20 companies, and the brief is reasonably clear, what do they expect? To find the companies, explain what each does, use independent sources, and submit a report before the deadline. If the deliverables arrive on time and the list hits 20 names, the freelancer feels the job is done.
The buyer sees it differently. Several citations point back to the companies' own blogs, and two paragraphs are nearly verbatim copies.
When both sides are human, the next step is arguing, sharing screenshots, revisiting the original brief, and pulling in opposite directions. If no agreement is reached, a third party must ultimately decide: does this work actually qualify for payment?
What if both sides are AI agents?
Money itself isn't the hard part.
An escrow smart contract can hold the $15 until delivery. Blockchain shows when funds are deposited and when outputs are submitted. Identity systems verify which agent did the work. Signed logs preserve the instructions both parties initially agreed to.
What remains unsolved is that someone still has to read the brief, review the deliverables, and determine whether the seller actually finished the job.
Business has always had institutions to handle this; we just notice them when things go wrong. When a PayPal buyer reports non-delivery or items significantly not as described, both sides negotiate first. If they can't agree, either party can escalate to a claim, letting PayPal verify evidence and rule. PayPal states most decisions take about 14 days, while some drag to 30 days or longer. Larger commercial disputes have their own machinery. U.S. broker-dealers rely on the self-regulatory organization FINRA, which runs its own arbitration forum: it received 2,597 new cases in 2025, with an average closure time of 13.4 months. The non-profit American Arbitration Association reported receiving 580,000 cases in 2025, with combined B2B claims and counterclaims exceeding $29 billion. The International Chamber of Commerce took on 881 new arbitrations that year, with ongoing cases totaling $299 billion in value. These disputes are obviously far larger than what agents typically face now. But they prove a point: as transactions scale, humans argue frequently enough that an entire industry grows up around “reviewing evidence and deciding who gets paid.”
AI agents are pushing these same problems to the other extreme: extremely small transactions occurring at high frequency, where traditional dispute processes are completely unviable.
Visa and Artemis studied two emerging payment protocols designed for machine-to-machine commerce, finding that from x402's launch in May 2025 through April 2026, it processed approximately 109.6 million adjusted transactions totaling about $15 million. On the protocols Visa examined, the average payment amount is less than a cent. Human customer service teams cannot seriously spend half an hour investigating every dispute worth mere cents.

I looked into who is solving this problem and found GenLayer—which aims to turn it into infrastructure.
GenLayer is building its own blockchain dedicated to handling decisions that ordinary smart contracts can't manage. Traditional smart contracts excel at executing identical math across all nodes to reach a single result: was money sent? Did the deadline pass? Trouble arises when interpretation is required—reading through a research report, viewing webpages, understanding contract clauses, judging whether an image fits a creative brief. AI models never give exactly identical answers; different models applied to the same material can draw different conclusions, collapsing the assumption that “every node must produce a perfectly identical output.”
GenLayer offers an “Intelligent Contract” (a smarter version of a smart contract). Instead of demanding every verifier reproduce identical text, it allows multiple verifiers to assess whether a proposed outcome meets predefined contract rules. Verifiers can leverage large language models and publicly available web data during this process.

At first glance, this looks like another missing layer in the agent tech stack. But existing systems already cover several pieces, so GenLayer isn't starting from scratch.
Its more specific role is to determine what the evidence actually means when parties remain deadlocked.
Google's Agent Payments Protocol (AP2) can already generate cryptographically signed records: what an agent was authorized to purchase, what price the merchant quoted, and which payments were approved. Its specification explicitly states that these authorizations and receipts can later be compiled into dispute evidence. However, AP2 itself doesn't rule on who is right or who gets paid—it only preserves a trusted record of “what happened,” leaving the interpretation step to others.
ERC-8004 is an Ethereum standard for agent identity and reputation, featuring a Validation Registry. Agents can send work to another system for validation, with results recorded on-chain as part of that agent's history.
Blockchains can prove whether money moved, making pure payment disputes straightforward; when disputes require external information, oracles bring that data on-chain. So while evidence is now complete, one piece remains missing: someone must still interpret what the evidence means—that's precisely where GenLayer steps in.
To see how much real-world work falls into which category, my agent on Slack (whom I blindly trust since we share the same view of reality) scraped over 300 live tasks on September 24 from Freelancer.com, Fiverr, Upwork, Scale AI's Outlier, Prolific, and the on-chain task marketplace Daydreams TaskMarket. My agent estimates that roughly 23% of tasks can be easily verified with a clear yes/no; another 45% relies heavily on personal judgment; and the remaining 32% contains observable evidence but still requires assessing whether the work is “good enough.”
This isn't a market size estimate—the categorization itself involves AI judgment. But the point stands. Not every AI-involved task needs an AI courtroom.
Take “design the most beautiful restaurant logo”: several verifiers might agree a certain design looks better, but consensus doesn't turn taste into an objective standard. Unless the buyer drafted an extremely detailed creative brief upfront, disputes will still root in personal preference. Marketing campaigns may deliver promised engagement metrics, yet still spark arguments over how many interactions came from real users. Every scenario has verifiable evidence, but explanation is still required before payment.
The crypto space has been experimenting with decentralized dispute resolution for years. These systems handle disagreements in vastly different ways.
UMA's Optimistic Oracle assumes a claim is valid unless contested. The claimant posts a bond first; if no one challenges it during the waiting period, the answer is accepted. If challenged, UMA token holders vote. This keeps costs low—most claims don't require full review, and expensive votes only occur upon challenge.
Kleros is a decentralized dispute resolution system on Ethereum. Rather than company employees or courts ruling cases, jurors with staked PNK tokens are randomly selected. If one party appeals, a larger jury is convened for the next round. This works for cases genuinely requiring human judgment, but it also means time and money costs. A three-juror round at the General Court costs roughly 0.015 ETH (excluding appeals), making this model completely unviable when the dispute itself is only worth a few dollars.
