MAQPNA

Banks and financial services

Agents for payments and KYC, with four eyes on every transfer.

A wrong transfer is irreversible and your regulator will ask who approved it. MAQPNA runs banking agents in isolated sandboxes on your infrastructure, holds high-impact calls for two approvers and records every decision in a ledger anyone can verify.

Why banking agents stall before production

The model is not the blocker. Control, evidence and accountability are.

  • Irreversible actionsAn agent that can initiate a payment or close an account can do it wrongly, or be talked into it by a crafted email or document.
  • Four-eyes rulesInternal controls need two people on high-value actions, and an agent is not one of them.
  • DORA and ICT third-party riskEvery MCP server and model endpoint an agent uses is an ICT third party you must register, monitor and evidence.
  • Customer data in promptsAccount numbers, IBANs and personal data must not leak into tool arguments, model prompts or logs.

How MAQPNA solves it

Controls the agent cannot bypass

  • Four-eyes approval

    Maker-checker for AI agents: a rule holds a transfer above your threshold for two different approvers. Approvers cannot approve calls made on their own behalf.

  • User approval on their own device

    The customer or employee the agent acts for confirms the action on their own device (CIBA).

  • Per-argument policy

    Typed conditions on amounts, currencies and counterparties. Default deny for anything you have not allowed.

  • Data loss prevention (DLP)

    The gateway finds and blocks or redacts IBANs, card numbers, keys and personal data in arguments, results and prompts.

  • Third-party register

    Export the DORA Art. 28 register of MCP servers and model routes your agents used, built from the ledger.

  • Kill switch

    Revoke an agent, session or user across the installation within about a second, with the reason recorded.

Architecture

A payments-operations agent, governed end to end

The agent runs in a confidential VM in your data centre. It reaches core banking only through the gateway, which checks every call against your policy before it leaves.

Your data centre · jurisdiction EU-DE · your keys (HSM)

Sandboxes

tier-2 (confidential VM, attested)

  • payments-ops
  • kyc-review
  • reconciliation

MAQPNA gateway

identity · policy · DLP · budgets · approval · audit

  • Allowedcore.get_accountread-only-tools
  • Held for approvalcore.create_transferfour-eyes above €10,000
  • Deniedcrm.export_customersdefault deny

Your systems

  • MCPcore banking
  • MCPKYC provider
  • modelin-country model route
Hash-chained audit ledger → write-once storage (WORM) · DORA third-party register export
Illustrative names. You write your own policies.

Evidence

Evidence for your supervisors

Controls and evidence, mapped. Not a certification.

  • SR 26-2 and OCC Bulletin 2026-13

    the US model risk guidance that replaced SR 11-7 on 17 April 2026. Most commentators read it as leaving generative and agentic AI out of scope, with a request for information to follow; one reads it the other way. Under either reading, the ledger shows which agents ran, which tools they called and who approved.

  • DORA

    ledger evidence of ICT operations, backups and recovery drills (maqpna dr drill), key rotation and the Art. 28 third-party register. MAQPNA runs on your infrastructure, so the runtime adds no hosted service to your ICT third-party chain.

  • MAS AI Risk Management Guidelines

    final since 7 October 2026, applying in phases from October 2027 and October 2028, with agentic-specific guidance due for consultation in 2027. Approvals, policies as versioned code and the ledger map to its human-oversight, change-management and inventory expectations.

  • RBI FREE-AI

    the committee framework asks for audit trails and attention to third-party AI providers. The hash-chained ledger and the third-party register help produce that evidence.

  • SEBI (expected, not yet published)

    final AI/ML rules are expected to require human oversight and kill switches, but had not been published by 31 October 2026. Held calls and the kill switch are the controls such rules describe.

  • EU AI Act Art. 12 and 14

    tamper-evident record-keeping and human oversight through approvals and the kill switch. Under Regulation (EU) 2026/1744, Annex III high-risk duties, such as credit scoring, apply from 2 December 2027. How the EU AI Act applies to agents.

  • GDPR

    DLP redaction, encrypted argument capture (off by default) and memory erasure with a signed certificate.

FAQ

Common questions

What does MAQPNA do for banks and financial services?

