Know the AI agent before it pays.

Businesses are handing their invoices to AI agents that read, decide and pay. When one of them pays, the bank cannot see which agent acted, on whose authority, or who answers for it. Agent Passport gives the bank those three answers before money moves, and a record afterwards that stands up on its own.

Customer dashboardBank dashboard

King's College Chapel and the River Cam, Cambridge
King's College Chapel across the River Cam, Cambridge

Hackathon partners

Organisers and regulators

C:\>DIR Cambridge Digital Innovation & Regulation InitiativeFinancial Innovation for ImpactNayaOnevouch.financeBIS Innovation HubGlobal Financial Innovation NetworkDigital Regulation Cooperation Forum and Regulator Knowledge ExchangeWomen Regulators NetworkNetworks for HumanityEthereum FoundationSwiftGoogleMoneyboxAnt InternationalEuroclearC:\>DIR Cambridge Digital Innovation & Regulation InitiativeFinancial Innovation for ImpactNayaOnevouch.financeBIS Innovation HubGlobal Financial Innovation NetworkDigital Regulation Cooperation Forum and Regulator Knowledge ExchangeWomen Regulators NetworkNetworks for HumanityEthereum FoundationSwiftGoogleMoneyboxAnt InternationalEuroclear

Ecosystem and academia

GFTNInnovate FinanceMENA Fintech AssociationIDBAfrica Fintech NetworkABFintechsAmerican Fintech CouncilTRMAlliance for Innovative RegulationFintech Alliance PHairdropdDFSAKWorld Alliance of International Financial CentersAFSIAnacofiCFTEFintech Association of Hong KongFintech Association of JapanFACEAlliance of Digital Finance and Fintech AssociationsGBBCInsurtech AustraliaFinTech WalesNCFApartnerCambridge Centre for Alternative FinanceE-Lab, King's College CambridgeEntrepreneurship Centre, Cambridge Judge Business SchoolOxford AI Governance Initiative and Oxford Martin SchoolUniversity of ManchesterUNRafSingapore Management UniversityLITE Lab, University of Hong KongCSOTechnical University of MunichNayaOneProductopediaAutraconGFTNInnovate FinanceMENA Fintech AssociationIDBAfrica Fintech NetworkABFintechsAmerican Fintech CouncilTRMAlliance for Innovative RegulationFintech Alliance PHairdropdDFSAKWorld Alliance of International Financial CentersAFSIAnacofiCFTEFintech Association of Hong KongFintech Association of JapanFACEAlliance of Digital Finance and Fintech AssociationsGBBCInsurtech AustraliaFinTech WalesNCFApartnerCambridge Centre for Alternative FinanceE-Lab, King's College CambridgeEntrepreneurship Centre, Cambridge Judge Business SchoolOxford AI Governance Initiative and Oxford Martin SchoolUniversity of ManchesterUNRafSingapore Management UniversityLITE Lab, University of Hong KongCSOTechnical University of MunichNayaOneProductopediaAutracon

The shift

AI agents are already buying. Last holiday season AI and agents stood behind a fifth of everything sold online, and the businesses behind those shoppers are handing over their invoices next. The question is no longer whether AI agents will pay, but whether anyone can see on whose authority.

Agentic commerce, global, $ billion: actual to 2025, forecast to 2030

2024 and 2025: online holiday sales influenced by AI and agents, Salesforce 2024, Salesforce 2025. 2030: agentic commerce spend, Juniper Research forecast; McKinsey puts the same year at $3 to 5 trillion.

The numbers

This is what happens when the person paying cannot see who is really behind a request. Of the £1.28 billion taken from UK customers last year, £576 million was authorised push payment fraud, up 19% in a year: payments the victims made themselves, in good faith, to someone they could not see. An AI agent sees even less than they did, and since 2024 the sending bank carries the loss.

Authorised push payment fraud, UK, £ million

Source: UK Finance, Annual Fraud Report 2026. Losses reported to UK Finance.

The problem

Today an AI agent's payment looks exactly like a person's. Nothing on it says which agent acted, what it was allowed to do, or who stands behind it. The bank sees an instruction and nothing else.

flowchart LR
  I["Invoice
from anyone"] --> A["AI agent
reads, decides, pays"] A --> B["Bank
sees only an instruction"] B --> M["Money moves
with no record of who acted"] B -.-> Q1["Which agent?
no signer"] B -.-> Q2["Was it allowed?
no mandate"] B -.-> Q3["Who answers?
no one can prove it"] class A agent class M out class Q1,Q2,Q3 q

Identity alone does not fix this. Knowing who an AI agent is tells you nothing about what it may do, and nothing about who answers when it goes wrong.

Our proposal

Rather than watching AI agents more closely, the future demands transparency by design. Every payment carries a provider's filing, a customer's signature and a bank's verdict, so authority is provable and revocable on every transaction.

  1. The bank registers the AI product. Once.
  2. The bank approves it for its list. Once.
  3. The customer signs the mandate. Once.
  4. The bank checks. Every payment.

The result

When the AI agent tries to pay the altered invoice, the bank holds it at the payee check for a person to review, and money never moves. What remains is a record the bank produced itself: who asked to pay whom, how much, which check stopped it and why, signed by the bank and chained to every earlier entry. The provider answers for its filing, the customer for its mandate, the bank for the check. A supervisor can read that record without operating any of it.

The customer's approval screen: a £12,000 payment to Ashby Ironmongery Ltd, requested by the AI agent, held for review because it is more than the £10,000 a payment the mandate allows, with Approve payment and Reject payment

Powered by

A passport is a promise. On vouch.finance the promise becomes money that can only be spent the way it was promised. When a customer signs its mandate, the bank mints an AI Voucher carrying the agent's limits, its payees and its passport. The bank runs the nine checks and the vouch policy engine authorises the settlement, so a refusal on one is a refusal on both. Suspend or revoke the passport and the voucher goes with it: the next instruction fails on every rail.

Powered byvouch.finance