Automated Comparison of Order Confirmations and Purchase Orders

Manufacturing Company

Problem

In manufacturing, customers often request order confirmations to ensure availability, correct data transmission, and delivery dates. These must be laboriously compared line by line with the associated purchase order, which is a time-intensive and error-prone manual process. With just-in-time/lean manufacturing, missing a quantity discrepancy can lead to production line standstill in the worst case, pricing discrepancies can lead to liquidity management problems. Catching errors early is paramount.

Solution

Using modern AI technology, we implemented an automation solution using the UiPath platform and OpenAI GPT AI models to automatically ingest order confirmations and compare them to the purchase order.

Process Steps

  1. 1
    Retrieve order confirmations from email attachments or a hot folder
  2. 2
    Digitize them using UiPath OCR
  3. 3
    Look up the purchase order number from the document using regular expressions and retrieve PO data from the ERP system
  4. 4
    Add a well-engineered prompt alongside the document and PO data to the context to allow the LLM to intelligently compare data
  5. 5
    In case of discrepancies send the discrepancy and extraction data, as well as the document, to a custom app in UiPath Action Center for human-in-the-loop review
  6. 6
    In case of no discrepancies or after human review, create the order confirmation in the ERP system ProAlpha

Advantages

  • Saves an average of about 5 minutes of manual handling time per purchase order, saving thousands of person-hours per year in total (estimated savings of about €150,000 per year for implementation costs <€50,000)
  • Reduced error rate by only displaying possible discrepancies to human operators
  • Low risk of AI hallucinations or malfunction due to structured input and output as well as just comparing data available in context

Challenges

  • Complex end-to-end solution
  • Robustness (ensure order confirmations aren't lost, AI sometimes doesn't return well-formatted data)
  • Many edge cases: mismatched descriptions, missing item codes, single OC for multiple POs, multiple units of measure, header data in items table, long item lists (>100 items)

Possible Extensions

  • Direct communication with suppliers in case of discrepancies using AI-generated messages
  • Tracking missing order confirmations by hooking into the purchase-to-pay process
  • Four-way match with invoice and delivery note for complete documentation
  • Same general approach for three-way match without order confirmation
  • Integration with additional ERP systems and e-procurement platforms
  • Optionally use UiPath Document Understanding for invoice data extraction instead of an LLM