Agentic payments are payments that an AI agent starts and completes on behalf of a person or a business, under a mandate that sets what it may buy, from whom and up to what amount. Moving the money is the part the industry has largely solved. The harder part comes afterwards: proving that a payment matched the mandate, and that the data the agent priced it from was the data it claimed to use.
That second question is now on the table at large banks. On 9 November 2026 Deutsche Bank and Ant Digital Technologies released a white paper, "Agentic Payments in Wholesale Finance", covering treasury, cross-border payments and trade finance. It argues for "bounded agentic execution" rather than open-ended autonomy, with limits that are explicit and can be revoked. Ant Digital Technologies president Zhuoqun Bian put the rule in one line: an agent "may propose but cannot extend its own authority". A bound only means something if someone can check, after the fact, that the agent stayed inside it. (For the wider picture of agents that move money, see our guide to
what agentic finance is.)
Agentic payments split into three layers
The International Monetary Fund (IMF) gives the clearest map. Its note "How Agentic AI Will Reshape Payments" (April 2026) separates agentic payments into three layers: intent, authorization and settlement. Agents reason, search and negotiate in the intent layer, where no money moves. Rule-based checks in the authorization layer decide whether a proposed action may go ahead: is this agent who it says it is, and is the payment inside its mandate? The settlement layer then moves the money, and once settled the payment is final in law and cannot be reversed.
The split is useful because each layer fails in its own way. In the intent layer an agent can misread a price, act on stale data or follow a hidden instruction planted in a web page. The authorization layer can approve a payment that looks valid against a mandate that was too loose, or check identity without checking what the agent based its choice on. Settlement does what it is told. The IMF points out that standing mandates replace payment-by-payment instructions, which makes it harder to trace a single payment back to what the user actually wanted.
Settlement shows that money moved; the proof of why it moved has to come from the layers above.
The rails agents pay on today
Agents already have two kinds of rails. Card networks run agent programs on top of existing card payments, such as Mastercard Agent Pay and Visa Intelligent Commerce. In September 2026 Ant International, Mastercard and Visa announced a joint effort on Know Your Agent (KYA) principles, so an agent checked on one network can be trusted on the next. It is an announced collaboration, not a published standard yet.
Stablecoin rails are the second kind. The best known is x402, an open standard that puts stablecoin payments inside ordinary web requests, so software can pay per request without an account or a checkout page. Mastercard's Agent Pay for Machines pilot, shown in October 2026, runs on x402 and settles on Base, Coinbase's Ethereum layer 2 network. Our post on
stablecoin stack layers shows where these rails sit.
The real volume is still small. TRM Labs measured x402 payments on Base, Solana and Polygon from May 2025 and published the results on 9 September 2026. The payments went through facilitators, the services that send x402 payments on-chain for the buyer. After TRM removed volume that did not look like real buying, such as addresses paying themselves, only between 0.6% and 7.5% of what was left could be traced to AI agents, depending on how strict the test was. Agent payments are early, which makes this the time to decide what their records should hold, before the volume arrives. TRM also names the gap: "The agent signs and a facilitator broadcasts, with no review in between."
What the settlement record must prove
When an agent payment is disputed or audited, four questions come up. Was the payment inside the mandate? What data did the agent price it from? Was the amount the one the agent agreed? Was the counterparty the one the agent checked? Rails answer some of these well and leave others to whatever the operator kept in its own systems.
| What has to be proven | Card rails with an agent program | Stablecoin settlement on-chain | Proof-backed record |
|---|
| The mandate | Where the program records it (for example Mastercard Verifiable Intent) | Only spend rules in the wallet or contract | The mandate as it stood when the agent acted, with a proof |
| The data the agent priced from | Not carried | Not carried | The query result the agent used, with a proof |
| The amount | In the card transaction | On-chain, visible to anyone | Linked to the settlement transaction |
| The counterparty | The merchant, through the card network | A wallet address, checked off-chain | The counterparty check the agent ran, with a proof |
Two points stand out. Card rails and stablecoin rails both show the payment well, and the card networks are building records of what the user authorized. Mastercard describes Verifiable Intent as "a tamper-resistant record of what a user authorised" that links identity, intent and action. But neither rail carries the data the agent read when it chose the price, the supplier or the moment to pay. That data sits in the operator's databases, and in a dispute the operator is the one who hands it over.
The second point is privacy. The data behind a payment decision is often the most sensitive part: a treasury position, a supplier's credit status, a customer's balance. Putting it on-chain to make it checkable is not an option for a bank or a payment service provider (PSP). The record has to let a counterparty or an auditor check it without seeing all of it.
Where Know Your Agent stops
KYA answers who the agent is and who stands behind it. Our post on
AI agent identity in finance covers that layer and why classic access control falls short for agents. BlackRock's paper "The Machine-Native Economy" (September 2026) describes the current pattern: know your customer (KYC), anti-money laundering (AML) and KYA checks generally happen off-chain, and only the verified result is passed on-chain to decide whether a transaction may go ahead.
That pattern settles eligibility. It does not settle what happened inside a single decision. An agent can be properly identified, hold a valid mandate and still pay the wrong amount because it read an old exchange rate. An industry white paper convened by the Monetary Authority of Singapore (MAS), Safeguards for Agentic Finance at Runtime (SAFR, October 2026), proposes a "tamper-evident, append-only record of every governance decision" and warns that what an agent reports about its own context should not be treated as self-verifying. Our post on the
AI agent audit trail looks at what such a record has to hold for an auditor to check it.
Where verifiable data fits in the settlement record
zkDatabase, the verifiable database built by Orochi Network, is designed to prove the data behind an agent payment: the mandate, the reference data and the query results the agent used, which neither card nor stablecoin rails carry. It generates a Zero-Knowledge Proof (ZKP) for every insert, update and query. A ZKP lets one party show that a statement about private data holds without revealing the data itself.
In an agent payment flow this works in four steps.
- Store the mandate and the reference data with a proof. The mandate, the spend limits, the approved counterparties and the price or rate data the agent will use are written to zkDatabase. Each change comes with a proof that it was stored and processed as claimed, and a short fingerprint of the whole data set is published on-chain, so any later change to the data shows up.
- Query before paying. When the agent prices a payment or checks a counterparty, zkDatabase returns the result with a proof that it matches the published state. The agent's policy (the rules code that decides whether the agent may act), or a smart contract, checks the proof before the payment goes to settlement.
- Link the proof to the payment. The record of the payment stores which proof and which data state the agent used, next to the settlement transaction on the card or stablecoin rail.
- Let the other side check. A counterparty, an auditor or a regulator checks the proof against the published fingerprint. They see that the mandate and the data the agent used match the published state, and which settlement transaction they are linked to, without seeing the treasury position or the full counterparty list.
The proof covers how the data was stored and read. It does not show that an exchange rate from an outside provider was correct, or that the agent's choice was a good one; those stay with the data provider, the model and the people who set the mandate. It is extra evidence the auditor can check, not a replacement for the audit.
Bring proof to your agent payment flow
Talk to the Orochi team about which part of your agent payment flow depends on private data, and how a proof for each query could sit next to settlement.