Enhancing AI Purchase Order Processing Software with Data Management

Customer Overview

Our client is a US-based distributor, partnering with a leading global manufacturer of fluid system components. They provide a wide range of high-quality products, including fittings, valves, regulators, gauges, hoses, and tubing, to meet the needs of various industries. The distributor works with numerous customers, offering quotations in high volumes, and ensuring the timely delivery of products against different types of Purchase Orders (POs), followed by the timely sharing of invoices to receive payments on time.

Project Overview

To process lengthy Inquiries and POs in high volumes, we developed a PO automation software using an agentic framework and multiple AI agents in phase 1 of this project. For phase 2, the client sought to increase the AI system’s capabilities to process different types and formats of POs. They wanted to provide sales insights, such as identifying customers with declining sales. They aimed to monitor the AI system’s performance, introduce a fallback mechanism, and give top management clear visibility into how the AI system was performing. Finally, they intended to automate invoice generation and sharing to avoid delays in receiving payments.

Challenges

The distributor faced challenges with limited data insights, PO processing limitations in the AI system, a lack of AI performance monitoring, and manual invoicing delays.

• Data was stored in silos across systems and databases, making it difficult to generate custom data insights for top management, sales, and CSR teams.
• The PO Automation Software lacked capabilities to identify, categorize, and prioritize PO types like revised POs, blanket POs, Government POs, and back-dated POs, and could not process POs in formats like Excel and Emails.
• The client also wanted to accelerate the PO processing speed of the AI-based system.
• The sales team wanted to spot quotations that did not convert to POs, identify customers whose POs or sales were declining, and find out which fluid systems the customers intended to build to pitch their components.
• There was no mechanism to monitor the performance of the AI-powered PO automation software. When the AI failed to process a PO, there was no fallback mechanism for manual CSR intervention.
• The top management wanted complete visibility into the AI system’s performance, as well as the performance of the CSR and sales teams.
• The manual process of generating invoices and emailing them to clients caused delays in payment collection, directly affecting profitability.

Solution

Building a data management system to help generate custom data insights, enable processing of POs of different types and formats, facilitate AI-based Purchase Order Processing Software’s monitoring, and automate invoice generation and sharing.

• In phase 1, we built a data support system for the multi-agent AI PO automation software. Turned it into a complete data management system in Phase 2 to centralize data into a data lakehouse and enable analytics and visualization.
• Synchronized SAP customer data with a read-only database for AI to run queries and detect DAPS codes (Government POs), contract codes (Blanket POs), repeat entries (Duplicate POs), and date markers (Back-dated POs). A rule-based process then classified and prioritized POs. Converted Excel and email POs into PDF and used separate logic to process each PO type.
• Added functionality to convert quotations directly into POs within the AI-based Purchase Order Processing Software, reducing processing time and saving storage space.
• Streamlined customer data and applied vector embeddings to compare processed POs with quotations and report which quotations didn’t result in sales.
• Combined database PO categorization data with Zendesk’s customer info to analyze each customer’s PO volumes and values over time and identify customers with declining sales.
• Analyzed customer interactions (emails, messages, call transcripts) to extract references to fluid systems and help the sales team pitch relevant components.
• Implemented AI-based Purchase Order Processing Software’s monitoring for SLA-based processing times, number of POs processed/rejected, issue types, number of AI issues detected/resolved, etc. Enabled notifications for the development team on AI failures and the CSR team to process POs when automation failed.
• Build dashboards for top management to view the AI system’s performance using Grafana (visualization) and our data management system (synchronization), tracking processing time, POs handled, and AI failure rates. Integrated CSR and sales KPIs at individual and team levels for complete performance visibility.
• Automated invoice generation and sharing workflow by syncing delivery messages from partners (FedEx, UPS) as events in Zendesk, and making them trigger invoice generation in SAP, which in turn triggers Zendesk to email invoices to customers and notify the CSR team.

Benefits

Our efforts in enhancing the AI Purchase Order Processing Software using a data management system and adding advanced functionalities offered the following benefits:

• Centralized data and real-time dashboards cut report generation time from hours to minutes, reducing manual effort by ~60%.
• Automated PO processing across varied formats and types minimized manual intervention and reduced errors by ~30–35%.
• AI monitoring and fallback alerts minimized downtime, allowing the CSR team to take over and keep PO processing continuous.
• Automated invoicing cuts cycles by 2–3 days, helping the distributor collect payments faster and reduce payment delays.

Technology

• Apache NiFi
• AWS Titan Embeddings
• SAP
• AWS Bedrock
• Amazon SQS
• Amazon API Gateway

Industry

• Manufacturing