Why governance is key to Deutsche Telekom's new AI-centric architecture

“There's a lot of excitement about AI, but it's creating a lot of fragmentation because there are a lot of agentic AI pilots and proofs of concept happening everywhere,” according to Shekhar Kulkarni, Global Chief Architect, Deutsche Telekom. His words sum up the experience of many technology leaders across the industry.

In the second half of 2025, Deutsche Telekom decided to address this problem head on by developing the Magenta AI-centric Reference Architecture (MARA), a holistic blueprint to embed AI into our operating model and enable AI transformation at scale across IT and networks, explains Kulkarni in an interview with TM Forum’s Insight.

“Because technology is evolving so fast, you don't want to mandate every single thing, because then you also effectively kill innovation,” says Kulkarni. “You want to get the right balance between innovation and standardization, and that balancing act is essentially done through governance.”

Deutsche Telekom has incorporated many lessons from its initial trials and deployments of agentic AI into MARA.

“The biggest lesson is that you need a platform and capabilities that allow you to experiment faster, with clear explainability behind the decisions agents are taking.”

AI agents that can explain their actions are an obvious necessity to a regulated company responsible for the security and privacy of customer data. But explainability also helps overcome internal obstacles to scale and innovation: “If you cannot explain agent behavior, then you cannot improve,” points out Kulkarni.

One example of the hurdles Deutsche Telekom wanted to overcome is the stagnation in the deflection rate, which is the rate at which AI handles tasks or inquiries without needing human intervention.

Mapping capabilities

The starting point for building MARA was creating a capability map, and this involved asking and answering many questions about how agents should be deployed and what their impact will be.

“We started from customer interactions,” explains Kulkarni. “How do we understand the customer intent? How do we orchestrate those intents across multiple systems? How do we deploy these agents? How do they access the back ends? Are APIs the only way? Do these agents now need skills and tools?,” he asks.

"Most of these capabilities are new to enterprises like ours, and they either need to be built or bought and integrated into an already complex enterprise landscape."

Next was the development of a technology blueprint to put these capabilities into production in a way that avoids both fragmentation and duplication.

To this end, MARA defines a structured, layered, AI-first system, notes Kulkarni, with the following layers: multimodal interaction, agent, multi-agent orchestration, AI access and model mediation, knowledge and intelligence, governance, trust and control, and edge and runtime infrastructure.

MARA’s structure sets out to enable agents to connect to back-end systems in a consistent, harmonized manner, operate as determined, deliver the right outcomes, and explain their output. MARA also defines the integration of security and privacy policies, and how to manage language model gateways, as well as the use, cost, and performance of models.

“The whole blueprint now consists of all the different capabilities that are really needed to adopt AI at scale,” says Kulkarni. “We are able to govern AI agents in a manner that we want to and add to our policies, whether it's security, whether it's the EU AI Act, or even our own guardrails,” he adds.

Certain capabilities are mandatory, such as the Model Context Protocol (MCP) gateway so that backend capabilities are exposed as tools to various AI agents, or the Model Gateway to ensure that various models are available in a secure and cost-efficient way. “Some of these capabilities are still maturing in the market, which is why MARA distinguishes between mandatory, preferred, and exploratory components,” says Kulkarni.

Security is a critical consideration, and agents will need to be carefully authenticated.

“Agents must be governed like digital employees, they operate under Zero trust, they must be authenticated and authorised," Kulkarni explains. “This is where some of the underlying capabilities, like MCP gateways and the model gateways, can come into play.”

Longer term, Kulkarni says he would also like to see policy by design, “so that policies can be embedded as a code into everything that you do. We are very much at the beginning of that journey, but we have a vision where every policy decision can be applied by agents, from the design to the deployment, and that's how we'll automate it.”

Working with vendors

Deutsche Telekom wants to keep its options open when choosing AI and software vendors. “You don't want to build everything; you want the ability to innovate with different vendors and different products, but in a controlled manner so that we can properly govern these agents as they are deployed.”

Magenta AI-centric Reference Architecture (MARA) therefore has “control points” to manage AI vendor integration both from a security and data privacy perspective and to avoid technology lock-in.

For example, Deutsche Telekom expects to work with multiple agentic frameworks, be it opensource versions such as LangChain, or vendor-specific ones from companies such as Salesforce. In addition, it will need to manage agentic use of a variety of language models.

MARA therefore requires each vendor to comply with the same standardized guardrails. These include ensuring that suppliers’ own agentic frameworks “are not directly coupling with our back ends, whether it's data sources or APIs. We ask them to go through our own MCP gateway. We expose these tools,” says Kulkarni.

In addition, vendors must provide access to the logs and metrics that can explain agents’ actions.

