Skip to content
unmatrixai

Lending, onboarding, and compliance

AI agents for banking: lending, onboarding and compliance

AI agents for banking are systems that carry regulated work such as KYC refresh, credit file preparation, reconciliation and complaint handling from intake to a decision a person can approve, with every step logged. Unmatrix AI builds them inside the bank's own environment, under its model-risk and data-residency rules, with forward deployed engineers sitting with the operations team that owns the workflow.

Last updated

Workflows

What do we build for banking?

The workflows we usually start with, and the architecture each one is built on.

  • KYC and periodic review

    Gather registry, sanctions and ownership evidence, compare it with the file, and route only the changes to a compliance analyst.

    Built as

    Autonomous agents (ReAct)

  • Lending file preparation

    Extract and verify income, collateral and covenant data from documents, flag gaps, and assemble the credit memo for the underwriter.

    Built as

    LangGraph workflow agents

  • Reconciliation breaks

    Investigate unmatched transactions across ledger, bank feed and trade systems, classify the break and draft the journal entry.

    Built as

    Autonomous agents (ReAct)

  • Complaints and regulatory correspondence

    Classify, draft and route responses with citations to policy and the customer's record, with deadlines tracked.

    Built as

    RAG agents

Constraints

What shapes the build?

Model risk management
Agents are documented, evaluated and monitored like any other model under an SR 11-7 style framework. We deliver the model card, eval results and monitoring hooks your MRM team expects.
Data residency
Deployment in your cloud region or on-prem, with open-source models where hosted APIs are not approved.
Auditability
Every read, decision and write is logged with the model version and retrievable for an examiner.
Segregation of duties
Agents draft and recommend. Approval rights stay with the roles that hold them today.

Worked example

A typical build: periodic KYC refresh for corporate clients

A commercial bank refreshes thousands of corporate KYC files a year. Analysts spend most of the time collecting evidence that has not changed.

  1. 01

    Trigger

    A file reaches its refresh date and enters the queue with its risk tier.

  2. 02

    Collect

    The agent pulls registry filings, LEI data, sanctions and PEP screening and current ownership from the sources your policy names.

  3. 03

    Compare

    It diffs the evidence against the file. Unchanged files are completed with an evidence pack attached.

  4. 04

    Escalate

    Director changes, ownership changes and screening hits route to an analyst with the difference highlighted.

  5. 05

    Record

    The case is updated in Fenergo with the evidence, the decision and the analyst's name where one was involved.

Analysts see only the files with changes, with the evidence already assembled. The refresh backlog clears and every decision is examiner-ready.

Integration

Which systems do we work with?

  • Temenos
  • FIS
  • Finastra
  • nCino
  • Fenergo
  • Salesforce Financial Services Cloud
  • ServiceNow
  • World-Check

Anything with an API or a file interface can be integrated. These are the systems we meet most often in banking.

Questions

What banking teams ask

Can an AI agent make credit or compliance decisions?

It should not, and ours do not. Agents gather, verify, classify and draft. The decision stays with the accountable person, with the agent's work attached.

How do you satisfy model risk management?

We treat the agent as a model: documented purpose, an eval set, performance thresholds, monitoring and change control. The MRM team gets the artefacts before go-live, not after.

Does customer data leave the bank?

No. Models run in your cloud account or on-prem, and nothing is used for training.

Which core banking systems can you integrate with?

Any with an API or a file interface. Temenos, FIS, Finastra and nCino are common, alongside Salesforce and ServiceNow.

Related

Next step

Pick one workflow. We will be at your office in two weeks.

A thirty-minute call to find the right first workflow, followed by a written scoping note. No deck, no pilot.

Book a call