The Challenge: POs That Don't Fit in Rows
Orders arrive as free text
Emails, scanned faxes, and PDFs from hundreds of suppliers (pole hardware, transformers, PPE), each phrased differently.
Manual keying is slow
Staff re type PO numbers, line items, and totals into the ERP. Minutes per document, thousands of documents per month.
Errors surface late
A mistyped quantity or price isn't caught until invoice matching fails, or a storm response crew is short on materials.
Our Demo Document: A Real World PO
…purchase order number PO-2026-04417, dated August 25, 2026. Bill to Midwest Supply Co., Carmel, IN. Ship to our Westfield warehouse. - 40 cases 16oz clear cups, item #CUP-16CL, at $28.50 per case - 12 boxes nitrile gloves (large), SKU GLV-NTR-L, $9.75 each - 5 rolls stretch wrap, item WRAP-80, unit price $22.00 Payment terms are Net 30. Requested delivery by Sept 5, 2026. Total should come to $1,367.00 before tax.
What makes this hard for software?
The Pipeline at a Glance
Amazon S3
Raw PO lands in the bucket; event fires
AWS Lambda
Orchestrates extraction on each upload
Bedrock / Textract
Free text → strict JSON fields
Validate
Types coerced; totals cross checked
Curated S3
Parquet tables: header + line items
Athena SQL
Query instantly via Glue Catalog
Optional Destinations
Demo Part 1: Land & Extract
Drop the document in S3
Raw email text lands in s3://po-pipeline/raw/. An S3 event notification triggers Lambda automatically. No schedules, no polling.
Scanned PDFs & images
Amazon Textract AnalyzeExpense: purpose built for invoices and POs, returns fields with confidence scores.
Free form text (our demo)
Amazon Bedrock with Claude: a prompt asks for strict JSON, handling any phrasing a supplier invents.
Demo Part 2: Validate & Structure
Trust, but verify
Two clean Parquet tables
| po_number | vendor | total |
|---|---|---|
| PO-2026-04417 | Midwest Supply | 1367.00 |
| po_number | sku | qty | price |
|---|---|---|---|
| PO-2026-04417 | CUP-16CL | 40 | 28.50 |
| PO-2026-04417 | GLV-NTR-L | 12 | 9.75 |
| PO-2026-04417 | WRAP-80 | 5 | 22.00 |
Demo Part 3: Query in Seconds
SELECT h.po_number, h.vendor_name, l.sku, l.quantity, l.unit_price FROM po_header h JOIN po_line_items l ON h.po_number = l.po_number WHERE h.po_date >= DATE '2026-08-01';
Where Utilities Put This to Work
Storm restoration
Emergency material orders placed by phone and email under pressure become trackable data the same day, supporting FEMA cost recovery.
Grid materials
Transformer, conductor, and meter POs from dozens of vendors normalize into one spend view for sourcing negotiations.
Contractor invoices
The same extract validate pattern three way matches contractor invoices against POs and receipts automatically.
Audit & compliance
Every extracted field links back to the source document in S3, giving a clean trail for rate case and regulatory audits.
What This Is Worth
- Spend analytics across all suppliers, including the long tail that never used EDI.
- Price variance alerts when a vendor's unit price drifts from contract.
- Delivery performance history to inform storm season stocking decisions.
From Demo to Production
Pilot (2 to 4 weeks)
Your documents, your environment: 100 real POs through the pipeline, accuracy measured against your team's own keying.
Integrate
Connect curated output to your ERP (SAP, Oracle, Maximo) and your BI layer; wire quarantine review into existing workflows.
Expand
Extend the pattern to invoices, receipts, and contractor timesheets. Same architecture, new prompts.
