Aralab

Invoice processing sounds simple until you're doing 900 a month from suppliers who each format their invoices differently. Traditional OCR breaks on anything non-standard. The problems were structural:

• No standard format. Every supplier had their own layout — different column positions, different naming conventions, different levels of detail.
• Three-way matching is fragile at volume. Matching invoice lines to PO lines to warehouse receipts requires tolerating description mismatches, partial deliveries, and pricing rounding — at speed, without errors.
• Discrepancies need judgment. Not every mismatch is an error. Some need approval, some escalation, some supplier follow-up. The system had to know the difference.

We built an autonomous AI invoice agent on Claude Sonnet 4.5 and GCP:

• Extraction — Claude reads any PDF invoice format and extracts line items, totals, supplier details, and dates without needing a template
• Supplier identification — NIF normalisation and fuzzy matching maps supplier names to ERP records even when they don't match exactly
• Three-way matching — deterministic logic for clean cases, LLM-assisted matching for ambiguous ones (partial descriptions, quantity splits, unit discrepancies)
• Validation — zero-tolerance price recalculation on every line; any arithmetic inconsistency is flagged
• Human-in-the-loop dashboard — exceptions surface for review with full context; clean invoices post automatically

LangFuse handles observability across the full pipeline.

The finance team went from processing invoices to reviewing exceptions. That's a different job — one that requires their judgment rather than their time.

The 6-week timeline and single-engineer delivery are as significant as the outcome. This wasn't a large infrastructure project.