AWS for Industries

Optimize Your Supply Chain Processes with Generative AI on AWS

According to a recent study by Gartner, there are three key supply chain macroeconomic trends that influence how the top 25 supply chain leaders prioritize their investments. These are supply chain organization design and talent, AI-driven advances to realize productivity and process optimization, and elevating supply chains to anti-fragile state, all of which help achieve supply chain excellence.

Businesses across various industries have long struggled with supply chain issues. These challenges include scattered data across fragmented sources, lack of comprehensive visibility into the entire supply chain, and inefficient supply chain processes. Procurement teams face many challenges when placing purchase orders (POs), such as inaccurate data from manual data entry, supplier challenges from capacity constraints or production delays, cost overruns, and lack of transparency. In this blog post, we will discuss how to leverage AWS generative AI capabilities to streamline supply chain procurement processes to improve operational efficiency, enhance transparency, compliance and risk mitigation, and reduce overall procurement costs.

AI-Powered procurement process optimization

AWS Artificial intelligence (AI) and Machine Learning (ML) capabilities help streamline procurement processes end-to-end, including automating PO generation and approval, and facilitating the invoice-to-payment workflow.

AWS generative AI can make generating and approval of POs more intelligent. It can automatically draft a PO based on supplier catalogs by generating contextual details such as item descriptions, quantities, delivery timelines, and terms. The orders are then automatically routed to an approval workflow. Invoice processing automation using ML capabilities streamlines extraction of invoice data such as line items, quantities, and amounts from scanned or emailed invoices. This data is fed into a data integration pipeline to automatically match invoices to POs and route them for automated review and approval.

The optimized procurement process reduces manual PO creation and matching checks, reduces errors, and ensures orders are placed based on up-to-date data and business constraints. This significantly reduces manual data entry, accelerates invoice-to-pay cycles, and ensures compliance with procurement controls, creating a seamless and efficient system.

Solution overview

The proposed solution integrates several AWS services to automate the procurement process from start to finish. Figure 1 illustrates the solution architecture.

Figure 1 – Technical solution architecture

Agents for Amazon Bedrock orchestrate interactions between foundation models (FMs), data sources, software applications, and user conversations via action groups that invoke AWS Lambda functions. The main interaction with an Amazon Bedrock Agent is via chat. The conversation is powered by a large language model (LLM) that you can select from the list of supported models. Using a reasoning technique called ReAct (Reasoning+Action), the model can break down complex tasks into multiple steps. The model then decides on the next action by going through each of the steps and observing the results. It uses previous step outputs for inputs to subsequent steps.

An action group defines actions that the agent can help the user perform. The solution has three action groups: purchase-order-creation, purchase-order-sending-to-approval, and invoice-sending-to-payment. Each action group can contain one or more actions, with functions written in Python code. To fulfill the user request, the Amazon Bedrock Agent will execute these actions by calling them as APIs from the Lambda function.

Figure 2 – Amazon Bedrock Agents action groups

To define an action group, you need to specify the parameters that the agent must obtain from the user for each action in the action group to be carried out. The corresponding Lambda function should contain the logic for each action: how the input parameters are handled and what the output is. Each action is independent; it can call various AWS services and perform code operations. For example, when a new purchase need arises, the agent will perform a draft PO generation by:

  • Retrieving templates stored in an Amazon S3 bucket.
  • Extracting relevant product details, quantities, and vendor information from supplier data maintained in Amazon DynamoDB.
  • Generating a PO number according to the company’s standards.
  • Calculating subtotals and totals including taxes and discounts.
  • Formatting the PO with the appropriate payment terms and shipping instructions.

The generated PO is then routed through an approval process using Amazon Simple Email Service. The Amazon Bedrock Agent can also use additional supplier metrics such as quality rating, on-time delivery percentage and defect rate in PO generation. Once the order is shipped, a vendor sends an invoice. Before it can be sent for payment, the checking and matching process can be done by the agent. The agent will perform the following actions:

  • Parsing documents using the computer vision service Amazon Textract to automatically extract invoice data like line items, quantities, and amounts.
  • Matching the invoice data to the original PO, along with goods receipt, flagging any discrepancies that require manual review.
  • Calling text-generation foundational models from Amazon Bedrock to explain programmatic matching results and enrich the email content.
  • If there is a match, attaching the invoice to a generated email and sending it for payment.

Typically, a company possesses a large number of documents with business rules and regulations. The Amazon Bedrock Agent can be augmented with this information by querying the Knowledge Bases for Amazon Bedrock. This technique is called the Retrieval Augment Generation (RAG). To effectively feed the documents to a knowledge base, they can be stored in an S3 bucket. The vectorized version is then created as an Amazon OpenSearch Serverless vector search collection. Once a knowledge base is created and the documents are ingested, the knowledge base searches among them to find the most useful information and uses it to answer natural language questions.

By combining the natural language understanding, reasoning, and generation capabilities of Amazon Bedrock Agents with other AWS AI/ML services, organizations can drive significant efficiencies and cost savings across their supply chain procurement operations. The agents simplify the job for procurement teams, automating repetitive tasks and surfacing insights to make more strategic decisions.

Optimized procurement workflow

The following sections give an example of a customer working through a streamlined procurement workflow, including PO creation, PO approval, and invoicing.

