Generative AI offers great perspectives for better user and employee interaction, e.g., in sales trainings. The AWS team of T-Systems created a new team member for a European original equipment manufacturer (OEM) – a GenAI-based sales trainer. The sales trainer avatar enables considerable flexibility of sales trainings and helps the customer to save costs.
Acceleration in the time to market is an important topic for the Automotive industry – as new models that cater better to the demands of customers and drivers give the OEMs a competitive edge. Thus, a trend towards faster releases of new vehicle models using new processes and technologies is observable. Beyond the “usual” competition for market potential, the entire Automotive industry is undergoing a fundamental changeover to software-defined, electric vehicles. The proportion of value added by digital components in vehicles is continuously increasing.
Besides market and data analytics or insights, which other tools can OEMs utilize to create competitive advantages? In 2024, artificial intelligence, and especially GenAI, come to mind. “For a European OEM’s sales team, we created an AI-powered tool that exploits AWS GenAI capabilities,” says Artur Schneider, AWS AI expert at T-Systems.
GenAI is ideal for efficiently implementing innovative, interactive training formats. With the AWS-based avatar, we have created an easily reproducible solution for sales and marketing.
The acceleration of automotive development not only has consequences for production, but also for marketing and sales. The increasing number of models, the wide range of customization options, and vehicle specifications posed a major challenge for the sales management of a European OEM: How can the sales staff stay up to date with information and details on new vehicle models?
Previously, traditional approaches such as face-to-face training resulted in high travel costs and logistical challenges. But training staff for conducting online trainings, for example, is also in short supply. At the same time, different sales employees have different training needs, which the OEM was unable to meet with standard training courses. Video sessions, for example, lack interactivity with no option for real interaction through real-time Q&A.
The aim was to create an interactive, intensive learning experience with individual content on products and sales methods that would allow every employee to perform their individual training. At the same time, the sales management wanted the training program to be easy and efficient to organize. The launch of a new vehicle model was to be used to test the suitability of artificial intelligence for the training courses.
“To create a training experience that is as realistic and personal as possible, we developed a digital avatar together with the OEM. It takes on the role of the trainer in the training sessions and provides the participants with the necessary knowledge about the new model,” says Artur Schneider, explaining the solution approach.
The solution consists of several components. The avatar itself is running in Unreal Engine, while the powerful cloud platform from AWS takes care of the rest. First of all, AWS offers strong scalability based on training needs, including the necessary GPU resources to run an AI avatar. In addition, AWS also offers a wealth of ready-to-use natural language processing and AI tools that can be easily integrated into functional solutions: Amazon Transcribe transforms audio to text, Amazon Polly reverses text to audio.
But understanding language is of course not enough to hold training courses. This is where AI, especially GenAI, takes center stage: a Large Language Model (LLM) is operated in the background. It is provided by Amazon Bedrock. The pre-trained LLM receives specific context and product information via RAG (Retrieval Augmented Generation). A knowledge base is created in Amazon Bedrock, which then generates an Amazon OpenSearch vector database containing the necessary documents. When a user asks a question, a relevant query is created in the vector database, and the necessary information is sent to the LLM to generate the most appropriate answer. The avatar, therefore, at any time has the latest information on the vehicle model.
The technical architecture also contains the essential accompanying features in the background: user authentication and management, for example, is done with Amazon Cognito. The user interface and interaction logic are hosted within a Flask application running on an ECS cluster (container). Flask provides a robust platform for developing interactive web applications. This setup communicates with the backend services and displays the training content to the users.
The core of the interactive training experience is powered by Unreal Engine servers running on G5 EC2 instances. These servers render the digital avatar and provide a high-fidelity, immersive training environment. Multiple Unreal servers are deployed to handle the rendering load and ensure smooth performance during high traffic periods. Thanks to the pixel streaming feature of Unreal Engine, the avatar can be integrated and operated on a web page. Therefore, users connected to the Flask app can access the avatar after authenticating with Amazon Cognito. The AWS capacities in the background ensure real-time communication with the sellers – without long waiting times.
Training participants can talk directly to the avatar using natural language processing and output. This creates an individual discussion. Beyond that, the online availability of the training service allows sales staff to plan their training times according to their personal schedule. The OEM makes the training concept more flexible and offers its employees a new and intensive learning experience, regardless of the availability of trainers.
The customer did not need any prior AI knowledge to develop the avatar. The T-Systems team delivered the solution from a single source. By using predefined modules, it was possible to create an initial prototype within a few weeks, which was continuously optimized with internal data.
