Fraud Detection Agent

Organisations lose an average of five percent of annual revenue to fraud. The Association of Certified Fraud Examiners quantifies this in its 2024 Report to the Nations, based on 1,921 investigated cases with a total damage of USD 3.1 billion (approx. EUR 2.9 billion). Most of these cases were not discovered by internal controls but by tips. Rule-based checking catches what it knows. What it does not know stays invisible - often for years.

Sampling-Based Audits Fail Against Deliberate Concealment

Sampling-based auditing rests on a core assumption: if a sufficiently large share of transactions is correct, you may infer the same for the whole. Fraud invalidates that assumption. A phantom vendor posting small amounts just below the approval threshold for 18 months never shows up in any sample. Threshold splitting - an invoice for EUR 9,900 (USD 10,700) instead of EUR 10,000 to bypass the approval tier - looks unremarkable in isolation.

Only full-population analysis makes these patterns visible. Not through stricter rules, but through statistical anomaly detection: which vendor has no purchase orders from procurement yet regular payments from AP? Which cost centre posts on Friday evenings when nobody reviews? Rule-based systems never ask these questions because nobody formulated them as rules.

AI-Generated Documents Shift the Threat Landscape

Until 2024, forged invoices were detectable by craftsmanship - wrong fonts, missing stamps, inconsistent tax IDs. That has fundamentally changed. AI-generated documents are today visually indistinguishable from real ones, and anti-fraud professionals across industries report a marked increase in GenAI-generated forgeries over the past two years. Chris Juneau, SVP at SAP Concur, put it plainly: “Do not trust your eyes.”

For Finance departments, this means a new line of defence. AI-generated fakes survive visual review and often pass rule-based validation as well. What gives them away are metadata inconsistencies, atypical document structures and statistical anomalies in context - such as a new vendor whose first invoice exactly matches the amount pattern of an existing vendor. This analysis requires AI trained on document authenticity.

Ten Decision Steps Separate Signal from False Alarm

The Decision Layer breaks fraud detection into a chain of ten decisions with three different decider types. Three steps are rule-based: duplicate detection, expense fraud against defined thresholds, and segregation-of-duties violations against the authorisation matrix. Five steps use AI analysis: phantom vendors, posting anomalies, document authenticity, round-tripping in payment networks, and aggregated risk scoring. Two steps rest with the human: the escalation decision and the final assessment of whether an alert is justified.

A concrete scenario: a mid-sized manufacturer processing 40,000 inbound invoices per year. The agent detects that a packaging supplier has been submitting invoices with the same net amount for six months but slightly varying item descriptions. Simultaneously, network analysis reveals the supplier’s bank account is linked to an employee in procurement. Neither signal alone constitutes proof. Combined, they produce a risk score that triggers escalation to the compliance officer.

The Compliance Officer Decides - Not the Algorithm

Six of ten decision steps use AI analysis. That makes this agent the most AI-intensive in the entire catalogue. Yet no algorithm decides whether a suspicion becomes an investigation. The Decision Layer documents every alert with trigger pattern, affected transactions, risk score and timestamp. The compliance officer assesses on this basis whether a false positive is present or an investigation is warranted.

This principle is not optional. ISA 240 obliges the statutory auditor to assess fraud risks. A demonstrably documented detection system - with logged alerts, escalation paths and investigation outcomes - is a concrete signal of the internal control system’s effectiveness. (US: SOX Section 404 imposes comparable documentation requirements for internal controls over financial reporting.) Comprehensive alert documentation strengthens not only the defence but also audit readiness under professional auditing standards and the requirements for early risk detection systems.