Deutsche Telekom is replacing some conventional telephone menus with AI agents in its customer service, according to its chief AI officer, Kartik Sheth. At a HumanX session reported by The Next Web, he said the share of interactions that never reach a human was rising by one or two percentage points a month. That is an operational signal, but it leaves an investor's central question unanswered: are customers getting their problems solved at lower total cost?
The distinction matters because a call that stays with a bot can mean a quick router reset, an abandoned attempt or another call later. The report does not publish the underlying call counts, a current overall automated-resolution rate or a reconciliation to service spending. Sheth's remarks therefore support the claim that the company is changing how calls enter its service system; they do not, on their own, establish a quantified earnings benefit.
The reported gain measures handoff, not the whole outcome
The conference program identifies Sheth as Deutsche Telekom's chief AI officer and lists a September 24 discussion on deploying agents at enterprise scale. In the original account, he said that when AI is introduced for a new use case, it initially handles no more than a fifth of those calls, with performance improving as the system is tuned. He also described reaching higher shares as progressively more difficult because the remaining cases are more complex. These are management descriptions of deployments, not a published, independently checked group-wide series.
The phrase "never reach a human" describes a handoff boundary. It does not tell the reader whether the customer's issue was resolved, whether the answer was correct, how long the interaction took or whether a repeat call followed. A rising containment share could represent productive automation, but it could also conceal rework if the wrong calls are held back. The useful comparison would track the same type of problem through resolution, escalation and repeat contact, alongside the no-human rate.
Nor does a percentage-point monthly rise imply an indefinitely straight line. The starting level and mix of requests are not supplied in the report, and the hardest calls may be the least suitable for automation. Extending the quoted pace into a forecast of future staffing or margins would add assumptions the evidence does not support.
Old systems determine the savings path
Sheth gave practical examples: agents can help with router checks, reboots and bill explanations, while more complicated problems go to experienced staff. He also described adapters and caching used to connect newer agents to older systems. A successful bot needs reliable access to the actual service action; fluent speech without a working backend would not fix a customer's connection. Conversely, routing a straightforward case directly to an agent could save waiting and free staff for difficult cases.
The company's 2025 annual report on data and AI already described using language models for some network-complaint tickets and improving its Frag Magenta chatbot to handle less scripted questions. It presented reduced incoming calls to advisers as a future aim. The conference account is more specific about voice-entry routing, but it should be read as a continuation of that service strategy, not evidence that every call centre has been replaced.
This internal service change is also different from two products that might be confused with it. The company separately markets CoMind, a voice and chat offering for business customers, and has discussed an AI assistant added to ordinary phone calls. Revenue from selling a product, cost savings in its own support lines and features for subscribers are three different economic channels. The new comments primarily concern its own customer-care routing.
A prior service metric offers a different test
In its 2025 efficiency disclosure, Deutsche Telekom reported a first-call resolution rate of 76.0%, up from 74.1% in 2024. That measure asks whether an issue was solved at the first contact. It is not interchangeable with Sheth's 2026 no-human measure: the periods, populations and definitions differ. It does, however, show the kind of customer-outcome metric needed to judge whether automation is improving service rather than merely shifting the queue.
The same disclosure says the group has invested in digital touchpoints and an AI-supported service platform to simplify workflows. Those investments carry integration and operating costs. Savings from fewer routine human contacts would have to outweigh model, platform, monitoring and escalation costs, as well as any extra expense from repeat calls. No public calculation in the cited material isolates that net effect for the newly described voice agents.
There is a credible positive case: customers with simple tasks may avoid menus and hold times, while agents focus on cases where judgment matters. The countercase is that automating easy contacts leaves human teams with a harder and potentially longer workload, and that mistaken answers produce more rework. Both mechanisms can operate at once. An investor needs matched operating and quality evidence before assigning either one a dominant financial effect.
The investment case requires matched measures
A stronger future disclosure would report the number and type of calls offered to AI, the share genuinely resolved without repeat contact, transfer rates, customer complaints and cost per resolved case. It should distinguish a pilot or country launch from a group-wide deployment. A consistent first-call resolution series and service-cost trend would help reveal whether more containment coincides with better outcomes and lower spending.
For now, the confirmed event is Sheth's reported description of a rollout and its management-tracked no-human share. Deutsche Telekom's earlier annual report confirms a broader automation strategy and supplies an older quality benchmark. Net profit, customer loyalty and employment consequences remain unquantified by these sources; treating them as achieved would run ahead of the evidence.