Purchase order creation

Example Corp wants to create a PO to purchase products from their supplier, AnyCompany. Previously, if the procurement specialist wanted to create a PO, they had to go through templates or drafts stored locally or in a shared directory. The procurement specialist had to manually fill in information such as supplier details, date, PO number, items to procure, taxes and discount information, subtotal and total amounts, and any additional notes. Company-specific guidelines had to be found and inspected. The required details had to be changed and input manually, which could cause mistakes and delays. Now, the procurement specialist can ask the agent to create a PO. Interaction with the agent happens via chat in natural language. The specialist needs to provide a template name, vendor name, and country name. The template includes product and supplier information. Some of these parameters can be optional, depending on the company’s procurement process. The remaining part of the workflow will be handled automatically by the agent.

Figure 3 – Agent interaction to generate Purchase Order

Purchase order sending for approval

According to Example Corp’s workflow, once a PO is created, it must be sent to the Global Supply Chain department for approval. Previously, the procurement specialist needed to manually create and send an email for approval including the PO and supplier metrics like quality rating and defect rates. Now, the agent can compose an email to send for review to the global supply chain department with the PO draft and supplier metrics automatically included.

Figure 4 – Agent interaction to send Purchase Order for approval

Invoice sending for payment

Once the goods are delivered, Example Corp receives an invoice from AnyCompany to pay. Before payment, Example Corp needs to confirm the received items and the quantities match the order placed.

Previously, the supply chain department had to manually verify that the PO, invoice, and receipt of goods matched. This time-consuming process had to be completed before the invoice could be sent to the accounting team for payment. Now, the procurement specialist can ask the agent to send invoice for payment, including the PO number and the name of the department they need to send the invoice to in the request. The agent handles match checks automatically. By streamlining the review process and sending the invoice automatically to the appropriate team, the agent accelerates the payment process.

Figure 5 – Agent interaction to send invoice for payment

Key benefits for your business

Optimizing supply chain procurement processes using AWS generative AI capabilities can deliver several key outcomes for organizations:

Improved Operational Efficiency: Automating repetitive tasks like PO creation and invoice processing eliminates manual effort and human errors, and improves productivity. This enables procurement teams to focus on more strategic activities like supplier relationship management and process optimization.

Reduced Procurement Costs: Streamlined invoice processing and approvals accelerate the invoice-to-payment cycle, improving working capital management. Additionally, process automations free up procurement specialist time to conduct supplier negotiations which lead to better pricing, payment terms, and service levels.

Enhanced Compliance and Risk Mitigation: Automating the procurement process helps enforce procurement policies, contract terms, and regulatory requirements.

By leveraging AWS generative AI capabilities through Amazon Bedrock and other complementary services, organizations can drive tangible improvements in efficiency, visibility, cost savings, and risk management across their supply chain procurement operations.

Complementary solutions to solve additional supply chain challenges

The solution we illustrated can be enhanced with additional capabilities to improve supplier communication and negotiation. Amazon Lex chatbots can be used to facilitate natural conversations with suppliers regarding order changes, expedited shipping, and pricing negotiations. Amazon Transcribe can be used to analyze call recordings with suppliers and identify areas for process improvement or cost savings opportunities. You can leverage Amazon Comprehend to sentiment-analyze supplier emails and identify potential issues or opportunities proactively.

These enhancements will help improve supplier relationships. Agents can now engage in more consistent, responsive, and personalized communications with suppliers. Analyzing interaction patterns helps identify top-performing suppliers and opportunities to strengthen partnerships. Automated scorecarding and performance feedback allows for more constructive supplier management.

Furthermore, you can perform predictive maintenance for your supply chain assets using AWS capabilities. Collect sensor data from supply chain equipment and infrastructure using AWS IoT Core. Use Amazon SageMaker to build machine learning models that can predict when assets like forklifts, conveyor belts, or storage tanks may require maintenance. Integrate these predictive insights with Amazon EventBridge to automatically trigger work orders and parts ordering.

Conclusion

The supply chain procurement optimization solution, powered by AWS generative AI, demonstrates the transformative impact of intelligent automation on procurement processes. By utilizing Amazon Bedrock agents and seamlessly integrating AWS services like Amazon Lambda, Amazon S3, Amazon Textract, and Amazon DynamoDB, organizations can improve efficiency, transparency, and cost savings.

The opportunities for further enhancements, such as integrated supplier negotiations, predictive maintenance, and advanced analytics, are the next steps in transforming end-to-end supply chain processes with generative AI. We encourage you to explore how this solution can optimize your organization’s procurement workflows. Contact our team of AWS experts today to schedule a consultation and learn more about harnessing the power of AI and ML to drive supply chain excellence.

Soonam Kurian

Soonam Kurian

Soonam Kurian is a Principal Solutions Architect at Amazon Web Services specializing on partner solutions. She is focused on data analytics, artificial intelligence, and machine learning. In her current role, she works with Global System Integrators and Independent Software Vendors providing architectural guidance to launch strategic industry solutions on AWS. In her spare time, she enjoys reading and being outdoors.

Darya Petrashka

Darya Petrashka

Darya Petrashka is a Data Scientist at SLB with 5 years of experience, focusing on supply chain projects in data analysis, NLP, and generative AI. She is passionate about using data for problem-solving, with a strong interest in classical machine learning, NLP, and AWS services. An AWS Community Builder and Authorized Instructor, Darya actively shares her expertise through public speaking at various industry events, including AWS Community Days, AWS Cloud Day, and PyCon. A dedicated learner, Darya continually hones her skills by participating in workshops, courses, and tech schools.