More time for value creation: Roland Berger relies on AI agents and Microsoft Azure

The challenge: Finding valid answers to complex questions

For management consultants Roland Berger, having the right information in the right place at the right time is essential. That’s why employees have access to the company’s in-house knowledge base. But searching through documents in huge databases and knowledge repositories to find non-customer-specific information, such as method and industry knowledge, employee expertise, or specific study results, was at times laborious and time-consuming.

“Generative artificial intelligence can provide support here by automatically finding and processing information. That means it has the potential to change almost every workplace and speed up processes enormously,” explains Michael Schich, Vice President of Tech & AI Innovation at Roland Berger.

To harness this potential and make knowledge quickly and easily accessible within the company, Roland Berger set up an AI platform called the Knowledge Hub. “We didn’t want our consultants to have to keep trawling through countless studies for information; we wanted them to get relevant results simply by asking a single question,” Schich says. “But it wasn’t long before our colleagues were telling us that the Knowledge Hub had reached its limits, especially when it came to more complex issues that involved bringing knowledge together from different sources.” For example, if consultants asked the Knowledge Hub for information from a single study, it was able to give them an answer without any problems. But if they wanted to obtain filtered knowledge from three different studies, they sometimes had to ask up to three separate questions. In addition, each answer would come back with only its respective source information displayed, making it very time-consuming to refer back to the documents for review purposes.

But users also made it clear what enormous potential they felt an AI-based knowledge hub had. And so Schich and his team decided to redesign the Knowledge Hub to include various Microsoft Azure services and additional AI agents. The idea was that this would make the Knowledge Hub even smarter.

The solution: Efficient, automated processes thanks to AI agents and Azure services

Roland Berger replaced the prototype with “reasoning acting agents,” or ReAct agents for short, which are able to develop an answer to a question iteratively and autonomously. “This approach is particularly suited to processes, where getting from the question to the answer isn’t overly complex. However, that wasn’t really ideal for our use cases—which is why we decided to use graph-based agents, each of which is an expert in a different area or process,” Schich says. These agents take user requests using natural language and break them down into individual steps or tasks. Each task is assigned to a single agent, which works iteratively with the other AI agents. The agents are connected to each other by a graph and share a process memory. To complete their task, the AI agents pinpoint and retrieve knowledge from Azure AI Search. Finally, they hand over their results to the AI agent that received the request. This agent now generates a valid response based on the results. In the process, it doesn’t draw only on internal SharePoint pages: depending on the data classification, the AI agent decides for itself whether and which external data sources and databases, online encyclopedias, or search engines, such as Microsoft Bing API, to consult. The agent also creates the associated queries itself, and is responsible for independently classifying, evaluating, and summarizing the answers. Continuous feedback from users ensures that the high-quality requirements are met. Quality assurance is also carried out by means of evaluation in Azure AI Foundry.

For the AI agents to work effectively, it’s essential that the data corpus in Azure AI Search always be up to date. To that end, it integrates knowledge in the form of data in various formats originating from classified databases and those that have been explicitly made available. Azure AI Document Intelligence makes this knowledge usable by extracting all the information from the various formats. The process is coordinated in Azure Databricks. Processing is parallelized to make the large volumes of data manageable. Azure Database for PostgreSQL is used to store and manage user-specific information, and Azure OpenAI provides access to pre-trained AI models for analysis.

Now, when consultants are looking for information contained in three different studies, they can do so in several ways: either the AI itself selects a suitable AI agent for this task using an automated selection process, or the users themselves visit the user interface, which was developed in-house, to select the AI agent specialized in this scenario. This user interface is provided via Azure Web Apps. Once they’ve entered their question, users receive information and references as well as citations from the original sources—regardless of whether the knowledge is in-house or from third parties. “By using the graph-based AI agents, we were able to reduce response times to an average of eight seconds. The fact that answers come with sources and citations also increases employees’ trust in generative AI and acceptance of these applications,” Schich says. “Now, thanks to automated processes, our consultants can take the hours they used to spend searching through documents and invest that time in activities that add value. Particularly in an industry like ours, where efficiency plays a major role, that’s incredibly valuable.”

“AI agents are revolutionizing process automation and greatly reducing the time and effort required for data preparation. This progress is accelerating the take-up of AI in various application developments, and it opens up new opportunities and efficiency gains,” adds Maria Mikhaylenko, Senior Partner, Global Managing Director at Roland Berger. Among these new possibilities are AI agents that suggest the perfect team for a new project based on employees’ profiles, capacity, and expertise, or ones that take brainstorming to a whole new level. Schich says: “Our consultants can simply type in the topic they want to work on. The AI agent can then automatically create a structure with individual chapters, independently search for relevant information, and add key topics from its own internet research—deciding for itself which to include. The end result is a fact-based briefing.” However, rather than simply outputting this briefing as an answer, the AI agent uses it to create different personas. All of this happens out of users’ sight in the background. “This way, on the basis of just one input, a team made up of different personas is always ready to offer individual responses, each with their own perspective on the topic. This gives employees new ideas and food for thought,” he says.

According to Schich, the versatile support provided by the AI agents is very popular with Roland Berger’s employees: “The AI agents are already being used by around 70 percent of our workforce. This clearly indicates the advantages of the solution and shows that it’s being successfully integrated into everyday working life.” To ensure that this remains the case in the future, Roland Berger is offering its employees training and education to support process integration. Participants learn how to get the most out of the AI agents using a range of state-of-the-art prompting strategies.

These steps enabled Roland Berger to turn a simple “chat with your data” solution into a complex multi-agent environment. Developing the AI agents was always an iterative process: “We continuously made improvements based on our experience and on user feedback,” Schich explains. But following this successful implementation, the team is not resting on its laurels. Microsoft services and AI capabilities are constantly evolving, and user feedback is also continuously opening up new applications and solutions. One example of this is how the agents developed in-house are being integrated into Microsoft 365 Copilot to make work easier where it’s actually being done. “By combining AI agents and Copilot, we should be able to eliminate the existing system discontinuity and provide our consultants with support that’s tailored even more closely to their working environment,” Schich says. Another example is agent systems that have greater freedom to collaborate with each other. For Roland Berger, this is a promising approach for the future that will permanently change the way its consultants work, further simplify access to knowledge, and make processes even more efficient.