Detecting and Preventing Fraud and Anomalies in Finance with Agentic AI

Introduction: How Agentic AI Detects and Prevents Fraud in Finance in Real-Time

Deepak: Good evening, everyone. My name is Deepak, and I'll be moderating today's session. Our panelists are Jon Naseeth, Dave Sackett, and Rajeev Pathak. This will be a 60-minute session, with 45 minutes of discussion and a 15-minute Q&A at the end. Feel free to drop your comments as we progress through the session.

Let me start by introducing our panelists. Jon Naseeth is the CFO and founder of Cantu Capital Link. With a background in AI and machine learning at Google, Jon excels in delivering solutions that address social and economic needs. Welcome, Jon.

Next, we have Dave Sackett, VP of Finance for Personal Technologies, a FinTech speaker, and a Forbes writer. He shares insights on AI, blockchain, and e-commerce and is committed to lifelong learning and kindness. Dave also co-founded AI1 to boost e-commerce sales. Hi, Dave, it’s great to have you.

Lastly, we have Rajeev Pathak, CEO and co-founder of Hyperbots Inc. With over 30 years of experience in building technology products and businesses for global markets, Rajeev has managed the technology business for Wipro as GM and vertical head, handling a business of over $100 million. Welcome, Rajeev.

Good morning, Jon and Dave, and good evening, Rajeev. Hello to everyone in the audience.

In the next hour, we will explore how agentic AI plays a huge role in detecting and preventing fraud and anomalies in finance and accounting. Our discussion will be divided into three sections:

1. Real-world frauds and why traditional defenses fall short.

2. How agentic AI detects, decides, and prevents fraud and anomalies in real-time.

3. Practical tips for integrating agentic AI into finance workflows.

We'll conclude with time for questions and comments. Let’s dive in.

Real-World Financial Frauds and Why Traditional Systems Fail to Prevent Them

Top 5 Costliest Finance Fraud Types Detected by AI

Jon: First, thank you for the introduction and the opportunity to be here. The most impactful and costliest fraud types can vary. A lot of times, things that make it into the news aren't the costliest. As a certified fraud examiner, I’ve seen that large-dollar incidents often don’t make the headlines because companies prefer to avoid public exposure.

From a cost perspective, here are the top five:

• Vendor Payment Fraud: Fake or inflated invoices are common and cause significant losses. Recently, I saw that the U.S. government is finally implementing proper vendor payment authorization after an issue surfaced.
  
  

• Revenue Leakage (Order to Cash): Hidden discount abuse often leads to financial losses.
  
  

• Business Email Compromise: This type of fraud is a significant threat.
  
  

• Indirect Tax Carousel Schemes: Occur when wrong tax codes are applied; intentional fraud if done deliberately.
  
  

• FP&A Model Manipulation: Financial Planning and Analysis (FP&A) can manipulate data, which can lead to fraud, especially when presented to investors.
  
  

These types of fraud happen frequently, and the consequences can be severe.

How Mid-Market Companies Become Vulnerable to Vendor Payment Fraud and How AI Can Help

Deepak: Very interesting. Rajeev, staying with P2P for a moment—what makes mid-market companies a soft target for duplicate payment or over-billing scams?

Rajeev: Mid-market finance teams are usually very lean. In P2P or accounts payable, a small team may be managing hundreds of vendors and thousands of invoices, so bandwidth is limited. Human-driven processes increase the risk of errors or fraud.

Vendor information often moves between departments through email in Excel files. Files can be intercepted, modified, and sent with altered information that appears legitimate, such as a fake invoice.

In the absence of significant oversight, there is a risk at nearly every point in the processing pipeline. Until there's deep automation and robust security, these systems are highly vulnerable to fraud.

Deepak: Dave, why do rules-based receivable systems miss revenue leakage and fictitious credit memo fraud?

Dave: Sales orders may be recorded to delay or shift revenue outside the audit scope. Credit memos can be backdated to move transactions to a future period. Fraudsters may use small credit memos below materiality thresholds to avoid detection. If unnoticed, these actions escalate over time.

Treasury Fraud Detection with AI: Protecting Bank Portals in Real-Time

Deepak: Jon, treasury is a hot target. Can you share a case where traditional bank portal approvals still fail?

Jon: Someone spoofed my lender’s call and used AI to mimic my voice to approve a fake loan. We now use a keyword system to confirm identity. Without real-time AI monitoring, such fraud can slip through easily.

Deepak: Rajeev, what tax-related anomalies slip past standard ERP tax engines?

