AI Agents
Discovery Express

Finance & back office

Fraud Review Agent

Works the fraud queue: gathers the evidence, writes the case and leaves the decision to an analyst.

In productionReviewed September 3, 2026
Internal apps · Fraud Review
Demo data

  1. Rules engine

    Alert 88231: card present abroad 40 minutes after a local purchase.

  2. Agent

    The travel notice was filed six days ago and the device is the usual one. My recommendation is benign, confidence 0.82, with both sources linked.

  3. Case file 88231 written · 7 sources linked · read only on the core · no action taken on the account

  4. Handoff to a person

    The analyst closes it in under a minute. Any block, unblock or report is executed by a person, never by the agent.

Internal appsFraud rules engineCore banking / policy adminCard processor
KPI it moves
Review time down 55% to 70%Measured
Channels
Internal appsSlack / Teams
Integrates
  • Fraud rules engine
  • Core banking / policy admin
  • Card processor
  • Your CRM
  • Compliance workspace
A fraud analyst at two monitors at night above Paseo de la Reforma
Finance & back office

What it does

The evidence gathered and the case written. The decision stays with your analyst.

Where it lives

Your cloud, your repo, your IP.

1. Receives 2. Looks up 3. Decides 4. Acts 5. Escalates 6. Learns

Systems it integrates

  • Fraud rules engine
  • Core / policy admin
  • Card processor
  • Your CRM
  • Compliance workspace

Fraud Review Agent

03Decides

Channels

  • Internal apps
  • Slack / Teams

A person on your team

01

Receives

Takes the alert from the rules engine with its score and the reason it fired.

02

Looks up

Pulls the account history, the device, the geography and prior cases that look like this one.

03

Decides

Weighs the evidence for and against, states a recommendation and its confidence in plain language.

04

Acts

Writes the case file with every source linked, so the analyst reads instead of hunting.

05

Escalates

Sends every block, unblock and report to a person. It never touches the account itself.

06

Learns

Compares its recommendation against the analyst's decision weekly and reports where it disagrees.

KPI it moves

Measured, with the source.

Every figure has a row in the Proof Ledger: source, date, owner and client permission.

Proof Ledger
Measured

Measured

55% to 70% lower

Time per manual review

Range observed across deployments of this agent, Sep 2026

Measured

92% to 99%

Queue cleared within SLA

Range observed across deployments of this agent, Sep 2026

Measured

35% to 50%

False positives closed without customer contact

Range observed across deployments of this agent, Sep 2026

Reviewed September 3, 2026

How it is governed

Every action has an owner, an eval and a limit.

Set in the Blueprint, enforced by the Govern layer of the platform.

Govern · Approval queue
Demo data

Waiting for a person

Block card ending 4419 · score 0.91

Case file ready · the analyst decides

Weekly eval

Passing

Agreement with the analyst decision

94.2%

Last 8 runs
91.1% → 94.2%

Evidence traceable to its source in every case

Limit, enforced

0 writes to the core today

Human approvals

  • Every block, unblock or account restriction
  • Any regulatory report before it is filed

Typical evals

  • Agreement with the analyst decision on a weekly sample
  • Evidence completeness and source traceability

Limits

  • Read only on the core; it recommends, it does not act
  • No PII leaves your cloud at any point

Time to production

8 to 12 weeks

of pilot to production under a KPI Contract, after a Blueprint of 1 to 2 weeks

The agent already runs in production elsewhere, so the 2 to 4 weeks of tuning inside the pilot go to your policies, your data and your systems, not to the model.

DaCodes delivery standard, Sep 2026. Integration is the long pole, not the model.

How we deploy
The four phases
Commitment gate

  1. 01Discovery Express

    2 sessions, no cost

  2. 02Blueprint

    1 to 2 weeks

  3. 03Pilot

    8 to 12 weeks

  4. 04Scale & Managed

    Continuous

  • Pilot, first 8 weeks
  • Up to week 12 of the pilot
  • This agent: 2 to 4 weeks of tuning, inside the pilot
  • Commitment gate: the KPI decides Scale

Proof and demo

What backs the label.

Proof Ledger note

In production

Measured in a real deployment and logged in the Proof Ledger with source, date and owner.

Reviewed September 3, 2026

Where this agent sits on the maturity policy

  1. Reference architecture
  2. Pilot
  3. In productionToday

The Discovery Express shows this agent on your data, live. Two sessions, no cost.

See this agent with your data

Related agents

Same function or industry.

All agents
A data steward comparing two printed customer records with a highlighter
Reference architecture

Finance & back office

Master Data Agent

KPI it movesTypical target

Duplicates down 80 to 95%

FAQ

Direct answers.

No. It has read access to the core and writes only the case file. Every block, unblock and report is executed by a person, and the platform enforces that.

No, it works the queue your engine produces. The engine keeps deciding what to flag; the agent removes the hours spent assembling evidence for each flag.

Its recommendation is compared with the analyst decision every week, and disagreement is reported by segment. Cases where it was wrong become eval cases.

Which process do we automate first?

Two discovery sessions, no cost, and you leave with two or three quick wins whether or not you build with us.

We reply in under one business day.

Loading the form