Fraud Detection Agent
Detect duplicate invoices, phantom vendors, expense fraud and AI-fake invoices.
Detects duplicate invoices, phantom vendor patterns, unusual posting patterns, AI-generated fake invoices, expense fraud and round-tripping via ML analysis and escalates suspected cases to the compliance officer.
Score Dashboard
What This Agent Does
Fraud in financial accounting causes billions in damages worldwide. Methods are becoming more sophisticated: beyond classic duplicate invoices and phantom vendors, AI-generated fake invoices are increasingly used - deep-fake PDFs indistinguishable from real invoices at first glance.
The Decision Layer combines rule-based and ML-based detection. Exact duplicates and segregation-of-duties violations are detected rule-based. Phantom vendor patterns (vendor without genuine business relationship), unusual posting patterns (Friday evening, threshold splitting), expense fraud and round-tripping use ML anomaly detection. AI-generated fake invoices are detected by LLM analysis of document authenticity.
The result: every transaction receives a risk score. Suspected cases are escalated to the compliance officer. False positives are assessed by humans - the investigation decision always remains with the human.
Micro-Decision Table
Detect duplicate invoices Is there a duplicate or slightly varied invoice? Rules Engine Vendor
Exact duplicates = R, variants (slightly changed vendor) = A
Decision Record
Challengeable: Yes - rule application verifiable. Objection possible for incorrect data or wrong rule version.
Challengeable by: Vendor
Phantom vendor detection Are there vendors without genuine business relationships? AI Agent Vendor
Pattern analysis of order history and vendor activity
Decision Record
Challengeable: Yes - fully documented, reviewable by humans, objection via formal process.
Challengeable by: Vendor
Unusual posting patterns Are there postings at unusual times or with threshold splitting? AI Agent Auditor
ML anomaly detection against historical behaviour patterns
Decision Record
Challengeable: Yes - fully documented, reviewable by humans, objection via formal process.
Challengeable by: Auditor
Detect AI fake invoices Is the document an AI-generated forgery? AI Agent Vendor
LLM analysis of document authenticity, metadata check
Decision Record
Challengeable: Yes - fully documented, reviewable by humans, objection via formal process.
Challengeable by: Vendor
Detect expense fraud Is there a duplicate submission or inflated amount? Rules Engine Employee
Rule violations = R, pattern recognition = A
Decision Record
Challengeable: Yes - rule application verifiable. Objection possible for incorrect data or wrong rule version.
Challengeable by: Employee
Round-tripping detection Are there circular money flows? AI Agent Auditor
Network analysis of payment flows
Decision Record
Challengeable: Yes - fully documented, reviewable by humans, objection via formal process.
Challengeable by: Auditor
Segregation-of-duties violations Is the requester, approver and payer the same person? Rules Engine Auditor
Authorisation matrix matching
Decision Record
Challengeable: Yes - rule application verifiable. Objection possible for incorrect data or wrong rule version.
Challengeable by: Auditor
Calculate risk score How high is the fraud risk of this transaction? AI Agent
ML-based scoring from all detection modules
Decision Record
Challengeable: Yes - fully documented, reviewable by humans, objection via formal process.
Alert to compliance officer Must a suspected case be investigated? Human Auditor
Investigation decision requires human judgement
Decision Record
Challengeable: Yes - via manager, works council, or formal objection process.
Challengeable by: Auditor
False positive assessment Is this a genuine suspected case or a false alarm? Human
Judgement in assessing the overall picture
Decision Record
Challengeable: Yes - via manager, works council, or formal objection process.
Decision Record and Right to Challenge
Every decision this agent makes or prepares is documented in a complete decision record. Affected parties (employees, suppliers, auditors) can review, understand, and challenge every individual decision.
Prerequisites
- Access to transaction data from ERP (postings, orders, payments)
- Access to vendor master data and order history
- Authorisation system with SoD matrix
- Configured thresholds for risk scores and escalation
Governance Notes
GoBD-relevant: fraud detection processes tax-relevant transaction data. The results - especially suspected cases and investigation outcomes - are sensitive data and must be treated confidentially.
For professional secrecy holders (Paragraph 203 StGB), suspected cases must not be disclosed to third parties. LLM inference for document authenticity checking must run in EU data centres. The agent reports suspected cases exclusively to the internal compliance officer. The investigation decision always remains with the human.
§203 StGB-relevant data is encrypted end-to-end and never passed to AI models in plain text.
Process Documentation Contribution
Infrastructure Contribution
The Fraud Detection Agent is the most A-intensive agent in the entire catalog. It uses the anomaly detection of the ICS Monitoring Agent and transaction data from all AP/AR agents. The ML scoring framework is reused for risk assessments in other domains. The document authenticity check becomes the standard for all incoming documents.
Builds Decision Logging and Audit Trail used by the Decision Layer for traceability and challengeability of every decision.
Related Agents
ICS Monitoring Agent
Monitor internal control system - four-eyes, segregation of duties, detect control gaps.
Procedural Documentation Agent
Keep procedural documentation automatically current - detect changes, generate drafts, close gaps.
Annual Statement Preparation Agent
Prepare annual financial statements - orchestrate checklist, draft notes, answer auditor queries.
Frequently Asked Questions
How high is the false positive rate?
In the initial phase, the false positive rate is typically 15-25%. With increasing training volume and feedback loops, it drops to 5-10%. Human assessment of every suspected case ensures no unjustified consequences are drawn.
Can the agent also detect internal fraud cases?
Yes. Segregation-of-duties checks, threshold splitting and posting time analysis explicitly target internal patterns. Round-tripping detection identifies money flows potentially used to conceal internal transactions.
Are detected suspected cases automatically reported to authorities?
No. The agent reports suspected cases exclusively to the internal compliance officer. The decision on further steps - internal investigation, criminal complaint, reporting to supervisory authorities - remains with the human. For Paragraph 203-relevant cases, additional confidentiality requirements apply.
Implement This Agent?
We assess your finance process landscape and show how this agent fits your infrastructure.