ProcureGuard is presented as a cloud-native fraud detection system designed to catch a classic procurement abuse: splitting a large purchase into smaller invoices so each one slips beneath approval thresholds. In the Medium post, its creator argues that this kind of tactic can leave finance teams exposed because conventional accounts payable tools tend to assess invoices one by one, rather than as part of a broader purchasing pattern.
The system is built around four autonomous ...
Continue Reading This Article
Enjoy this article as well as all of our content, including reports, news, tips and more.
By registering or signing into your SRM Today account, you agree to SRM Today's Terms of Use and consent to the processing of your personal information as described in our Privacy Policy.
agents running on Google Cloud. The first agent handles invoice ingestion, pulling in raw JSON or scanned documents, validating required fields and normalising dates and currency values. A second agent then scans transaction history over a rolling 30-day period to identify whether invoices tied to the same vendor or purchase order have been structured to evade authorisation rules. The post says this analysis is powered by Vertex AI and Gemini 2.5 Flash.
A third agent acts as a challenger, testing whether a suspected pattern could instead be explained by legitimate procurement behaviour such as milestone billing or separate departmental orders. That adversarial step is meant to reduce false positives, which remain a persistent weakness in many rule-based detection systems. A fourth agent compiles the findings into an evidence package for auditors, compliance staff and senior finance executives, including a risk assessment and a record of the supporting transactions.
The architecture combines Python services with event-driven infrastructure. Invoices are triggered through Google Cloud Pub/Sub and passed to a Flask-based microservice hosted on Cloud Run, where the pipeline can run asynchronously. The system also uses BigQuery for audit logging and Firestore for live case state. The creator says the deployment is designed around least-privilege access, using Application Default Credentials rather than stored local key files.
The post also places ProcureGuard in a wider Google Cloud context. Google’s own procurement document tools are designed to extract data from invoices and receipts, while its fraud defence products focus on real-time risk analysis across large volumes of user interactions. In that sense, ProcureGuard appears to combine document extraction, behavioural analysis and case management into a single procurement-specific workflow.
According to the post, the system was tested against a synthetic dataset of 300 invoices, including normal transactions, legitimate multi-invoice purchases and planted split-invoice cases. It claims perfect precision, recall and F1 score in that benchmark, with no false positives. Those results, if reproduced in live environments, would be notable, though synthetic evaluations often flatter systems that have yet to face the complexity of real procurement data.
Even so, the broader appeal of the project lies in its framing of fraud detection as a reasoning problem rather than a static rules exercise. Instead of asking only whether one invoice crosses a threshold, the system tries to infer intent from linked payments, vendor history and approval context. That is a more ambitious model of financial control, and one that reflects the growing interest in agentic AI systems for enterprise risk management.
Source: Noah Wire Services
