Boost automotive productivity with process automation facilitated by Generative AI

Enterprises today are continuing to explore new solutions to increase efficiency, particularly around process automation. That efficiency is pertinent to multiple industries, but it is especially notable in the automotive sector. Automotive research and development (R&D) teams need to address costly, time-consuming, and error-prone processes, for example, in generating request for proposal (RFP) documents.

The development cycle for new vehicles is complex and can last for up to 5 years. During that time, thousands of components—such as infotainment modules—are designed, built, and tested. Part of the work of the development cycle is conducted by the original equipment manufacturer (OEM)—that is, the vehicle-producing company—but large portions of the effort are solved in collaboration with suppliers.

Automotive companies often use RFPs when awarding various types of work to vendors. The use of RFPs empowers automotive companies to use the expertise and capabilities of external vendors while also maintaining control over the selection process and making sure that chosen vendors align with relevant requirements and objectives.

Based on a given RFP, potential suppliers generate offers detailing their approaches through a two-step process. First, the suppliers’ technical approaches are evaluated by the department that issued the RFP. If a bid is approved by the issuing department, the OEM’s procurement team reviews the price. Depending on the number of applicants, the process can last anywhere from several weeks to months before the final suppliers are selected.

In theory, an initial RFP should contain all required information and context for the request, including fully defined technical specifications. But in practice, an RFP may contain only a short brief with vague descriptions. RFPs that are not issued under a strict process may miss critical information derived from previous projects and overall historical knowledge.

To address those challenges, Accenture, Amazon Web Services (AWS), and Intel have collaborated on an RFP generation solution using generative artificial intelligence (AI). In considerably less time, the RFP generator automatically gathers critical company information and pre-populates content in the RFP. Users then receive automated feedback thanks to a verbal exchange with the knowledge base, empowering them to focus on critical tasks elsewhere. Improved RFP quality streamlines supplier assessment by eliminating the need for back-and-forth communication with suppliers, making the entire process more efficient.

In this blog, we dive into specific features and benefits of the RFP generator, which is powered by Intel-based instances for Amazon Elastic Compute Cloud (Amazon EC2)—secure and resizable compute capacity for virtually any workload—and Intel® Tiber™ AI Studio.

Solution overview

The solution had to match a few essential parameters. First, it needed to augment the capabilities of automotive procurement employees, meaning that it needed to complement existing RFP-generation processes as opposed to fully replacing human-fulfilled tasks. Second, the tool had to use generative AI to deliver significant performance improvement at a price point that would make deployment at scale financially viable.

At the solution’s core, we put Intel® Tiber™ AI Studio, a cloud-agnostic machine learning platform. Intel Tiber AI Studio has a container-based architecture that helps facilitate seamless deployment across multiple environments, supporting hybrid cloud strategies and ease of migrating workloads. That adaptability extends to the creation of heterogeneous compute pipelines, where different stages can run on optimized architectures, such as preprocessing on CPUs and training on GPUs or HPUs, regardless of the provider.

We deployed Intel Tiber AI Studio using different AWS services. For orchestration, we used Amazon Elastic Kubernetes Service (Amazon EKS), a managed Kubernetes service to run Kubernetes in the AWS cloud. For DNS management, we used Amazon Route 53, a reliable and cost-effective way to route end users to internet applications. We integrated worker nodes from Intel® Tiber™ Developer Cloud, such as 4th Gen Intel® Xeon® Scalable Processor and Intel® Gaudi® 2 AI accelerators from Habana Labs, and served our self-hosted large language models (LLMs) as API endpoints through Intel Tiber AI Studio. That arrangement empowered us to successfully integrate the endpoints into the RFP generation and ticket-creating process.

Moreover, we delivered this solution using a self-hosted LLM (Mistral 7B) that is supported by Accenture and Intel’s AI Playground.

The details

Figure 1. RFP context diagram

Figure 1 illustrates the solution’s workflow for generating RFPs. The process begins with user input through a Streamlit-based frontend interface. That data is then processed by a Python backend for prompt engineering, making sure that the input is prepared for further processing. The backend connects with an endpoint to Intel® Tiber™, where a Kubernetes web service is deployed. Key elements of the solution are as follows:

Cache check: Helps verify if the same query was sent previously.

Retrieval query: Augments the prompt with relevant data.

PDF/Word export: Generates an RFP step by step, reviewing each document subsection sequentially.

The solution enforces a specific structure by chapter and subsection to generate one section at a time. That means the solution can provide individual instructions for each section and establish a standard for writing future RFPs.

Again, the challenge was not how to automate the process completely but how to create a tool that would more seamlessly complement an employee’s own process for creating context-specific RFPs—in as little time as possible. To do that, we used a self-hosted Mistral 7B model that met accuracy needs with a sufficiently large context window, a robust MLOps tool, and vector storage to house past RFPs.

Our model selection criteria were set according to a range of factors around data security, cost, and ability to handle a sufficiently large context window. Of utmost necessity was the incorporation of an open source model so that we could control deployment scenarios and provide greater data security by avoiding any dependency on a third-party provider. We looked at a variety of model sizes, and with this, the cost of deployment and selection of hardware to support the model endpoint were key considerations. We spent time comparing models as large as Llama 2 70B and models as small as 7–8 billion parameters. When we looked at the effectiveness of responses generated through our RAG pipeline, we found that we could get a high quality of response from a smaller model capable of handling a sufficiently large context window. That requirement was met by Mistral 7b, which we ultimately selected. Thus, in the end, a smaller model met our needs while being more cost-effective.

Key decisions

In sum, the key decisions that we made were as follows:

LLM selection: We selected Mistral 7B over other models in view of factors like greater security, lower cost, and ability to handle sufficiently large context. We spent time comparing other models, such as Llama 2 70B, but ultimately a smaller model met our needs while being more cost-effective.

Performance optimization: We used Intel Extension for PyTorch to help increase inference speed on Intel hardware.

Cross-cloud model for development and deployment: We used Intel Tiber AI Studio to stitch together workflows that might use components in different clouds (for example, fine-tuning an Intel Gaudi 2 AI accelerator in the Intel Tiber Developer Cloud) while still having our model-serving inference requests done on AWS.

Customer benefits

Model ownership and security: Keeping data in a “walled garden” helps increase overall security.

Time and cost savings: Potential time and cost reduction in RFP generation contributes to a more streamlined research and decision-making process.

Scalability and consistency: AI-powered solutions can better handle increasing volumes of RFPs and error tickets without compromising on quality.

Improved accuracy and compliance: The ability to integrate historical data and industry regulations reduces human error and helps customers meet automotive industry standards.

Faster time to market: Automating administrative tasks can accelerate the overall R&D process, potentially leading to faster product development and time to market of new automotive technologies.

Conclusion

Our pilot customer benefited from an industry-leading solution that streamlined its work process without introducing net-new AI costs that might have offset gains in productivity. By adopting the solution, the automotive manufacturer can stay ahead of its competition and meet growing industry demand.

Accenture’s generative AI solutions, powered by AWS and Intel technologies, offer improved efficiency and cost savings in automotive environments. The three companies have a long history of collaborating to drive business transformation through advanced technology solutions, which have spanned AI, cloud transformation, and edge computing. Our custom engagement makes use of Accenture’s deep industry expertise, Intel’s hardware and software innovation, and the AWS cloud and AWS services to provide a seamless customer solution. Reach out to us if you’d like to learn more.

Accenture’s AWS Partner Profile | Contact Accenture | Intel’s AWS Partner Profile | Contact Intel