brianletort.ai
← Library

U.S. Department of the Treasury, Bureau of the Fiscal Service

Company-reportedThreshold is the controlExceptions onlyQueue elimination

Near-real-time machine-learning prioritization of potentially fraudulent Treasury checks

That potentially fraudulent checks must be found through slower post-payment reviews rather than prioritized in near real time.

Financial crime · United States

Collections: Queue eliminated · Regulated autonomy

An editorial scene for U.S. Department of the Treasury, Bureau of the Fiscal Service contrasts review payment and bank information through existing fraud processes. with scores and prioritizes potentially fraudulent treasury checks in near real time. in the near-real-time machine-learning prioritization of potentially fraudulent treasury checks workflow.

Executive brief

The operating-model shift, in one view.

The model did not replace fraud investigators; it changed which transactions reached them first and compressed the time available for recovery.

AI value · Fraud and improper-payment recovery attributed to expedited ML check-fraud identification

Company-reported

$1 billion recovered in FY2024; Treasury separately reported $375 million recovered in FY2023

Treasury-reported attribution. The agency does not publish model precision, false-positive rate, or a causal decomposition from other process changes.

Before

Review payment and bank information through existing fraud processes. → Pursue suspicious payments after identification.

After

Scores and prioritizes potentially fraudulent Treasury checks in near real time. → Investigate prioritized cases and expedite recovery actions.

Human boundary

The model prioritizes risk; authorized officials and financial institutions decide whether and how to investigate or recover funds.

Why it matters

That potentially fraudulent checks must be found through slower post-payment reviews rather than prioritized.

This case is company-reported. Use it for the operating-model shift; do not treat the numbers as independently measured.

How the work changed

Before

How the work ran before the change.

  1. Step 1 of 2

    Payment-integrity analysts

    Review payment and bank information through existing fraud processes.

    ControlRules, referrals, and manual prioritization

  2. Step 2 of 2

    Recovery teams

    Pursue suspicious payments after identification.

    ControlRecovery procedures

What changed

That potentially fraudulent checks must be found through slower post-payment reviews rather than prioritized.

Decision rightHuman sets the threshold; AI decides each instance

After

How the same work runs now.

  1. Step 1 of 2

    Machine-learning fraud process

    Scores and prioritizes potentially fraudulent Treasury checks in near real time.

    ControlRisk-based screening

  2. Step 2 of 2

    Analysts and partner institutions

    Investigate prioritized cases and expedite recovery actions.

    ControlHumans determine investigative and recovery action

Process model built from the published workflow evidence for U.S. Department of the Treasury, Bureau of the Fiscal Service. Every step, actor, and control appears in full below.
Every step, actor, and control

Exception path

Flagged checks proceed through human review and institutional recovery processes; the public source does not disclose thresholds or appeal procedures.

Work removed

  • Undifferentiated review of lower-risk checks
  • Delay between suspicious-pattern emergence and recovery prioritization

Decision authority

The model prioritizes risk; authorized officials and financial institutions decide whether and how to investigate or recover funds.

Before

  1. 01

    Payment-integrity analysts

    Review payment and bank information through existing fraud processes.

    Control: Rules, referrals, and manual prioritization

  2. 02

    Recovery teams

    Pursue suspicious payments after identification.

    Control: Recovery procedures

After

  1. 01

    Machine-learning fraud process

    Scores and prioritizes potentially fraudulent Treasury checks in near real time.

    Control: Risk-based screening

  2. 02

    Analysts and partner institutions

    Investigate prioritized cases and expedite recovery actions.

    Control: Humans determine investigative and recovery action

Work that left the path

  • Undifferentiated review of lower-risk checks
  • Delay between suspicious-pattern emergence and recovery prioritization

Human role before

Analysts detected and prioritized suspicious checks through existing processes.

Human role after

Analysts work a machine-prioritized queue and retain investigative, recovery, and law-enforcement decision authority.

AI roleMachine-learning risk detection that strengthens and expedites identification of potentially fraudulent Treasury checks.

Outcomes

Fraud and improper-payment recovery attributed to expedited ML check-fraud identification

Company-reported

Before the enhanced AI process at the start of FY2023→$1 billion recovered in FY2024; Treasury separately reported $375 million recovered in FY2023

Fiscal years 2023 and 2024 · U.S. Treasury checks processed through the Office of Payment Integrity

Treasury-reported attribution. The agency does not publish model precision, false-positive rate, or a causal decomposition from other process changes.

What leaders can reuse

Anti-pattern

Equating dollars recovered with model precision or allowing risk scores to trigger adverse action without human review.

Questions

  1. 01Does the model only prioritize, or can it block payment?
  2. 02What precision and appeal data should be public?

Portability conditions

  • High-volume transaction data
  • A lawful human investigation and recovery process
  • Monitoring of false positives and disparate effects

Reputation risk

medium

Evidence and authority

What the public record supports.

Current · updated

2 primary; publication outcomes are reported.

Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 37f86dbf3a4048b3

Related transformations

More in Financial crime →

Audi Neckarsulm · Threshold is the control

WPS Analytics full-population inspection

Quality can move from ultrasound sampling to AI screening of every weld.

John Deere (Deere & Company) · Threshold is the control

In-season post-emergence herbicide application (See & Spray targeted spraying)

That a sprayer cannot tell a crop plant from a weed while moving through a field, so the only reliable way to control weeds is to broadcast herbicide across the entire field regardless of where weeds actually are.

Michigan Medicine / Epic Sepsis Model v1 · Threshold is the control

Real-time proprietary sepsis alerts

A deployed alert can add workload while missing most cases.

State Bank of India · Threshold is the control

AI straight-through processing for low-value cheques

A bounded low-value tier can move to touchless processing while assurance shifts to sampling.

Sources

Read the evidence, freshness, caveat, and version policy.