Trust Agent for Payments: Verifiable AI Actions Across Card and Real-Time Fraud Workflows

The unverifiable decision is the real fraud risk

In payments, the model that catches fraud is no longer the hard part. The hard part is proving, months later and to a regulator's satisfaction, exactly why an automated system blocked a card, held a real-time payment, or cleared a flagged transaction. Most AI fraud tooling produces a score and an outcome. It does not produce evidence. When a dispute, a chargeback, or a supervisory review lands, the firm is left reconstructing a decision the machine made and cannot account for.

That gap is where regulated payments businesses stall. A model can be accurate and still be undeployable, because accuracy without an evidentiary chain becomes a liability the second a customer complains or an auditor asks. Trust Agent closes that gap. It makes every automated action in a fraud workflow verifiable and signed, so the decision and its proof are the same artifact.

Trust Agent is one of the Studios that run on Mickai, the sovereign AI operating system: AI that regulated businesses own and run inside their own walls, on-prem and air-gapped, with every action written to a tamper-evident, post-quantum-signed audit record we call the OAR. Built and live.

What "verifiable" actually means in a fraud workflow

A fraud decision is not one event. It is a chain: a transaction arrives, features are computed, a model scores it, rules and thresholds apply, a human or an autonomous policy acts, and an outcome is written back to the card or payment rail. Each link is a place where a firm later needs to answer a question. What data did the system see. Which model version scored it. Which policy fired. Who or what authorised the action. When.

Trust Agent signs each of those links as it happens. The inputs, the model and version, the policy applied, the actor, and the timestamp are bound into a record that cannot be altered after the fact without detection. The signatures are post-quantum, so the evidentiary chain holds against the cryptographic threats that are coming, not only the ones that exist today. The result is that an autonomous fraud action is no longer a black box with a log bolted on. The proof is generated at the moment of decision and is inseparable from it.

This matters most on real-time rails, where there is no settlement window to pause and review. An instant payment is held or released in milliseconds. Trust Agent lets a firm run that decision autonomously and still walk into a supervisory meeting with a signed, ordered account of every action the system took, reconstructable transaction by transaction.

Card and real-time, the same evidentiary spine

Card fraud and real-time payment fraud have different tempos and different attack patterns. Card disputes play out over weeks through scheme rules and chargeback cycles. Real-time fraud is a one-shot decision against an irrevocable transfer, with authorised push payment scams as a dominant loss vector. Firms typically run separate stacks for each, with separate logging, and reconcile them by hand when a regulator wants a unified picture.

Trust Agent puts both on the same evidentiary spine. The same OAR record format covers a blocked card authorisation and a held faster payment. A fraud operation can show consistent, signed reasoning across the entire payments surface, and an auditor sees one chain of evidence rather than two incompatible ones. It also means the firm can deploy autonomous policies on real-time rails with the same confidence it has on card, because the proof obligation is satisfied identically in both.

This sits naturally alongside Nemesis, our fraud and AML Studio, and Plutus, our finance Studio, so the same signed record carries through from the fraud decision into reconciliation and reporting. The evidentiary chain does not break when the workflow moves between functions.

Why this is built sovereign, and why that is the wedge

Payments fraud data is among the most sensitive a firm holds. Card numbers, transaction graphs, victim details, and the model behaviour that protects them are exactly the assets that compliance regimes ring-fence. A large share of the market is legally constrained from running this work on public-cloud AI. Around 850,000 UK businesses, roughly 15 percent, and around 5 million across the EU sit under rules that bar it: PRA SS2/21, UK GDPR special-category data, the EU AI Act's high-risk classification, the NIS Regulations, and the cross-border exposure of the CLOUD Act.

That is the wedge. The sovereign AI market is moving from about USD 40 billion in 2025 toward an estimated USD 148 billion by 2032, and the regulated payments segment is squarely inside it. Mickai is built for the buyers who cannot send this data anywhere, and Trust Agent gives those buyers something they could not previously have: autonomy and auditability at once, inside their own walls.

