Interview: Where is the future of AI heading?

No other field is developing as rapidly as Artificial Intelligence (AI). What was considered innovative yesterday is often already outdated today. In this interview, Christoph Bräunlich, Head of AI at BSI Software, provides an outlook on the most important trends for 2025, such as AI agents and reasoning models like DeepSeek-R1, which is already available in the BSI Customer Suite only one week after its release.

Which changes are you expecting to see regarding AI?

Christoph: The development of AI models such as DeepSeek-R1 (DeepSeek), o3 (OpenAI) and Opus (Anthropic) is in a state of change. While the scaling of data and model size has been crucial for improving Large Language Models (LLM) in the past, it now seems that a plateau has been reached. Even new models show less significant leaps through scaling, which raises doubts as to whether scaling alone will continue to be sufficient as a driver of progress.

In addition, the challenges for smaller competitors are increasing. Take, for example, the immense power consumption involved in training and running AI models: The training process of a single AI model is estimated to use the annual power consumption of around 12,500 households – a burden that, in the future, only large AI forges will be able to handle. At the same time, we hear of large providers sometimes charging 1,000 dollars per question for simple queries from their chats. The logical conclusion of this development is that a change from “bigger” to “smarter” will be necessary, as we can see very clearly with the example of DeepSeek-R1.

What does “smarter” mean in concrete terms?

Christoph: There are several possible solutions. One approach is Agentic AI, also known as agent-based Artificial Intelligence. These are AI models that can pursue goals autonomously and are supplemented by engineering. The agents provide the AI models with tools in order to be able to plan, calculate and execute actions.

Is Agentic AI already available for BSI Software customers?

Christoph: Of course. At BSI Software, we have been working intensively with Agentic AI for a long time, responding quickly to innovations – especially thanks to our agnostic use of models. This means that we don't rely on just one model provider, but integrate any models we want, relying on open source for maximum flexibility. A current example of this is DeepSeek-R1, which we integrated a week after its release and are already using for even more powerful AI agents. However, the best model is of no use to us if it is not deeply integrated. To this end, we developed the concept of the so-called lambdas with which AI Agents can perform any given action in the entire BSI Customer Suite – regardless of whether a new data record needs to be created in BSI CDP, an interface called up in BSI EIP or a business case carried out automatically in BSI CRM. The BSI Customer Suite also offers the option of easily configuring brains yourself. Overall, Agentic AI helps make processes within the BSI Customer Suite more efficient, simpler and more sustainable, and we are convinced that the AI-driven automation of business processes represents a significant efficiency factor for our customers. At the same time, we naturally take into consideration that the involvement by the “human factor” remains possible. We can automate everything – depending on their business model, our customers decide for themselves whether they want to automate using AI agents or deliberately employ a real person in certain areas.

Which technologies are available to AI agents?

Christoph: In addition to calling up tools via lambdas, many other patterns are available, such as retrieval-augmented generation (RAG), which gives AI models extended access to external information. The combination of “information retrieval” and “large language model” is already being used successfully by BSI customers. Another pattern is the so-called test-time compute, which is used by DeepSeek-R1, for example. Here, AI models are given additional processing time during execution so that they can generate and evaluate several potential solutions in order to develop the solution iteratively. This way, the LLM output is scrutinized one more time. Another approach that the agents can use is the multimodality of models. These are models that can not only work with text, but also with images, voice and videos. This way, whole new applications close to human conversations can be enabled. Or documents can be analyzed and evaluated.

How difficult is it to build AI agents?

Christoph: A project can be time-consuming, of course, but a functioning agent can be created very quickly. At our last BSI Customer Summit, a competition was held to see who had the best AI idea. The winning idea was an AI agent that compares third-party contracts with an insurance company’s own contracts and presents arguments in favor of the company’s own insurance product. The configured agent was presented just hours later – still at the Customer Summit.

AI Trends 2025 in a nutshell?

Christoph:

1. Progress in AI development by scaling ever larger models is slowing down – even with increasing computing power. This leads to a reorientation and to engineering technologies such as test-time compute.
2. Therefore, it will become increasingly important in the future to link AI models in a clever and – ideally – easily configurable way. General solutions such as o3 from Open AI are cost-intensive. Specialized solutions are targeted, economical and sustainable.
3. Within BSI’s CRM/CX software, intelligent solutions are created with just a few clicks and fully automated. For that purpose, BSI Software continuously invests in the engineering and integration of the right AI models.

Thank you very much for these fascinating insights!