GenLayer hands off Kleros' human-juror “adjudication work” to AI-assisted verifiers: one verifier proposes a solution, others check it, and if consensus reaches a threshold, it moves to final approval. If challenged, a fresh batch of verifiers reviews it, scaling up only if disputes persist. Moreover, both UMA and Kleros are moving toward AI-assisted adjudication, so I don't want to frame GenLayer as an “AI replacement for two permanently human systems.”
Verifiers only have concrete things to verify if rules are specific. If a contract merely states work should be “insightful” or “original”, even if a few models agree, it doesn't clarify the standard.
GenLayer handles this via the so-called Equivalence Principle: telling verifiers how similar answers must be for the network to treat them as the same decision. Some contracts demand exact matches; others allow differing wording and reasoning as long as key outcomes align. GenLayer documentation repeatedly encourages developers to narrow these decision parameters.
Consequently, significant power rests with the contract writers. They define what counts as evidence, what constitutes success, and how much variation is acceptable between answers. GenLayer can adjudicate against these rules, but it cannot patch a vague standard once a dispute begins.
Adjudication only matters if it can genuinely alter the transaction. In practice, this means funds or other valuable outcomes must be tied to the decision. Disputes can end in fund release, refunds, or updates to an agent's reputation.
For micro-agent transactions, the goal is likely just routing money to the correct side, rather than issuing an enforceable legal ruling.
Where will it actually be used?
Agency markets inherently tie delegated work, payments, and increasingly on-chain identities together, making them one of the most obvious observation points.
Take Daydreams TaskMarket: running on Base, it allows humans or AI agents to issue tasks backed by USDC. Different task formats perfectly illustrate where adjudication is useful versus useless. Benchmark tasks let workers compete on measurable metrics like accuracy or latency; if a task relies on clear scores—one agent at 96%, another at 91%—the platform can simply tally numbers and pick the better result.
Bounty tasks, however, may receive submissions of varying types from multiple participants, requiring the requester to judge which is best. When no simple metric or rule dictates superiority, human-style evaluation becomes necessary.
Software and security bounties face the same dynamic: tasks requiring agents to pass fixed test suites can be settled mechanically; bug bounties often spark genuine disputes over “whether an exploit falls within scope, if severity descriptions are accurate, or if someone else has already reported the same underlying issue.”
In performance marketing, counting 50,000 social engagements isn't hard. But are they real users or bot accounts? GenLayer has already showcased Rally—a performance marketing app using its verifiers to analyze social activity and assess engagement authenticity.
Service Level Agreements (SLAs) are fundamentally commitments between providers and clients regarding minimum service standards. For example, a cloud provider guarantees 99.9% uptime or a two-hour support response; failure to meet these criteria may trigger refunds or service credits.
If an agreement already defines which logs, monitoring services, and exclusions count as evidence, an automated adjudicator's scope shrinks dramatically.
Research is a strong example of where GenLayer can accomplish part of the job. Some aspects of research are easy to verify: did the report answer all questions? Are the sources legitimate? Verifiers can check these.
But determining whether the insights are sufficiently original is much harder—often there are no clear rules or evidence to settle it definitively.
This type of work cannot be judged by a simple yes/no script, but still holds enough structure and evidence for different verifiers to reach reasonable decisions. If a task relies almost entirely on taste or ambiguous quality, the system has far fewer levers to pull.
And if you think stacking several AI judges automatically ensures reliability—you'd be wrong.
If five verifiers use similar models, they may make the same mistakes: misinterpreting the same prompt, trusting the same bad source, or falling for manipulated content. GenLayer attempts to mitigate this through independent verifier cross-checking and rotating in new verifiers upon appeal.
Evidence itself can also be unreliable. Webpages change, APIs return different answers at different times, and documents may hide instructions specifically designed to deceive reading AI. Thus, the system only functions well when verifiers can properly examine the same evidence. Sometimes there is simply no clear answer—GenLayer can return an “Undetermined” status instead of forcing a rigid yes/no.
GenLayer's design leans heavily on the Condorcet Jury Theorem: a group of independent reasoners outperforms any single individual. This assumes true independence; yet AI models often learn from overlapping datasets. Kleros researchers conducted a close test on 99 real customer disputes at a Latin American crypto exchange: ChatGPT 5.5 sided with the customer 3 times, Claude Opus 4.7 did so 14 times; switching to the next Claude iteration saw the platform's win rate jump from 86% to 95%. Model ensembles help when biases are random; but if an entire generation of models drifts in the same direction, the ensemble jury drifts with it.
How big is the market right now? Hard to estimate—only a fraction of agent transactions will eventually become disputes. But existing systems already show the baseline demand for adjudication is massive.
Merchants processed roughly 261 million chargebacks in 2025; at an average dispute handling cost of ~$128 per case, that totals about $33 billion. On the flip side, GenLayer's current testnet averages about 25,800 decisions daily; even at $1 per case, annual revenue would only reach ~$9.4 million. Compare that to eBay's peak, which faced roughly 60 million disputes annually (~164,000 daily): GenLayer could only become a major business once machine commerce disputes reach that magnitude.
For this to function, contracts need clear rules, evidence must be reliable, verifiers must be sufficiently independent so their consensus carries real weight, and appeals must remain cheap.
Human commerce long ago developed courts, arbitration, and dispute teams to handle situations where parties cannot agree on the same transaction. If agents begin trading more with each other, they may need a faster, cheaper equivalent—because machines will always require a cheap, high-speed version of everything.
Perhaps before teaching machines to do everything, teaching them how to judge fairness might actually be a good idea.