The moat under it is an IP estate of 104 filed UK patent applications, roughly 2,340 claims, held by Mickai LTD. Filed, not granted, which establishes priority and a prior-art position across the verifiable-action and sovereign-audit architecture. As a third-party momentum signal, Micky Irons was ranked number 4 on Crunchbase as of June 2026, with the Mickai company profile in the top 1 to 2 percent globally.

The dual-buyer thesis

Trust Agent serves two buyers from one architecture. The first is the regulated payments firm that needs to deploy autonomous fraud workflows it can defend. The second is the platform that wants to offer verifiable, sovereign AI actions to its own regulated customers and would rather own the capability than build it. Mickai is positioned as an ally to that second buyer, not a competitor to the model labs. The category we lead, verifiable autonomous action under regulatory constraint, is one a hyperscaler would rather own than reproduce, because the IP estate and the standing audit architecture are the hard part, and they already exist.

A category being built to scale

Trust Agent is live today, running on the same sovereign substrate as the rest of the Mickai Studios, with the OAR as the connective tissue across every one of them. We are building to scale from a UK base, with Birmingham manufacturing secured for the hardware path. The economics follow the architecture: a Y5 revenue path to billions at high gross margin, underwritten by the patent estate and the dual-buyer thesis. This is a category a hyperscaler would want to own, and the substance is already in place.

If verifiable AI for payments fraud is a problem you own, I would like to talk.

Micky Irons, founder and CEO of Mickai. Contact: micky@mickai.co.uk

FAQ

What does Trust Agent actually do in a payments fraud workflow? It makes every automated decision verifiable and signed at the moment it happens. Inputs, model version, policy applied, actor, and timestamp are bound into a tamper-evident, post-quantum-signed audit record, so an autonomous block, hold, or clearance carries its own proof for later review.

Does it work for both card and real-time payments? Yes. Card fraud and real-time payment fraud run on the same evidentiary spine and the same OAR record format, so a firm can show consistent, signed reasoning across its entire payments surface and present one unified chain of evidence to an auditor.

Why does sovereignty matter for fraud AI specifically? Fraud data is among the most regulated a firm holds. A large share of UK and EU businesses are legally constrained from running this work on public cloud under regimes like PRA SS2/21, UK GDPR, the EU AI Act, and the CLOUD Act. Mickai runs on-prem and air-gapped, inside the firm's own walls, so autonomy and compliance no longer trade off.

Are the patents granted? They are filed, not granted: 104 UK patent applications with roughly 2,340 claims held by Mickai LTD. Filing establishes priority and a prior-art position across the verifiable-action and sovereign-audit architecture.

Frequently asked questions

What does Trust Agent actually do in a payments fraud workflow?

It makes every automated decision verifiable and signed at the moment it happens. Inputs, model version, policy applied, actor, and timestamp are bound into a tamper-evident, post-quantum-signed audit record, so an autonomous block, hold, or clearance carries its own proof for later review.

Does it work for both card and real-time payments?

Yes. Card fraud and real-time payment fraud run on the same evidentiary spine and the same OAR record format, so a firm can show consistent, signed reasoning across its entire payments surface and present one unified chain of evidence to an auditor.

Why does sovereignty matter for fraud AI specifically?

Fraud data is among the most regulated a firm holds. A large share of UK and EU businesses are legally constrained from running this work on public cloud under regimes like PRA SS2/21, UK GDPR, the EU AI Act, and the CLOUD Act. Mickai runs on-prem and air-gapped, inside the firm's own walls, so autonomy and compliance no longer trade off.

Are the patents granted?

They are filed, not granted: 104 UK patent applications with roughly 2,340 claims held by Mickai LTD. Filing establishes priority and a prior-art position across the verifiable-action and sovereign-audit architecture.