Just a few years back, robotic process automation (RPA) was the poster child for workplace efficiency: these are software robots that are used for frequently recurring, rule-based tasks like moving customer data from one database to another or to automatically send emails to customers. Then large language models (LLMs) arrived, making it possible for machines to handle messy input, and pull answers from data on their own, which opened up smarter ways to automate decisions and save time. Now, AI agents are on the job, promising companies cleaner operations and faster results. According to a recent US survey1, nearly four out of five organizations say they’ve introduced AI agents somewhere in their workflow – though most are still figuring out how to move from experiments to company-wide impact. The challenge ahead is practical: to find out how process intelligence can deliver in places where legacy systems still dominate.
In general, an agent is a system that can act autonomously. Agentic systems have been around for quite some time now. Just think about RPA, driverless cars, or even thermostats. They also sense and decide without constant human oversight. But the rise of LLMs has given these systems a special ‘spice’, a new kind of intelligence – the ability to think, act, and observe. To interpret context, understand goals and plan actions. Unlike classical automation, AI agents pursue their own goals and orchestrate workflows. They access tools, data sources, collaborate seamlessly with humans and other agents, and learn from their results. At the heart, it’s not just about automation – it’s about true process intelligence running through the company.
Agents make everyday life easier by handling complex, judgment-based tasks, sorting through lots of information, helping us decide quickly, and personalizing services. This frees up more time for what we care about, but the real impact comes from their ability to continuously learn and adapt even in dynamic environments, further optimizing how things get done across the whole organization. Just imagine the complexity behind our network at Deutsche Telekom: thousands of servers, cables, antennas firing across various frequency bands – all running at the speed and scale of 5G. The real challenge we were facing here was to detect problematic, congested cells and adjust performance to prevent service problems: the RAN Guardian agent2 helps exactly here by monitoring the radio access network (RAN) around the clock. It functions as a multi-agent system, meaning that several specialized agents work together to ensure optimal network performance.
An example: the agent scans public sources, such as social media, identifies upcoming events in Germany, and classifies them based on their scale and location. Then, another agent steps in the process and checks how well the mobile network can handle the traffic during each event and suggests optimization measures if needed. Finally, an additional agent can make adjustments by e.g. reallocating network resources and also documents the whole process to further improve and optimize the network for future events. The RAN Guardian agent uses multiple data sources and works with modern AI models like Google Gemini 2.0. When an area becomes quieter, the AI reduces capacity there, saving energy. That’s an important step toward a “self-healing” mobile network.
Or let’s take our AI Engineer that supports software developers. The agent is proficient in modern programming languages like Python and TypeScript, but also in legacy ones such as COBOL or C++, and helps developers in generating, testing, and documenting codes. With it, applications that used to take months to develop can be programmed within minutes. Because in software development, ‘magic’ is not coding; it’s understanding what the customer wants, what problem to solve with the code, and how it fits into the existing IT system – and developers can focus on exactly these.
At the heart, it’s not just about automation - it’s about true process intelligence running through the company.
Frederick Peters, Chapter Lead of the Digital Enabler Team at T-Systems
The catch is, most companies still rely heavily on old IT systems: core banking platforms, ERP modules, HR databases – many of them designed about 20 years ago. These systems are reliable but stubborn. They weren’t build to feed data to machine learning models, let alone to cooperate with AI agents. At the same time, companies are heavily investing in AI systems to meet rising customer expectations, automate processes and boost productivity. This creates a tricky mix: Old systems are rigid and isolated, while AI requires data, transparency, and structure to be effective. It’s no wonder that 85% of companies admit they find it hard to scale AI effectively.3
And the gap is growing and growing, especially as LLMs keep expanding what AI can do. Analysts predict that by 2028, a third of enterprise software will include AI agent features.4 But the reality: legacy isn’t going away. The question is: How can companies close this gap and bring AI agents into their existing IT reality?
To build a solid foundation for agentic AI, businesses need to master orchestration. This means that tasks and information are handed off seamlessly, decision-making stays transparent, and processes flow end-to-end without interruption. This allows AI agents to be integrated into existing workflows. Old systems remain in place but gain a smart layer enabling transparency, flexibility, and control. For example: An urgent customer request comes in, the AI instantly analyzes the context and hands off just the right pieces to the old HR database for records, or the ERP for supply data, all in real time. Meanwhile, humans step in only when their insight is needed, or a judgment call is required.
Instead of silos, orchestration shapes a true flow. A great example on how this can be done is our HRCules, an AI-driven HR platform at Deutsche Telekom. The whole, fragmented HR ecosystem needed to be modernized by replacing a 15-year-old legacy IT system, without disrupting business operations. With the Pega platform, which combines AI and low-code capabilities, we built an orchestration layer between old and new systems. That lead to an efficiency increase in HR operations of around 80%, leading to leaner, more transparent processes and significantly increased employee satisfaction. This project shows how companies can lay the foundations for agentic AI.
The potential is obvious, but the real drivers of progress and responsible use happen behind the scenes.
AI remains an existential topic for our economy and every single company. Thanks to our strategic partnerships and extensive experience in realizing AI and automation projects for our customers across various industries, we are prepared to help companies in shaping their future in the agentic era. In close partnership with NVIDIA and SAP, we’ve recently announced a decisive milestone for the digital sovereignty of Germany and Europe: In Munich, we’re building one of Europe’s largest and most modern AI factories, the Industrial AI Cloud. It will be equipped with 10,000 NVIDIA GPUs of the latest generation (Blackwell), and with that we are increasing the AI computing power in Germany by about 50 percent. As early as the beginning of 2026, major customers, SMEs and public authorities will be able to use our AI factory for various applications: to simulate production facilities, crash tests, carry out digital wind tunnel tests for cars and aircraft, train robots or develop and operate their own AI models. The Industrial AI Cloud is embedded into our T Cloud ecosystem and ensures full data sovereignty, and compliance with the strictest national and European regulations.
1 PwC’s AI Agent Survey, 2025, pwc.com
2 https://www.telekom.com/en/media/media-information/archive/ai-agents-for-mobile-network-1099054
3 The Automation Gap: Making Legacy Systems And AI Work Together, Jakob Freund, 2025, Forbes Technology Council
4 Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 2025, gartner.com
Driving Efficiency With Agentic AI, Saurav Gupta, 2025, business-reporter.co.uk.