Rajeev: ERP tax engines validate sales tax based on dictionaries, but can’t check context. Outdated exemption certificates or misclassified invoices can lead to penalties. Agentic AI can continuously monitor these anomalies and flag errors in real-time.

Dave: Overly optimistic forecasts, such as reducing churn unrealistically, can mislead stakeholders. Agentic AI can monitor FP&A models, detect discrepancies, and alert finance teams before errors impact decision-making.

Jon: Unusual approval requests late at night or early in the morning can indicate anomalies. AI can spot these subtle signals that humans might miss.

How Agentic AI Detects, Decides, and Prevents Fraud in Real-Time

Differences Between Agentic AI and Predictive Models for Fraud Detection

Dave: Standard predictive models calculate probabilities and require human intervention. Agentic AI owns the full loop, makes decisions based on policies, and prevents fraud in real-time, acting like a junior accountant or financial analyst executing tasks automatically.

Rajeev: Stacking multiple agents in specialized roles produces even stronger fraud detection outcomes.

Rajeev: AI agents can read invoice emails rigorously, eliminate unauthorized vendors, check for duplicate invoices, and cross-check for manipulations. Matching agents can perform automatic two-way and three-way matching against ERP ground truth, flagging inconsistencies in real-time.

O2C Fraud Prevention with Agentic AI

Dave: Agentic AI verifies credit memos against warehouse management systems and ledger entries. If a credit memo has no basis, it stops the process and flags it, preventing fraud in real-time.

Jon: AI identifies incorrect vendor tax codes or risk scores. Shared databases and cross-system monitoring ensure tax anomalies are flagged before they result in penalties.

Dave: Agentic AI performs flux analysis between FP&A models, detects unauthorized changes, and ensures forecasts remain accurate. It reduces manual auditing needs and prevents hidden manipulation.

Practical Tips for Integrating Agentic AI into Finance Workflows

Jon: Clean master data—vendor, customer, and IDs—is essential. Databases must communicate properly, and audit trails should be enabled to maintain traceability and minimize risk.

Dave: Deploy AI where fraud detection pain is highest—high-volume transactions with significant risk. This ensures early ROI and impactful results.

Rajeev: Balance cost and risk. Focus on functions with historical risk like AP tax compliance or large invoice processing tasks.

Rajeev: Communicate AI agents as fraud prevention and productivity tools, not replacements. Demonstrate benefits like faster invoice processing or more accurate month-end closing to gain trust and adoption.

Jon: Maintain transparent audit trails for all agent actions. Ensure policies, rules, and system access are documented so auditors and regulators can verify processes.

Rajeev: Transparency in AI rules and system interactions avoids black-box issues. Auditors need clarity on agent operations.

Dave: Agents can be orchestrated externally, reading multiple databases without direct manipulation. This allows for real-time fraud detection and decision-making across disintegrated systems.

Rajeev: AI agents identify correlations and anomalies across systems, preventing large-scale fraud proactively.

Jon: KPIs include number of anomalies detected, money saved, and false positives prevented. Prevention itself is often the best measure.

Rajeev: Different organizations will define KPIs differently; measure impact relative to risk and operational efficiency.Dave: Balance trial-and-error with internal resources to train the system. Gradually reduce false positives and improve detection accuracy over time.

Rajeev: Teams must monitor AI outputs, handle exceptions, and formulate policy for AI-driven processes. Trusting AI while maintaining oversight is a key skill for modern finance teams.

Q&A: Real-World Examples of Fraud Detection with Agentic AI

Jon: AI verifies data in ways humans cannot, catching subtle fraud patterns like below-threshold transactions or manipulated records.

Critical Data Sources for Effective AI Fraud Detection

Jon: Monitor all relevant internal and external data sources, drilling down into areas that show red flags. Fraud often follows predictable patterns.

Rajeev: External data sources require extra attention as fraudsters often exploit payment channels.

Jon: A technology company overpaid rebates due to multiple returns being claimed repeatedly. Multi-variable AI algorithms identified discrepancies that rule-based systems missed.

Rajeev: Implement a maker-checker model where AI agents detect anomalies and humans verify, and vice versa. Collaboration ensures checks and balances in fraud prevention.

Conclusion: The Future of Fraud Prevention with Agentic AI

Deepak: Thanks to Jon, Dave, and Rajeev for sharing insights. The future of finance will rely on AI agents to prevent fraud, detect anomalies, and improve operational efficiency while humans maintain oversight.

All: Thanks. Goodbye!