Timeline
10 weeks
Team
4 engineers + 1 PM
Client
Mid-sized manufacturer
The Problem
Finance teams spent hours manually processing invoices: extracting line items, validating against ERP data, and reconciling totals. The process was error-prone, slow, and consistently delayed month-end close cycles.
Our Approach
Process Mapping (Weeks 1 to 2)
Shadowed the finance team through a full month-end close to map every manual step. Catalogued the 14 invoice formats in circulation and the ERP fields each needed to land in.
Document Intelligence Pipeline (Weeks 3 to 5)
Built ingestion and extraction on Azure Document Intelligence with custom-trained models for the client's vendor invoice formats. Added a verification layer using Azure OpenAI to catch edge cases the layout models missed.
ERP Integration & Reasoning (Weeks 6 to 8)
Integrated bidirectional sync with the client's ERP. Added AI reasoning for 3-way PO matching, tax calculation validation, and currency conversion checks before any record was written back.
Chat Interface & Rollout (Weeks 9 to 10)
Shipped a conversational query layer so finance users could ask 'Show me all unreconciled invoices over $50K from Q3' in plain English. Rolled out to the AP team with a 2-week supervised period.
The Solution
We built an AI-powered document processing platform using Azure Document Intelligence and Azure OpenAI to extract, validate, and reason over financial data. A natural-language chat interface lets finance teams query invoices conversationally, and automated ERP cross-checking eliminated manual reconciliation.
Why This Tech Stack
Azure Document Intelligence was the fastest path to production-grade extraction without us training OCR models from scratch. Azure OpenAI's function calling made ERP tool-use deterministic and auditable, which is critical for finance workflows where every write-back is traceable.
The Outcome
Reduced invoice processing time from hours to seconds. Month-end close cycles accelerated significantly, and the finance team recovered capacity previously lost to manual data entry, which now goes to higher-value analysis.
Key Metrics
4 hrs → 12 sec
Avg. invoice processing time
98.5%
Extraction accuracy
6 days → 2 days
Month-end close acceleration
~60%
Finance team capacity recovered