Private AI workflows for banks

Overview

Metis supports banking teams with source-grounded workflows for KYC, AML, payments and trade operations. Human approval gates keep accountability with named reviewers, in line with the expectations of regulated institutions.

The workflows run on the Metis ARX private core, so customer data, documents and decisions stay inside the bank's own environment. AI handles the high-volume, rules-driven processing work, and people keep every judgement that carries regulatory or financial risk.

Why banks need a different approach

Banks are expanding their use of AI, but they operate under strict requirements on governance, accountability, auditability, data protection and consumer protection. The board owns every automated decision, and that responsibility cannot be passed to a vendor. A workflow is only useful to a bank if it can show who approved what, on which evidence, and under which rule.

Metis is built around that test. Each workflow is governed by workflow state, hard rules, audit logs and human approval gates. Answers and drafts come only from retrieved sources and carry citations. If no source exists, the system says so and does not guess.

The Metis Agent Operating Model

Instead of one AI trying to do everything, Metis organises work as a team of specialised agents. Each agent handles a defined step, and a workflow orchestrator controls the order, the rules and the hand-offs. A named human reviewer approves the steps that carry risk.

The four workflows below have been designed as end-to-end flows. Exact steps are configured to each bank's policies.

KYC

Agents read onboarding documents, including scanned forms, IDs and handwritten entries, using AI OCR.

Extracted fields are checked against the bank's rules, and missing or inconsistent items are flagged.

Customer documents and policies are summarised, with citations to the source.

A named reviewer approves, rejects or requests more information before the case moves forward.

AML

Agents triage alerts, assemble the supporting data and draft the case narrative.

Hard rules decide which cases must be escalated.

Investigators review the evidence and decide on actions such as blocking a payment, requesting information or closing a case as a false positive.

The audit trail records every step for later examination.

Payments

Agents read payment instructions and supporting documents, and check them against rules and screening results.

Exceptions are routed to a human queue with the full context attached.

Release or hold decisions stay with authorised staff.

Trade finance

Agents reconcile large document packs, such as invoices, bills of lading and SWIFT messages, against letter-of-credit terms.

Discrepancies are flagged with the exact source page and paragraph.

Trade operations staff confirm the findings and decide on the next step.

Where humans stay in control

Human approval gates are a design feature, not an add-on. They sit wherever a decision carries regulatory, financial or customer risk. The reviewer is a named person, the evidence they saw is attached, and their decision is logged. Supervisors are notified when exceptions, low-confidence results (below 90% confidence) or rule mismatches appear.

Built on a private core

Air-gapped deployment on on-premises hardware, with no external calls during inference.

Document vectors and model parameters kept locally.

Role-based access control by department or branch.

An immutable audit trail covering queries, outputs and document access.

PII masking, prompt-injection protection and data-leakage prevention.

Certification to ISO/IEC 27001 and ISO/IEC 42001.

Controls map to the expectations of HKMA, including its high-level principles on AI and its technology-risk guidance, as well as PCPD requirements on personal data. This describes how Metis controls map to those frameworks. It is not regulatory endorsement.

Banks that want cloud AI for general tasks can add Metis AERIX. Its policy routing keeps customer data on ARX and sends only public-safe work to approved cloud models.

What it looks like in practice

An onboarding case arrives with a scanned application. The agents extract the fields, run the checks and screen the customer. They flag one missing proof-of-address document and draft a summary with citations. A named compliance officer reviews the summary, requests the missing document and later approves the case. The bank can later show the full chain: what was received, what the AI did, which rule applied and who decided.

Getting started

Banks can begin with a pilot on a single workflow, and add KYC, AML, payments and trade flows as they scale. A typical implementation runs about 16 weeks from setup to go-live, including rules configuration, user acceptance testing, staff training and hypercare.

To explore the banking workflows or arrange a demo, contact enquiry@metis.info.

About Metis

Metis is a private and hybrid AI platform for banks and other regulated enterprises.