“If you don't have a good way of effectively observing, collecting each and every interaction within the agent, the workflow, each and every log, each and every trace, you are always fighting the next iteration,” says Kulkarni.

Pointing out that every agent needs to be registered, Kulkarni would like to see a shared agent registry across Deutsche Telekom’s operating companies. He would also like to see a common agentic marketplace “to make sure that everybody understands what is already deployed, and they can reuse it if required.”

Measuring Agentic AI’s business value 

Although Deutsche Telekom has high ambitions to use AI everywhere, Kulkarni readily recognizes that the full potential of Agentic AI is yet to be realised even though early signs of structural cost reallocation, reduction of technical debt, time-to-modernize acceleration are emerging.

In the meantime, Deutsche Telekom treats AI as an enterprise-grade capability with measurable KPIs, notes Kulkarni. Its AI quality and governance metrics include task success rate, hallucination rate, cost per transaction, and escalation rate.

And there have already been noteworthy breakthroughs in how AI helps business teams.

These include the use of AI voice and chat bots to serve customers and in automating how the company serves enterprise clients. “In the area of customer experience and B2B we are already seeing the operational cost going down because of the automation,” says Kulkarni.

When it comes to enterprise services in particular, agentic AI is helping Deutsche Telekom streamline the management of leads, contracts and pricing. Other benefits include higher customer conversion rates and average revenue per user as a result of greater personalization.

In the live production network in Germany, Deutsche Telekom’s RAN Guardian uses agents to proactively automate and optimize radio access network performance.

Out with the old

Another key area where agentic AI is making a sizeable difference is in consolidating architectures and retiring legacy IT.

“That's where the automation and the acceleration will come into play,” says Kulkarni. “In some of the European operating companies, they are able to replace old legacy applications within three months which they had planned to do in 2027.”

It’s not just a matter of decommissioning legacy software through reverse engineering, explains Kulkarni. There is also forward engineering, where an agent works within guardrails to deliver new target architecture.

“Traditionally trouble ticket management is largely standardized and very static, whereas in the agentic world you can use the power of AI in terms of analytics and reasoning to make that process more intelligent and adaptive.”

However, the use of agentic AI “needs to be selective,” and focus on adaptive processes, those that require multi-step reasoning, and decision-heavy workflows, says Kulkarni. “Determinism is required in certain billing or revenue assurance processes,” and transactional, deterministic processes are best handled as components.

New ways of working

The development of what Kulkarni describes as a “living stack” will reshape roles within technology divisions.

“Architects, developers, designers, testers will all be assisted by this living stack, which provides an exact view of what's deployed in the production,” he says. “What you model is what gets deployed, and then everything is in a controlled closed loop.”

Kulkarni gives the example of an API designer, who previously had to discuss with other people to understand the implications of API design specification change. “Now with AI, we consolidate all that knowledge into a model which they can use to really understand the impact much more quickly,” he says.

The API designer can therefore use AI to generate examples and “focus more on understanding end to end problem that we are trying to solve,” Kulkarni adds.

Enterprise architects meanwhile will have time to focus on business issues, roadmaps, governance, while solution architects focus on end-to-end solution delivery: “They would have to consider more capabilities, in terms of governance, security, model integration, MCP integration, which they've never done in the past.”

What’s next?

MARA is still in its relative infancy.

Deutsche Telekom is building on its use of TM Forum’s Open Digital Architecture (ODA) components and Open APIs, which Kulkarni describes as “a good starting point that allows you to really quickly move.” MARA treats AI agents as first-class architectural participants that orchestrate ODA components rather than replacing them. It integrates TM Forum Open APIs via standardized interfaces and many of them enabling tools and skills that AI agents can consume.

But Kulkarni says operators need more standardization beyond components and APIs. One of the areas where he is seeking greater industry standardization is around ontology, which Telstra, for example, has described as a shared language of how to model knowledge for a telco business.

“The semantics behind that data is critical for agents to perform,” says Kulkarni. “Standardization will allow us to really create this layer … because ultimately semantic has to be common.”

“For us, Magenta AI-centric Reference Architecture (MARA) is not just about deploying AI agents - it is about redefining enterprise architecture for an AI-first world. We believe that agentic AI will only deliver sustainable value when it is governed, observable, and integrated into the operating model by design."

MARA, he believes, provides that backbone, "enabling innovation at speed, while maintaining trust, compliance, and architectural integrity across Deutsche Telekom and our companies in the European footprint. Our ambition is clear: to move from fragmented AI experiments to a scalable, enterprise-grade, AI-centric architecture that positions Deutsche Telekom as a leader in AI-driven telecom transformation.”