AP automation auto-posts 72% of invoices to QuickBooks

Junior bookkeepers were spending ~40% of their time on AP intake — manually re-keying vendor name, amount, GL code, and due date into QuickBooks Online. Published error rate was ~2% (consistent with APQC's manual-AP benchmark).

What we built

• Email-in mailbox per client → S3 drop → Bedrock Data Automation classification (invoice / statement / credit memo / utility bill / not-an-invoice).
• Custom blueprints per common vendor (top 50 vendors cover ~80% of volume); generic invoice blueprint as fallback.
• Vendor master deduplication agent that catches “Acme Inc / Acme Corp / Acme Corporation” as the same entity using fuzzy match + LLM judgment with citation.
• Confidence-routed HITL: fields with confidence ≥0.92 auto-post to QBO; below routes to Amazon A2I with the original PDF and the agent’s extracted fields side-by-side.
• Three-way match agent on clients where POs exist; flags discrepancies for human review rather than auto-resolving.

Caveats

Free-text line items on construction invoices break three-way matching. Vendor master deduplication is harder than it sounds — we still review the dedup agent’s calls manually for the first three months on every new client. Handwritten utility bill amounts hit ~85% accuracy, never 99%; the confidence routing handles that gracefully but the per-document HITL rate on utilities stays higher than on system-generated invoices.

Stack. AWS Bedrock Data Automation · Amazon Textract for handwritten line items · Amazon A2I for HITL · QBO API · S3 · weekly eval review in Langfuse.

Auto-post rate ~72% across the client base. Junior bookkeeper time on AP intake fell from 40% to 15% of the week. Verified error rate fell to 0.7%. Invoice processing dropped from 18 hours to 35 minutes.