Invoice Processing & Approval

The accounts payable team receives supplier invoices by email and occasionally by post. Each invoice must be matched to an open purchase order in SAP, validated for amount and line items, routed for approval (invoices above €5,000 require two manager approvals), and then posted to the SAP payment run. The supplier receives a confirmation once payment is posted and the PDF is archived in SharePoint.

The process runs 6,000 times per year at an average of 22 minutes per invoice. About 20% of invoices hit an exception — most commonly a missing or incorrect PO reference or a price discrepancy — and require manual intervention before they can proceed. The remaining 80% follow a predictable, rule-based path that is well-suited to automation.

Invoice processing lands squarely in AI agent territory for three reasons. First, the inputs are unstructured: invoices arrive as PDFs in dozens of different layouts from hundreds of suppliers. Traditional scripted OCR fails on format variation; Claude's native document understanding handles it reliably. Second, the matching and validation logic involves fuzzy comparison — amounts within tolerance, PO references that are present but slightly malformed — that requires genuine reasoning rather than string equality checks. Third, the exception rate (20%) is high enough that the system needs to make sensible decisions about what to escalate rather than failing on anything unusual.

RPA was considered and rejected. The invoice PDF extraction problem alone would require a separate OCR layer with vendor-specific templates, making the solution brittle and expensive to maintain as supplier formats change. RPA is the right call when the process is entirely UI-based and inputs are already structured — this one is neither.

Your Azure + Python stack aligns well with the recommended approach. The agent is written in Python and can be containerised and deployed to Azure Container Apps or AWS ECS — the Bedrock API call is the only AWS dependency. Everything else (SAP, SharePoint, Exchange) uses standard REST APIs that work identically from either cloud.

Between AI agent (score: 9) and hybrid (score: 8), the difference is a question of risk appetite rather than technical capability. A pure AI agent handles the full 100% of invoice volume autonomously, escalating only genuine exceptions. A hybrid approach sends all flagged invoices to a human queue regardless of whether the agent is confident. For a first deployment, hybrid is the safer choice: AP staff stay in the loop on exceptions, which builds trust and surfaces edge cases faster. Once the team is confident in the agent's accuracy, the human gate can be narrowed to only the highest-value or most ambiguous cases.

RPA (score: 5) is not competitive here because the document extraction problem is genuinely hard for rule-based systems. If all your invoices arrived as structured EDI files, RPA would score higher. They don't.

The agent runs as a Python service triggered by an Exchange webhook when a new invoice email arrives. It calls Claude on AWS Bedrock with the attached PDF, instructing it to extract structured invoice data into a defined JSON schema. This gives you vendor name, invoice number, date, line items, VAT breakdown, and PO reference — all extracted natively without a separate OCR tool.

The extracted PO reference is used to call the SAP REST API and retrieve the matching purchase order. The agent compares invoice amounts to PO amounts with a configurable tolerance (default ±2%). If everything matches and the amount is below €5,000, the agent calls the SAP invoice posting endpoint directly. If the amount exceeds €5,000, it creates an approval request in SharePoint via the Graph API and waits for the approval callback before posting.

Exceptions — missing PO reference, amount outside tolerance, vendor not in SAP master data — are routed to an exception queue. This is a lightweight web interface (or a Teams bot) where AP staff see the flagged invoice, the agent's reasoning, and the available actions: approve anyway, reject, or request correction from the supplier. The agent resumes once a decision is recorded. Every run produces a structured JSON audit log stored in Azure Blob Storage for the 7-year retention requirement.