Agent Pageant submission type
maestro-mind
Full name
Magdalena Pozycka-Silva
Team name
RunForrest
Team members
@gil_silva, @magdalena_pozycka
How many agents do you use
multiple-agents
Industry category in which use case would best fit in (Select up to 2 industries)
finance
Complexity level
advanced
Summary (abstract)
In 2025, corporate expense fraud cost companies over $3 billion, a threat severely accelerated by the rise of Generative AI which allows bad actors to create flawless fake receipts in seconds. Traditional manual audits by Finance and Accounts Payable teams only cover about 10% of total corporate spend, leaving a massive financial vulnerability. To eliminate this blind spot, the Agentic Expense Fraud Detection solution leverages a specialized dual AI-agent architecture orchestrated by UiPath Maestro. It provides 100% autonomous audit coverage by semantically extracting receipt data, running GenAI forensic scans, and proactively messaging employees to resolve missing information. By fully automating the audit process, this solution guarantees policy compliance and catches fraud before payouts happen, escalating to a human manager only when absolutely necessary.
Detailed problem statement
The manual processing of employee expenses is fundamentally broken in the modern digital landscape.
The Financial Impact: In 2025, companies lost over $3 billion to expense fraud, with expense reimbursement schemes accounting for 13% of all reported corporate fraud cases.
The GenAI Threat: The core driver of this growing vulnerability is the accessibility of Generative AI. Fraudsters can now generate pixel-perfect, mathematically flawless fake receipts in seconds. Traditional OCR software cannot detect these synthetic artifacts, and human eyes easily miss them.
The Coverage Gap: Because human finance teams are buried under the sheer volume of daily submissions, they typically only have the resources to manually audit approximately 10% of the total spend. This limitation creates a multi-billion-dollar blind spot.
The Organizational Friction: When a discrepancy is found manually, the resulting back-and-forth communication wastes time, creates awkward friction that often drags in Human Resources, and leaves Internal Compliance teams anxious about the 90% of unchecked expenses.
Detailed solution
We replace unreliable manual sampling with complete, autonomous verification. The solution is an end-to-end workflow utilizing UiPath Maestro to orchestrate two specialized AI agents, completely automating the audit cycle.
-
The Trigger
The process begins when an employee submits a receipt and expense details via a custom UiPath Expense Submission App. -
Agent 1: The Receipt Processor (Data & Context)
Triggered instantly, Agent 1 acts as the Insight Specialist. Using UiPath Document Understanding, it performs 100% real-time semantic extraction. It doesn’t just read text; it understands the context of the document and classifies the expense type. It then writes this verified data directly into a centralized UiPath Data Fabric (interacting with the Users and Expenses entities) to establish a manipulation-free single source of truth. -
Agent 2: The Expense Validator (Audit & Action)
Agent 2 acts as the Integrity Officer. Powered by the anthropic.claude-sonnet-4-6 model, it runs advanced, multi-layer validation on the extracted data. It performs duplicate logic checks, cross-references proofs (like calendar events or hotel bookings), and runs GenAI forensic scans to detect synthetic pixels or mathematical impossibilities. -
Autonomous Resolution & Communication
Agent 2 operates with bounded logic to resolve issues without human HR or Finance intervention:
The Happy Path: If the proof is valid and policy-compliant, the expense is approved, and the employee receives a detailed Teams/Slack notification with the approval reasoning and confidence score.
The Exception Path: If a receipt is blurry or proof is missing, Agent 2 autonomously messages the employee via Slack or Teams to request the correct documentation, waiting up to 5 days for a valid reply.
- Human-in-the-Loop Escalation
Strict guardrails ensure the AI never acts beyond its authority. If the employee’s reply is insufficient, the 5-day window times out, the expense exceeds a hard financial limit (e.g., $2,500), or strong GenAI fraud is suspected, the agent seamlessly packages the context and escalates it to a human manager via a UiPath Task App. The manager reviews the AI’s forensic feedback and makes the final decision to approve or reject.
The Outcome: We are automating the entire audit process to enforce policy and catch fraud before funds are ever released.
Narrated video link (sample: https://bit.ly/4pvuNEL)
Expected impact of this automation
Financial ROI: Achieves 100% autonomous audit coverage (up from the typical 10% manual sampling), catching GenAI fakes and preventing fraudulent payouts before funds are released.
Time Saved: Reduces manual expense review by 90% and completely eliminates manual data entry for the Finance and Accounts Payable teams.
Reduction of Manual Tasks: Automates the repetitive back-and-forth of exception handling. The AI agent proactively messages employees via Teams/Slack to resolve missing documentation without human intervention.
Improved Compliance: Ensures zero-bias, 100% consistent policy enforcement while creating a permanent, immutable audit trail for every processed expense.
Accelerated Cycle Times: Approves compliant “happy path” expenses in seconds, dramatically improving the employee experience and eliminating approval bottlenecks.
UiPath products used (select up to 4 items)
UiPath-action-center
UiPath-agent-builder
UiPath-apps
UiPath-maestro
UiPath-studio-web
Automation Applications
Teams, Outlook
Integration with external technologies
Claude
Agentic solution architecture (file size up to 4 MB)
Sample inputs and outputs for solution execution
Inputs:
- Employee identification (Name/Email).
- Expense description.
- The raw receipt or proof of purchase (Image/PDF upload)
Outputs:
- Data record (Data Fabric) with all the extracted receipt details
- Calculated confidence scores and GenAI forensic fraud analysis results
- Autonomous Teams messages for missing info and final approval/rejection message.