AWS for Industries

Revamping procurement operations with generative AI–powered Offer Analyst solution on AWS

This blog post explores the transformative impact that advanced generative AI can have on procurement operations. Through an engagement with Boston Consulting Group (BCG) and Amazon Web Services (AWS) Professional Services BMW Group has successfully implemented and released its internal tool called Offer Analyst, a groundbreaking generative AI solution that empowers procurement specialists (e.g. purchasers) at BMW Group in their daily work. In the following sections, we will delve into how this innovative transformation is redefining efficiency and accuracy, setting new standards across industries.

Breaking down the traditional process

To understand the magnitude of this transformation, it’s essential to first grasp the traditional offers-processing workflow at BMW Group. This process can be broken down into three main steps:

  1. Documentation collection: BMW Group uses a specialized system for tender publication and supplier interaction when purchasing indirect goods and services. The offers are downloaded by the purchaser, usually in a Word document or PDF. Other supporting documents, such as pricing sheets in Excel, are also considered. All this documentation is then shared with the reviewers.
  2. Review and preselection: The purchaser and the department in charge of the project each review all the offers from different perspectives. The purchaser screens each offer for discrepancies to the tender and/or BMW Group’s terms and conditions. This step involves identifying discrepancies or risk spots. Meanwhile, the department responsible for the project assesses service coverage, quality, and initial price comparison to determine technical approval. This includes an in-depth comparison between the scope of the request for proposal (RfP) and the scope of the offers. As a result, a set of offers is preselected to continue the purchasing process.
  3. Offer Selection: The purchaser compiles feedback on the preselected offers and reviews analysis reports. This information is then used to prepare for discussions with the suppliers.

Despite the meticulous nature of this process, some key pain points have become increasingly significant in today’s technology-centric environment.

First, there are challenges with process efficiency. The traditional approach involves tedious proofreading and the comparison of multiple documents. This requires attention to detail and high interpersonal coordination efforts with multiple stakeholders, causing the process to become cumbersome and lengthy.

Second, the quality of outcomes could suffer. The mentioned intense manual effort in the review process increases the risk for errors, which in turn negatively impact the quality and accuracy of the subsequent procurement processes.

Lastly, the human element is increasingly important to consider. Streamlining processes can significantly enhance employee satisfaction by reducing manual and repetitive tasks. This allows employees to focus on more challenging and rewarding aspects of their roles, helping boost their morale and productivity.

Again, these challenges are not unique to BMW Group but are prevalent across procurement departments worldwide.

New player has joined the game: Generative AI

With its advanced text analysis features, generative AI can effectively process large volumes of information, perform precise text comparisons, and automate tedious tasks. This groundbreaking technology is set to transform operational workflows.

The offer analyst

To capitalize on the transformative potential of generative AI, BMW Group engaged with BCG and AWS Professional Services to create a revolutionary solution named Offer Analyst. This innovative tool uses advanced generative AI capabilities, augmented with procurement expertise, to streamline and enhance the offer evaluation process. Offer Analyst features an accessible, user-friendly interface that simplifies analysis, making it easy for users to navigate and use. The close collaboration with BMW Group’s procurement experts verified that the tool was custom built to address unique challenges faced by the procurement team, setting a new standard in procurement excellence.

Offer Analyst quickly revolutionized the traditional offer-review procedure. Here’s a detailed look at the updated review process:

Figure 1. Offer Analyst user flow diagram (by Boston Consulting Group)

Figure 1. Offer Analyst user flow diagram (by Boston Consulting Group)

  1. RfP document upload: The journey begins with the upload of the RfP document in PDF format. This document outlines the project’s requirements and expectations.
  2. Offers upload: Users upload the submitted offers and price sheets, centralizing all relevant documents for analysis.
  3. Offer information extraction: In this phase, the large language model (LLM) extracts bidder information from the previously uploaded documents.
  4. Initial analysis—standard criteria: The tool automatically compares the offers against the uploaded RfP and other relevant documents (e.g. terms and conditions) using standardized criteria derived from expert interviews. The analysis results are presented in an easy-to-digest dashboard, providing a comprehensive overview of how each offer aligns or misaligns with the requirements.
  5. Tailored analysis—ad hoc criteria: Recognizing the unique requirements of each project, the tool lets users adjust existing criteria or create new requirements. Using generative AI’s natural language processing capabilities, users can provide descriptions or prompts to guide the tool, asking it to analyze specific elements of the offer. This flexibility helps verify that the evaluation is tailored to the specific needs of each tender, enhancing relevance and precision.
  6. Download analysis: For further offline processing, users can easily download the analysis results in an Excel file.
  7. Interactive analysis—chat with your offer: To gain deeper insights, this feature lets users ask prompt-based questions that pertain to various aspects within the offer, moving beyond criteria-based analysis. This interactive approach provides a nuanced understanding of each offer, letting users extract insights and clarify doubts.

