Deutsche Telekom put a Gemini-based agent into its live mobile network — major-event response fell from hours to about a minute

Last Updated: October 16, 2026

The problem: a network that moves faster than its runbooks

Every mobile operator has the same structural problem. Traffic is not evenly distributed in space or time — it clumps around Christmas markets, Carnival parades, stadiums and festivals — and the operational tooling that detects degradation was built on static thresholds and rules written months earlier. Deutsche Telekom and Google Cloud put the diagnosis bluntly in their original announcement: as telecom networks become increasingly complex, traditional rule-based automation falls short in addressing real-time challenges.

The second, less obvious problem is fragmentation. A modern network spans radio access, transport and core domains, each with its own monitoring stack. When a service degrades, the signal that explains it often sits in a different system from the one that raised the alarm. Deutsche Telekom's later framing of this is that operational teams often rely on fragmented systems that slow root-cause analysis and resolution. Humans bridge those systems by hand, and that bridging is where the hours go.

Traditional network management approaches are no longer sufficient to meet the demands of 5G and beyond. We are pioneering AI agents for networks, working with key partners like Google Cloud to unlock a new level of intelligence and automation in RAN operations as a step towards autonomous, self-healing networks

— Abdu Mudesir, Group CTO, Deutsche Telekom, Deutsche Telekom newsroom, February 2025

What was actually built

RAN Guardian is not a chatbot bolted onto a network operations centre. According to the joint announcement, the agent was built using Gemini 2.0 in Vertex AI, with Cloud Run, BigQuery and Firestore providing execution, analytics and state. The division of labour matters: BigQuery holds the telemetry at scale, Firestore holds agent state, Cloud Run executes, and the model does the reasoning that rules used to do badly.

The published capability set is deliberately narrow. The agent performs autonomous RAN performance monitoring, analysing key network parameters in real time to predict and detect anomalies; AI-driven issue classification and routing, prioritising degradations using network monitoring, inventory, performance and coverage data; and proactive network optimisation, where it recommends or autonomously implements corrective actions, including resource reallocation and configuration adjustments. That is a bounded, largely reversible action space — a design choice worth copying.

What the production numbers say

The February 2026 release is where the operational evidence lives, and it is unusually specific for a vendor-adjacent announcement.

The Carnival example is the most instructive. Deutsche Telekom reports that during the February Carnival season the agent identified around 130 events and parades each expected to draw over 10,000 participants, served by 611 different mobile sites, all pre-checked for issues by RAN Guardian, with most also monitored live. Five sites actually hit peak loads and were adjusted during the event. That ratio — hundreds of sites watched, a handful touched — is what a well-tuned agent looks like. An agent that intervened on 200 of 611 sites would be a liability, not an asset.

From one domain to many: MINDR

Having proved the pattern in the radio access network, Deutsche Telekom generalised it. MINDR — Multi-Agentic Intelligent Network Diagnostics & Remediation — applies the same agentic principles at the service level across RAN, transport and core. It was developed with Google Cloud's Autonomous Network Operations framework, and built with Google Gemini models on Vertex AI, operating as a collaborative multi-agent system in which specialised agents collect and correlate network data, build a real-time end-to-end view of service performance, and support root-cause analysis and controlled remediation across domains. The company also stated an intention to use the A2A (Agent-to-Agent) protocol for coordination, subject to its availability.

MINDR extends AI-driven intelligence from individual network domains to an end-to-end service approach, enabling us to proactively diagnose and resolve issues before any customer impact. It represents another significant step toward autonomous, self-healing networks.

— Abdu Mudesir, Board Member, Product & Technology, Deutsche Telekom, Deutsche Telekom newsroom, February 2026

Note the sequencing, because it is the part most organisations get wrong: single agent, single domain, single country, in production for a quarter — then multi-agent, multi-domain, multi-country. Following the German rollout, RAN Guardian is scaling across Deutsche Telekom's European national companies, starting with the Czech Republic and Croatia.

