Bosch: balancing act between cost savings with AI and know-how protection

Bosch uses generative artificial intelligence (AI) based on numerous standard models such as GPT from OpenAI and LLaMA from Meta in order to tap into as much internal expertise as possible using natural language. Andreas Nauerz, Head of Department at Bosch's digital division, explained this at the AI Summit of the IT association Bitkom in Berlin. The aim is to make information from manuals and online resources more easily accessible and constantly available. However, according to Nauerz, Bosch quickly reached the limits of the current large language models: These could not do anything with technical terms in particular. It was also important to safeguard the company's intellectual property and digital sovereignty. Bosch is therefore also relying on a cooperation with the German AI start-up Aleph Alpha, in which the Group also has a financial stake.

In general, Bosch has identified a good 160 usage scenarios for generative AI (GenAI) within just a few months, explains Nauerz. Essentially, the focus is on the areas of search and summarization, bot technology, content creation and software development. The aim is to increase efficiency, quality and user-friendliness and to "get the numbers in shape". One of the tools developed for this is the internal chatbot AskBosch. This also supports the internal legal experts, for example, by evaluating data from thousands of documents. The feedback from employees in turn improves the voice assistants. AskBosch now has 70,000 users and saves around one million calls per month in customer service alone, which should reduce costs by around 84 million euros per year.

Bosch works in this sector with Microsoft (Copilot), AWS and Google, among others. There are also special applications such as Synthesia for AI-generated presentations with avatars, HeyGen for translations of training videos, for example, and Adobe Firefly for image creation. In terms of return on investment (ROI), however, the Group cannot "try out an infinite number of tools". In addition to AskBosch, the GitHub Copilot for programming is heavily used. 6000 of the approximately 50,000 in-house developers are already using this bot and have generated four million lines of code with its help. 30 percent of suggestions are accepted directly. This is equivalent to a time saving of 20 percent per programmer, which in turn translates into 37.5 million euros less costs per year.

PhariaAI as the operating system for Bosch AI

Nauerz emphasized that Bosch also wants to use GenAI to "increase the competitiveness of our products" and tap into new business potential. The standard models are not so good for this. Theoretically, they would have to understand special terms that do not appear in the training data. Furthermore, the common large language models are difficult to implement on small end devices. Experiments have shown that when a bot receives the not entirely untypical error message "Error 530" in an e-bike battery, for example, it suggests using a screwdriver of a certain brand and wearing "clean clothes". In this case, you would simply have to switch the e-bike off and on again. An image generator also failed when trying to depict a drill because of the company name and logo.

"Bosch's unique knowledge is crucial for every single step in the use of GenAI," says Aleph Alpha founder Jonas Andrulis. Incorporating this into a large language model would not work. There is also the risk that "GPT 17" will suddenly make completely new insider knowledge from Bosch available to all users worldwide. "You can create improvements by mixing models," explained Andrulis. However, more than ten percent of the partner's data is still not readable. Even changing the order in a question-and-answer list can lead to a deterioration of up to 30 percent.

According to Andrulis, what is needed is a technology "that works with people", maintains their control and at the same time retains the expert knowledge in the AI system. Aleph Alpha recently presented such a process with PhariaAI, which is intended to form a kind of operating system for GenAI. Bosch uses its own language model Pharia-1-LLM as a basis for this, among other things. It is clear that "no general-purpose model will be as smart as the experts from Bosch." In a test, however, PhariaAI performed significantly better than the competition in the task of teaching the special GenAI German. This means that Bosch's expertise can now be better "captured in the logic of the application" without it leaking out.

(mma)