Case study · 2024 · Finance · Lending
Document Automation for a Commercial Finance Team
A commercial finance operations team processed 8,500 invoices, statements, and funding packets a month by hand: read, re-key, post, chase exceptions. We automated the pipeline end to end — AI extraction tuned on their own document mix, validation rules that encode their controls, and RPA that posts clean documents into the ERP while humans work a properly triaged exception queue.
Screens are illustrative recreations — client interfaces stay confidential.
The challenge
Documents arrived in every format a counterparty could invent — scans, photos of printouts, spreadsheets exported to PDF — and each one crossed at least three pairs of hands between inbox and ERP posting.
The team's auditors were the second stakeholder: any automation had to preserve maker-checker controls and leave an evidence trail at least as good as the manual process it replaced.
What we built
Extraction tuned to their paper
Azure Document Intelligence custom models trained on the team's own document mix handle layout variety; low-confidence fields route to a human snippet-review instead of failing the whole document.
Controls as validation rules
The team's posting controls — PO matching, tolerance thresholds, duplicate detection, approval routing by amount — became explicit rules evaluated before anything touches the ERP, versioned and reviewable by audit.
RPA with a paper trail
Serverless orchestration drives posting into the ERP with every step logged — source image, extracted values, rule results, approver — one click from any ledger entry back to its evidence.
An exception queue worth working
Documents that fail rules land in a queue sorted by aging and amount with the failure reason attached — the team works judgment calls, not data entry, and queue patterns feed monthly rule tuning.
The results
- 82%
- Straight-through processing
- 1.1 days
- Median cycle time
- −92%
- Posting errors
posted with no manual touch, month 6
receipt to ERP posting, from 6.5 days
corrections per 1,000 postings
June reflects the pilot document types only; all document classes onboarded by September.
View the data as a table
| Straight-through rate | |
|---|---|
| Jun | 34% |
| Jul | 51% |
| Aug | 62% |
| Sep | 71% |
| Oct | 78% |
| Nov | 82% |
- 82% of document volume now posts straight through with no human touch, up from zero.
- Median cycle time from receipt to ERP posting fell from 6.5 days to 1.1.
- Posting errors dropped 92%, and the audit review of the automated pipeline closed with no control findings.
Client identities stay confidential; figures are rounded from end-of-engagement delivery reporting.
Stack & expertise
- Python
- Azure Document Intelligence
- Azure Functions
- SQL
- ERP integration
Demand Forecasting for a Grocery Co-operative
Store-level demand models and a planner review queue that halved stockout hours and cut fresh-produce shrink 14% across 18 stores.
Grounded Research Assistant for a Law Practice
A permission-aware retrieval assistant over 8,000 matter documents — every answer cited, every source access-controlled — that cut research time by two-thirds.
Vision Inspection on a High-Speed Packaging Line
Edge-deployed defect detection at 96 units per minute, with an operator feedback loop that pushed escaped defects down 84% in six months.