 

Technical solution overview

Figure 3. Offer Analyst solution architecture on AWSFigure 2. Offer Analyst solution architecture on AWS

The reference architecture above (figure 2), designed and implemented in a joint effort between BMW Group, BCG, and AWS Professional Services, demonstrates the key solution components, including the following:

  1. Frontend/user interface (UI): Users initiate the process by uploading documents in an application UI. These documents are handled by an Application Load Balancer, which routes traffic to targets based on the content of the request. Authentication is managed by the BMW Group Identity Provider, verifying secure access to the application.
  2. Document storage: Uploaded documents are stored in Amazon Simple Storage Service (Amazon S3), an object storage service that offers scalability, data availability, security, and performance. A rule is configured in Amazon EventBridge—a serverless event bus—to monitor the Amazon S3 bucket for PutObject events, which are triggered whenever a new document is uploaded.
  3. Integration layer: Upon detection of a new document, the Amazon EventBridge rule triggers a document ingestion function from AWS Lambda, a compute service that runs your code in response to events and automatically manages the compute resources. This AWS Lambda function retrieves the document from Amazon S3 and processes its contents by extracting the text.
  4. Generative AI layer: The generative AI services layer is responsible for parsing files and implementing the retrieval-augmented generation (RAG) mechanism using Anthropic Claude 3 Sonnet through Amazon Bedrock, a fully managed service that offers a choice of high-performing foundation models. The extracted text undergoes embedding using the Amazon Bedrock embedding models. The processed text and its corresponding embeddings are stored in Amazon Aurora Serverless PostgreSQL—an on-demand, autoscaling configuration for Amazon Aurora—equipped with the PGVector extension.
  5. API layer: The system also features a REST API endpoint managed by Amazon API Gateway, a fully managed service that makes it easy for developers to create, publish, maintain, monitor, and secure APIs at any scale. This lets the user interact with the uploaded documents. Amazon API Gateway routes requests to additional Document Chat and Criteria Analysis AWS Lambda functions, which respectively handle the document chat interactions and the criteria analysis capabilities. These functions use the stored embeddings and text data to provide such capabilities.
  6. Security, logging, and tracing: To help maintain high security and compliance, a best practice is to also implemented using the following services: Amazon Route 53, Amazon Certificate Manager, Amazon Elastic Container Registry, AWS Key Management Service, Amazon CloudWatch, AWS CloudTrail, AWS Secrets Manager

The Offer Analyst architecture uses AWS serverless cloud services that provide the tools for building scalable, robust, and secure solutions. Those services can automatically scale to handle increased traffic without the need for manual intervention and can manage thousands of concurrent requests, scaling up or down based on workload demands.

By designing a serverless solution architecture in accordance with AWS Well-Architected principles, BMW Group accelerated the delivery of Offer Analyst, focusing on the use case implementation. Combining their unique strengths and expertise, the teams efficiently spearheaded the delivery, with BCG supporting BMW Group’s driving of the generative AI logic implementation, prompt engineering, UX/UI, and solution release, while AWS Professional Services focused on the infrastructure, security, and automation of the solution.

Now that we’ve covered the high-level architecture and its highlights, let’s dive deep into the technical details of three critical components of the Offer Analyst solution.

Document ingestion
AWS offers a range of services that can be used to implement a vector database for RAG. Here, we built upon Amazon Aurora with PostgreSQL, facilitating near real-time ingestion of documents into the vector store. This lets the RAG model access and retrieve relevant information for use cases where low-latency performance is crucial for a seamless user experience. As alternatives to Amazon Aurora with PostgreSQL, AWS provides other services for RAG, such as Knowledge Bases for Amazon Bedrock, which you can use to give foundation models and agents contextual information from your company’s private data sources for RAG, and Amazon OpenSearch Service, which makes it easy for you to perform interactive log analytics, near real-time application monitoring, website search, and more. Each of these services offers unique features and capabilities, letting developers choose the solution that best fits their specific requirements. For example, Knowledge Bases for Amazon Bedrock is a fully managed service designed for building and deploying knowledge bases, while Amazon OpenSearch Service provides distributed search and analytics engines that can be used to store and query large volumes of data.

The AWS Lambda function data-ingestion-processor is based on a Docker image stored in an Amazon ECR repository. The function uses the LangChain Amazon S3FileLoader to read the file as a LangChain Document. Then, the LangChain RecursiveTextSplitter chunks each document, given a CHUNK_SIZE and a CHUNK_OVERLAP, which depends on the maximum token size of the text embedding model. Next, the AWS Lambda function invokes the embedding model on Amazon Bedrock to embed the chunks into numerical vector representations. Lastly, these vectors are stored in the Amazon Aurora PostgreSQL database. To access the vector database, the AWS Lambda function first retrieves the Amazon Aurora–managed access credentials from AWS Secrets Manager.

