Investigating flagged transactions

The process starts when the ERP flags a transaction for review. An analyst must then open four different systems: the ERP itself, the banking feed, the document archive, and email. They collect relevant documents, read through correspondence and transaction history, and piece together evidence to understand what happened. Once they have gathered everything, they write a short rationale explaining their findings and decide whether to escalate the case to the controller or close it. Every decision must be documented in a way that an auditor can reconstruct the reasoning years later, and any personal data must be handled in compliance with GDPR.

The team handles around 1,800 flagged transactions per year, spending an average of 70 minutes on each case. Most cases follow a standard pattern and could be resolved quickly, but the manual work of opening systems, hunting for documents, and writing up the rationale takes time. About 10 to 15 percent of cases are exceptions that require real judgment, either because the transaction pattern is unusual, because earlier data entry errors have surfaced, or because the information across systems is incomplete or conflicting. The analysts are skilled professionals, but they spend most of their week on routine work that feels repetitive and boring.

The hybrid approach is the right fit because this process sits at the intersection of high automation potential and non-negotiable human accountability. The client has excellent API access to the ERP, banking feed, and document archive, and their data engineer already integrates these systems nightly, so the technical foundation is in place. An AI can gather evidence from all four systems automatically, read the correspondence, classify the case as standard or exception, and draft a rationale that matches the team's style. That work alone will save the analysts an hour per case on routine tasks. But the compliance requirements make it essential that a human reviews and approves every decision before it is logged. Regulators expect to see a person's signature on these cases, and the audit trail must show that someone verified the reasoning, not just that a machine produced it.

The hybrid path delivers exactly what the client asked for: automation of the boring parts, with analysts kept in the loop for exceptions and final sign-off. The AI handles the 85 to 90 percent of cases that fit the standard pattern by drafting the rationale and flagging anything unusual, and the analysts spend their time reviewing the output and applying judgment where it matters. This is much faster than the current manual process, but it preserves accountability in a way that a fully autonomous AI agent cannot. An agent would be faster still, but it would expose the client to compliance risk if the model hallucinates a fact or misses a red flag, and there would be no human checkpoint to catch it before the decision is logged.

The client's preference for in-house hosting and their existing data engineering resources make this path even more natural. They can build the orchestration layer and the review interface themselves, host the workflow on their own infrastructure, and use their data engineer to extend the nightly integration to support real-time queries. They will need to integrate an LLM for reading correspondence and drafting rationales, but that can be done via API with models hosted externally or, if they prefer, with an open-source model they run locally. The hybrid architecture gives them control over the data flow, transparency into how decisions are made, and the flexibility to tune classification rules and prompts as regulations or internal policies change.

The tradeoff is that this approach does not eliminate human effort entirely. Analysts will still need to review every case, though the review will be much faster than the current manual investigation. If the client wanted to push further toward full autonomy, they could start by letting the AI auto-close the highest-confidence standard cases and only route borderline cases to humans, but that would require a longer pilot to prove that the model's confidence scores are reliable enough to trust without review. For now, the safer and more compliant path is to require human approval on every decision, and the ROI is still strong even with that gate in place.

The top three contenders for this process are Hybrid, AI Workflow, and AI Agent, but they differ sharply in how they handle accountability and compliance. The Hybrid approach scored highest because it threads the needle between automation and oversight. It uses AI to gather evidence from the ERP, banking feed, and document archive, then drafts a rationale and classifies the case as standard or exception. Crucially, it requires a human to review and approve every decision before it is logged. This keeps the audit trail clean, ensures GDPR compliance, and matches the client's explicit preference for keeping people in the loop on exceptions and final sign-off. AI Workflow scored well but sits in an awkward middle ground. If designed without mandatory approval gates, it risks compliance exposure in a finance context where regulators expect reconstructable reasoning and human accountability. If you add those gates, it effectively becomes the Hybrid approach, just with different branding. The AI Agent path scored lower because full autonomy is inappropriate here. An agent could technically perform the investigation and write rationales, but the compliance requirements and the need for human signatures on decisions make it too risky. The client has made clear that skilled judgment matters for exceptions, and no regulator will accept a black-box AI making final decisions on flagged transactions without human review. RPA and Traditional Code both fall short because they cannot interpret unstructured correspondence or handle the judgment required for the 10 to 15 percent of cases that do not fit the standard pattern. RPA would just automate the clicking, leaving the analytical work untouched. Traditional Code could aggregate data but would need brittle rule engines to handle exceptions, and maintaining that over time would be expensive and fragile. The Hybrid path wins because it automates the boring parts, preserves human accountability where it matters, and aligns with the infrastructure and skills the client already has.

The implementation starts with integrating the three API-enabled systems into a single evidence aggregation service. The data engineer already moves data between the ERP, banking feed, and document archive nightly, so extending that pipeline to support real-time or near-real-time queries is straightforward. When a transaction is flagged, the service pulls the transaction details from the ERP, retrieves related documents from the archive, and fetches corresponding banking feed entries. This milestone delivers a unified view of all evidence without forcing analysts to open four systems manually. Once the data flow is working, the next step is building the AI layer. An LLM like GPT-4 or Claude is integrated to read the aggregated evidence and correspondence, classify the case as standard or exception, and draft a rationale in the team's style. The model is fine-tuned or prompt-engineered using historical closed cases so it learns the decision criteria and writing tone the controller expects. The AI outputs a draft rationale, a confidence score, and a classification flag. This milestone takes the raw evidence and turns it into a structured recommendation, but nothing is final yet. The third piece is the human review interface. A lightweight web application presents each flagged case to an analyst, showing the AI-drafted rationale, the evidence summary, and the classification. Analysts can approve the rationale as-is, edit it, or escalate the case for deeper investigation. Every action is logged with a timestamp, user ID, and the raw evidence payload to satisfy audit trail and GDPR requirements. This interface is where the accountability gate lives, ensuring no decision is recorded without human sign-off. The final phase is a controlled pilot. The hybrid workflow runs in parallel with the manual process for two to four weeks, starting with cases the AI classifies as standard. Analysts give feedback on rationale quality, false positives in classification, and any compliance gaps. Prompts and confidence thresholds are tuned based on real-world performance before rolling out to all 1,800 cases per year. Once live, the system is monitored continuously, with monthly reviews to refine classification rules and update the model as regulations or internal policies evolve.