Challenge: Customer order requests arrived as PDF attachments by email. CSR teams manually reviewed each document, extracted the required details, and re-entered the information into Deposco (the warehouse management system).
This process was time-consuming, repetitive, and entirely dependent on manual effort.
The operational consequences were predictable:
- Data entry errors: Manual transcription from PDF to WMS introduced mistakes in quantities, SKUs, addresses, and delivery dates
- Processing delays: Each order required a CSR to open the email, read the PDF, identify the relevant fields, and type them into Deposco. During peak periods, orders queued up
- No scalability: Order volume growth meant hiring more CSR staff or accepting longer processing times
- Inconsistent extraction: Different customers sent orders in different PDF formats, making it harder to standardize the manual process
In a 3PL operation where speed and accuracy directly affect customer satisfaction and warehouse throughput, a manual order entry bottleneck creates downstream problems across the entire fulfillment chain.
Solution
We built a proof of concept to automate the full flow from email inbox to WMS order creation.
The solution consists of two connected processes.
Step 1: PDF Ingestion from Email
The first script monitors the orders inbox, identifies emails containing customer order documents, extracts the PDF attachments, and stores the raw content in a Google Cloud Storage bucket.
This step handles:
- Email filtering to isolate order-related messages from general inbox traffic
- PDF extraction and storage in a structured cloud location
- Metadata logging (sender, timestamp, attachment filename) for traceability
By landing raw PDFs in cloud storage, the ingestion step is decoupled from the extraction logic. This means the AI extraction process can be updated, retrained, or reconfigured without touching the email integration.
Step 2: AI-Powered Data Extraction and Order Creation
A second process is triggered when new PDFs land in the storage bucket. This process uses an AI model configured with customer-specific extraction rules to:
- Identify the correct data fields from each PDF (item codes, quantities, shipping addresses, requested delivery dates, special instructions)
- Map extracted values into the Deposco API JSON structure
- Write the structured output into Snowflake for audit and reporting
- Create a new customer order in Deposco via API for further operational review
The customer-specific extraction rules are critical. Different customers send orders in different formats: some use structured purchase order templates, others send freeform email bodies with attached spreadsheets converted to PDF, and others use legacy ERP-generated documents with inconsistent layouts.
Rather than building a single rigid parser, the AI model adapts its extraction logic per customer, recognizing field positions, labels, and formatting patterns specific to each sender.
Human Review Layer
Orders created through the automated pipeline land in Deposco with a review flag. CSR staff verify the extracted data before the order moves into fulfillment. This keeps a human in the loop during the proof of concept phase while still eliminating the manual data entry step.
The shift is significant: instead of reading PDFs and typing data, CSR teams review pre-populated orders and correct exceptions. This is faster, less error-prone, and more scalable.
Architecture
The end-to-end flow:
- Email inbox receives customer order PDFs
- Python script extracts attachments and stores them in Google Cloud Storage
- AI extraction pipeline reads PDFs, applies customer-specific rules, and structures the data
- Snowflake stores the extracted order data for audit, reporting, and exception tracking
- Deposco API receives the structured order and creates a new entry in the WMS
Each step is logged and traceable. If an extraction fails or produces low-confidence results, the system flags it for manual review rather than silently creating an incorrect order.
Next Steps
The proof of concept validated the core flow. The next phase focuses on:
- Expanding customer-specific extraction rules to cover more document formats and edge cases
- Improving accuracy across document variations (handwritten notes, multi-page orders, orders with non-standard SKU formats)
- Strengthening monitoring and exception handling so operations teams have clear visibility into extraction confidence scores, failure rates, and processing times
- Scaling to additional customers and higher order volumes
- Tighter integration with operational review workflows to reduce the time between order receipt and fulfillment start
Business Impact
Even at the proof of concept stage, the results point to significant operational improvement:
| Metric | Manual process | Automated process |
|---|---|---|
| Order entry method | CSR reads PDF, types into Deposco | AI extracts, CSR reviews pre-populated order |
| Processing bottleneck | Scales with headcount | Scales with compute |
| Error source | Manual transcription | Extraction model (correctable, improvable) |
| Audit trail | None (manual entry) | Full (every extraction logged in Snowflake) |
In the long term, this solution can significantly reduce manual order entry effort, improve processing speed, and increase data consistency. It creates a scalable foundation for handling growing order volumes while allowing teams to focus on exception management and higher-value operational activities instead of repetitive data entry.
Technologies and Tools
Python, Google Cloud Storage, OpenAI API, Snowflake, Deposco API