A wrong transfer is irreversible and your regulator will ask who approved it. MAQPNA runs banking agents in isolated sandboxes on your infrastructure, holds high-impact calls for two approvers and records every decision in a ledger anyone can verify.

What stops banks and financial services from putting AI agents into production?

Irreversible actions: An agent that can initiate a payment or close an account can do it wrongly, or be talked into it by a crafted email or document. Four-eyes rules: Internal controls need two people on high-value actions, and an agent is not one of them. DORA and ICT third-party risk: Every MCP server and model endpoint an agent uses is an ICT third party you must register, monitor and evidence. Customer data in prompts: Account numbers, IBANs and personal data must not leak into tool arguments, model prompts or logs.

Which controls does MAQPNA enforce?

Four-eyes approval: Maker-checker for AI agents: a rule holds a transfer above your threshold for two different approvers. Approvers cannot approve calls made on their own behalf. User approval on their own device: The customer or employee the agent acts for confirms the action on their own device (CIBA). Per-argument policy: Typed conditions on amounts, currencies and counterparties. Default deny for anything you have not allowed. Data loss prevention (DLP): The gateway finds and blocks or redacts IBANs, card numbers, keys and personal data in arguments, results and prompts. Third-party register: Export the DORA Art. 28 register of MCP servers and model routes your agents used, built from the ledger. Kill switch: Revoke an agent, session or user across the installation within about a second, with the reason recorded.

Which regulations does it supply evidence for?

SR 26-2 and OCC Bulletin 2026-13: the US model risk guidance that replaced SR 11-7 on 17 April 2026. Most commentators read it as leaving generative and agentic AI out of scope, with a request for information to follow; one reads it the other way. Under either reading, the ledger shows which agents ran, which tools they called and who approved. DORA: ledger evidence of ICT operations, backups and recovery drills (maqpna dr drill), key rotation and the Art. 28 third-party register. MAQPNA runs on your infrastructure, so the runtime adds no hosted service to your ICT third-party chain. MAS AI Risk Management Guidelines: final since 7 October 2026, applying in phases from October 2027 and October 2028, with agentic-specific guidance due for consultation in 2027. Approvals, policies as versioned code and the ledger map to its human-oversight, change-management and inventory expectations. RBI FREE-AI: the committee framework asks for audit trails and attention to third-party AI providers. The hash-chained ledger and the third-party register help produce that evidence. SEBI (expected, not yet published): final AI/ML rules are expected to require human oversight and kill switches, but had not been published by 31 October 2026. Held calls and the kill switch are the controls such rules describe. EU AI Act Art. 12 and 14: tamper-evident record-keeping and human oversight through approvals and the kill switch. Under Regulation (EU) 2026/1744, Annex III high-risk duties, such as credit scoring, apply from 2 December 2027. How the EU AI Act applies to agents. GDPR: DLP redaction, encrypted argument capture (off by default) and memory erasure with a signed certificate. MAQPNA supplies controls and evidence; it does not certify compliance.

Where does it run?

The agent runs in a confidential VM in your data centre. It reaches core banking only through the gateway, which checks every call against your policy before it leaves. The same runtime runs on a laptop, in your cloud account, on Kubernetes or air-gapped.

Does SR 26-2 cover agentic AI?

The US model risk guidance that replaced SR 11-7 on 17 April 2026. Most commentators read it as leaving generative and agentic AI out of scope, with a request for information to follow; one reads it the other way. Under either reading, the ledger shows which agents ran, which tools they called and who approved.

How do the MAS AI Risk Management Guidelines apply to AI agents?

Final since 7 October 2026, applying in phases from October 2027 and October 2028, with agentic-specific guidance due for consultation in 2027. Approvals, policies as versioned code and the ledger map to its human-oversight, change-management and inventory expectations.

Will SEBI require a kill switch for AI?

Final AI/ML rules are expected to require human oversight and kill switches, but had not been published by 31 October 2026. Held calls and the kill switch are the controls such rules describe.

Put your first banking agent under control.

Policy, approval flow and evidence, walked through with you.

More install options
curl -fsSL https://maqpna.com/install.sh | sh