Developer finds AI invoice agent fails badly on real-world messy enterprise data

A software developer tested an AI-powered invoice-processing agent on real corporate invoice data and found it failed in multiple critical ways that clean test scripts had not revealed. The agent misread legitimately correct invoices due to OCR noise and inconsistent number formatting from legacy vendor portals, requiring a dedicated normalization layer to fix. A memory module incorrectly auto-approved a mismatched invoice by over-relying on a similar past resolution, prompting the developer to add metadata guardrails that force human review when key identifiers like purchase order numbers do not align. The agent also entered infinite tool-calling loops when encountering undocumented surcharges, which was resolved by imposing circuit breakers that cap retries and escalate unresolved cases to human reviewers. The developer concluded that LLMs should handle contextual reasoning while deterministic code manages arithmetic, and that agent memory must be treated as a hint rather than a decision-making shortcut.