Customer support · Production

Deutsche Telekom builds open-source LMOS multi-agent platform for customer-facing AI across Europe

The problem

Deutsche Telekom needed to deploy generative AI customer service automation across 10 European countries with different languages and multiple channels (chat, voice, autonomous), but no framework existed capable of scaling across this complexity without requiring deep Spring Boot expertise or tolerating brittle monolithic prompts.

First attempt

Early RAG chatbots on LangChain required deep Spring Boot expertise and heavy boilerplate; LLM prompts were fragile so any change risked breaking the entire chatbot; and no JVM-ecosystem multi-agent framework existed that could meet Deutsche Telekom's scalability requirements.

Workflow diagram · grounded in source

Customer request via channel

trigger

“for all of the channels through which customers reach us, which is chat channel, voice channel, and also autonomous use cases”

Intent-based agent routing

routing

“In front of that, we have an agent router that routes each incoming request to one of those agents”

Input filter validation

validation

“We have, for example, here, this CustomerRequestAgentDetector. If a customer comes and they ask specifically for a human agent, then this will trigger this filter, and that process would be then triggered. We then also have a HackingDete…”

LLM agent processes request

ai_action

“this entire prompt is generated on each request, meaning we can customize it, tailor it for each customer, each NatCo, or each channel, which is a very powerful feature that we really lie on heavily”

Output filter and anonymization

validation

“we want to detect if the LLM is fabricating any information. Here, we can use regular expressions within this filter to extract all the links and then verify that these links are actually valid links that we expect the LLM to be outputting”

Fallback to FAQ RAG agent

routing

“we have this UnresolvedDetector. As soon as the LLM says it cannot answer a question, this filter will be triggered, and then we can do a fallback to another agent, which, in most cases, is the FAQ agent, which in turn holds our RAG pipe…”

Reported outcome

The LMOS platform has served over a million customer questions with an 89% acceptable answer rate, deflected more than 300,000 human-agent conversations at a risk rate under 2%, and outperformed vendor products on agent handovers by 38%.
Domain agent development time fell from 2 months to 10 days.

Reported metrics

Questions answeredmore than a million

Acceptable answer rate89%

Human-agent conversations deflectedmore than 300,000

Risk rateunder 2%

Show all 9 reported metrics

questions answeredmore than a million

acceptable answer rate89%

human-agent conversations deflectedmore than 300,000

risk rateunder 2%

agent handovers vs vendor products38% better

domain agent development timefrom 2 months to 10 days

use case enhancement development timefrom weeks to 2.5 days

agent development time (F9 initiative)12 days

use cases released in one month12

Reported stack

LMOSARCKotlinOpenAI APIsKubernetesIstioDockerHelmLangChain

Source

https://www.infoq.com/presentations/ai-agents-platform/?topicPageSponsorship=88befbbd-30f0-4d18-9d43-0bf2cb3e751d

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The LMOS platform has served over a million customer questions with an 89% acceptable answer rate, deflected more than 300,000 human-agent conversations at a risk rate under 2%, and outperformed vendor products on age…

What tools did this team use?

LMOS, ARC, Kotlin, OpenAI APIs, Kubernetes, Istio, Docker, Helm, LangChain.

What results were reported?

Questions answered: more than a million; Acceptable answer rate: 89%; Human-agent conversations deflected: more than 300,000; Risk rate: under 2% (source-reported, not independently verified).

What failed first in this deployment?

How is this customer support AI workflow structured?

Customer request via channel → Intent-based agent routing → Input filter validation → LLM agent processes request → Output filter and anonymization → Fallback to FAQ RAG agent.

The LMOS platform has served over a million customer questions with an 89% acceptable answer rate, deflected more than 300,000 human-agent conversations at a risk rate under 2%, and outperformed vendor products on agent handovers by 38%.Domain agent development time fell from 2 months to 10 days.