The other half: service automation at group scale

The network agent is the sharpest story, but it sits inside a much broader programme that Deutsche Telekom laid out at its AI Investor Day in Bonn on 5 October 2026. On the customer-facing side, the company reported that the chatbot Frag Magenta handled approximately 2.6 million customer service calls in the first half of 2026 alone, that in the US the number of customer service calls has dropped by 55 percent with AI agents there handling 40 percent of customer contacts, and that initial proactive-service use cases show a 30 percent decrease in complaints.

Frag Magenta itself long predates the generative-AI wave. Deutsche Telekom's own description notes it now helps with more than 380 different requests, can access individual customer data after login to perform tasks such as retrieving bills or changing bank details, and that in 2024, "Frag Magenta" handled nearly 10 million customer conversations. The chatbot resolves more than one-third of requests immediately. When it fails, it hands the human agent the transcript and a summary — a deflection design that treats escalation as a first-class path rather than a failure state.

The money — and where the evidence thins out

At the investor day the group set out targets rather than realised results: for 2027, Telekom expects AI and automation to generate gross savings of approximately 1.1 billion euros outside the U.S. compared to 2023 . By 2030, it aims to achieve savings of approximately 2.5 billion euros in indirect costs, also compared to 2023. On the revenue side, it expects roughly €250 million in AI-related revenue from businesses outside the US in 2026, rising to about €800 million by 2030.

What matters most is that technology delivers measurable value, for our customers and for our company.

— Tim Höttges, CEO, Deutsche Telekom, Deutsche Telekom newsroom, October 2026

Those are the right words. They are also, in the public record, applied unevenly. The network numbers are specific and dated. The savings numbers are multi-year targets carried in a release that explicitly flags its own forward-looking statements as subject to risks and uncertainties largely outside the company's control. Treat the first as evidence and the second as guidance.

What the public record does not establish

Four gaps are worth naming plainly. First, precision: the company reports over 100 autonomous remediation actions in month one, but not how many were later judged correct, nor the false-positive rate. Second, guardrails: the releases describe explainable remediation and controlled action, but not the approval thresholds, rollback mechanism or human-override design that sit behind an agent allowed to change live network configuration. Third, customer outcome: no CSAT, churn or complaint metric is attributed specifically to RAN Guardian — the 30% complaint reduction cited at the investor day relates to proactive service use cases in the connection-provisioning context, not to the RAN agent. Fourth, status: MINDR was described in February 2026 as having first production releases planned for later that year, and we found no subsequent Deutsche Telekom release confirming it reached production.

There is also an attribution question on the service side. The 55% drop in US customer service calls appears in the same paragraph as AI deployment, but the release does not isolate how much of that decline is AI-driven versus the product, app and process changes T-Mobile US has made in parallel. An honest reading is: directionally strong, causally unproven.

What an integration team should take from this

The transferable lesson is scope discipline. Deutsche Telekom did not start by asking a model to run the network. It picked one domain with dense, machine-readable telemetry; one recurring event class with a predictable shape (crowds generate traffic surges at known times and places); one bounded set of actions the agent was permitted to take; and one metric with a clean before/after. That is why a number like "hours to around a minute" is quotable at all — the baseline existed and was measured.

The second lesson is architectural sequencing. Pre-check, detect, classify, act is a pipeline, not a prompt. Most of the value in the Carnival example comes from the first step: anticipating which 611 sites matter before anything degrades. Teams evaluating model options often over-index on reasoning quality and under-invest in the retrieval and correlation layer that decides what the model even sees — here, BigQuery telemetry joined to inventory, performance and coverage data.

The third is that multi-agent is an upgrade path, not a starting point. MINDR's A2A coordination and cross-domain correlation only became tractable once a single-domain agent had earned operational trust. If you are considering a similar progression in your own operations, our research and services pages cover how we evaluate readiness — or you can get started with a scoped assessment.

We’re just getting started. We want to roll out across the entire company what’s working today in individual applications. This will enable us to grow faster, work more efficiently, and set new standards for quality.

— Tim Höttges, CEO, Deutsche Telekom, Deutsche Telekom newsroom, October 2026