Criteria analysis
To facilitate efficient side-by-side analysis of submitted RfP bids, BMW Group uses the large context window of Anthropic Claude 3 Sonnet through Amazon Bedrock (context window of up to 100,000 tokens). For an LLM call, BMW Group constructs a prompt, including a description of the criteria for the LLM to evaluate, the original RfP document, and the full text of the offer. To generate structured responses from LLMs, BMW Group builds upon the Instructor library when invoking Amazon Bedrock LLMs. For each offer, the LLM is instructed to evaluate the conformance of the offer against the criteria using traffic light logic, with the following possible classifications: red—no conformance, yellow—partial conformance, green—full conformance, and gray—unknown. Moreover, the solution provides a written explanation of the evaluation result and highlights the key passages in the original offer that support informed decision-making. The solution repeats the above procedure for each analysis criterion to produce a table with the full comparison of all offers across all specified criteria.

Document chat
To further enhance procurement specialists’ efficiency, Offer Analyst provides a chat-based interface for exploring and extracting specific information directly from the uploaded offer documents. This is a RAG use case where the solution retrieves relevant text chunks from an offer through a relevancy search in the vector database and then augments the LLM prompt that contains the question with the relevant text chunks for the LLM to produce a final answer to the user query.

Business outcomes: A transformative leap

The generative AI–supported Offer Analyst has helped transform BMW Group’s procurement processes, addressing key pain points in the traditional approach. Automating the review and comparison of offers helps reduce the time spent manually proofreading, letting employees focus on strategic tasks.

Automated compliance checks reduce the risk of errors, resulting in more reliable procurement outcomes. Detailed analyses also strengthens purchasers’ selection process, helping lead to more effective discussions.

Lastly, the new approach positively impacts the human factor—by alleviating hours of manual review, the solution boosts user satisfaction. Employees can now engage in more meaningful work, enhancing job satisfaction and motivation.

Looking ahead: The future of procurement at BMW Group

The successful implementation of Offer Analyst is just the beginning. BMW Group’s forward-thinking approach shows that they are well-positioned to use further AI and machine learning advancements. The broader vision for BMW Group’s procurement strategy includes continual improvement and adaptation, but by staying ahead of the curve, the company can help ensure that their procurement processes remain efficient, and effective.

Conclusion

The journey of transforming procurement operations at BMW Group, with the help of generative AI, is a testament to the power of technology and innovation. The generative AI–supported Offer Analyst is a shining example of how advanced technology can drive significant improvements in operational efficiency and accuracy, making the procurement more efficient and enjoyable for those involved.

To learn more about the power of generative AI and how to use it for building differentiated experiences, boosting productivity, and innovating faster, visit Generative AI on AWS.

Maik Leuthold

Maik Leuthold

Maik Leuthold is a Project Lead at the BMW Group for advanced analytics in the business field of supply chain and procurement, and leads the digitalization strategy for the semiconductor management.

Alvaro Dominguez

Alvaro Dominguez

Alvaro Dominguez is a lead AI engineer at BCG X and part of the leadership team for AI in automotive. He is an expert for building and industrializing innovative generative AI applications in the automotive sector.

Dennis Winter

Dennis Winter

Dennis Winter is a Data Scientist at the BMW Group, with a focus on analytics in supply chain and procurement. He develops cloud-native data analytics solutions.

Martin Maritsch

Martin Maritsch

Martin Maritsch is a Data Scientist at AWS ProServe focusing on Generative AI and MLOps. He helps enterprise customers to achieve business outcomes by unlocking the full potential of AI/ML services on the AWS cloud.

Nicola D'Orazio

Nicola D'Orazio

Nicola D'Orazio is a Senior Cloud Application Architect at AWS ProServe specialized in assisting automotive customers. He guides customers’ technical architecture and ease the adoption of new services and products when they become available.

Ole Eersink

Ole Eersink

Ole Eersink is an IT consultant at BCG Platinion. He specializes in industrial goods clients and is an expert on the design, build, and rollout of generative AI use cases in the automotive industry.

Shukhrat Khodjaev

Shukhrat Khodjaev

Shukhrat Khodjaev is a Senior Global Engagement Manager at AWS ProServe. He specializes in delivering impactful Big Data and AI/ML solutions that enable AWS customers to maximize their business value through data utilization.

Tobias Altenbuchner

Tobias Altenbuchner

Tobias Altenbuchner is a project lead at BMW Group for artificial intelligence across all business divisions. In his role as AI tech lead, he is responsible for the tech delivery of large-scale AI and generative AI projects.

Dr. Tobias Schmidt

Dr. Tobias Schmidt

Dr. Tobias Schmidt is a Partner at BCG and part of the leadership team for AI in Automotive. He is specifically focusing on analytics/AI solutions in supply chain management and volume steering and has supported OEMs and suppliers globally over the past 12 years.