What the technician actually sees
The old routine had several steps. A technician had to work out which elevator was affected, dig up the notebook with earlier service notes, and page through technical manuals. If he then found he needed another tool, he walked back to the car. Afterward he wrote up the job, often in that same car.
Microsoft says an agentic AI app now replaces much of that. It was built with Microsoft Foundry and Azure OpenAI in Foundry Models. Before Teichmann enters the building, the app briefs him on his phone or tablet. He learns which elevator and floor are involved, and whether the job is routine maintenance or a call-out such as a problematic door. That tells him which tools to bring. He can also search digital manuals and the unit's earlier service history from the same device.
After the job, a separate voice agent takes over. It asks him about the work in conversation, records his spoken answers, and asks follow-ups when it needs more detail. Those answers feed the record that other technicians draw on later.
Teichmann describes the arrangement in his own words. He says he is the boss at the job site and the AI is a buddy who gives advice. The reporting is consistent on this point: technicians inspect the equipment and use their training and judgment. Nothing in it describes an autonomous repair system.
The plumbing: a decade of connected-equipment data
The agents sit on a data foundation that took years to build. TKE began connecting elevators and escalators about ten years ago, using sensors and digital controllers. Microsoft says the data was pulled into a central platform using Azure, Azure IoT and Azure Databricks. The platform was developed further with Microsoft Industry Solutions Engineering, now called Microsoft Frontier Company.
Equipment variety made a common data environment important. TKE services units from other manufacturers as well as its own digital-native EOX and HELIX lines. Microsoft also notes that units can last for decades, so configurations and histories differ widely.
Structured versus unstructured data
Christian Jung, TKE's head of digital data and AI foundation, explains the shift. Earlier work covered only structured data, meaning IoT streams and records already sitting in enterprise systems. Technical documents and technicians' narrative notes were out of reach. Generative AI lets TKE extract, search and connect information across those sources. Jung says it also let TKE give technicians guidance on how to fix a problem.
This is a common pattern in enterprise AI projects. Sensor readings show what a machine is doing, but the explanation often sits in a manual or a previous technician's notes.
The reporting does not name a model version, a retrieval design, a grounding method or a validation benchmark. Readers should not assume any of those details.
Digital Operations Centers and the feedback loop
Microsoft says TKE has created nine Digital Operations Centers (DOCs). Analysts and technical specialists there watch the status of connected units, recorded events and error codes. Machine learning calculates "health scores" to help prioritize maintenance, and the teams compare data over time and across similar equipment. When they spot a pattern, they send recommendations to technicians.
The article gives one example. A door opens and closes several times before the lift moves, then behaves normally for days. A passenger would shrug it off. Connected-unit data can show whether the behavior keeps recurring. Tobias Lüsebrink, who heads digital transformation for TKE Germany, calls this abnormal operation that points toward a possible breakdown. The DOC can flag it and pass along the affected floor. The technician then inspects the unit and decides what to do.
Information also flows back. A technician may find that the signal does not match the problem on site, or that several faults produce similar symptoms. Debrief notes let TKE refine its troubleshooting and identify new patterns.
Lüsebrink adds a caveat. Technology cannot capture everything an experienced technician knows from the sound and feel of a ride. Teichmann puts it differently: the knowledge is in his colleagues' heads, and the application is how he passes it on.
Scale and company-reported results
The figures below come from TKE or Microsoft and have not been independently audited.
- Scale: According to Microsoft's feature, TKE has 25,000 service technicians who care for 1.4 million units worldwide. TKE says those units carry 1.5 billion passengers a day.
- DOC count over time: A TKE press release dated April 17, 2026 said the company had eight DOCs in operation, with more planned that year. The October Microsoft story says nine. These are statements from different dates, so the later number is the more recent one.
- U.S. pilot outcomes: The same release said that connected-unit implementation reduced unplanned service visits by 20,000 and cut callback rates by more than 40% in the US in 2025. It also reported a 33% drop in cancellation rates. TKE ties the visit and callback figures to cases where structured DOC-supported interventions were carried out. These are pilot results from the DOC model. They are not proof that the new technician briefing app produced the same gains everywhere.
Other write-ups of this story carry different numbers. One third-party article cites a 25% cut in emergency callouts and a 15% improvement in first-time fixes. Neither figure appears in TKE's own release or Microsoft's feature, so treat them with caution.
Corroboration and context
Microsoft's own materials from earlier in 2026 describe the same design. A Microsoft customer story says TKE built a real-time database giving technicians access to telemetry, service history and shared knowledge. A Microsoft Hannover Messe roundup says TKE's agents assemble contextual briefings before technician visits and capture insights afterward. A Databricks session summary describes the use of GenAI for service briefs and documentation, with human-in-the-loop validation and enterprise governance. That last detail is a useful hint about how the system is governed. The October feature itself does not spell out review, retention or privacy rules for voice debriefs.
What IT and developer readers can take from this
- Telemetry is not enough. Connected-asset data tells you what happened. The context that explains it often sits in manuals, ticket notes and people's memories. The TKE case is about making that unstructured material usable.
- The data work came first. The agents rest on roughly a decade of cloud connectivity and a common data platform. A Foundry deployment without that groundwork would have little to draw on.
- Capture is part of the design. A voice debrief lowers the effort of recording what was done. A system that learns only from structured fields misses most of what technicians know.
- The human stays in charge. The agents brief, search and ask questions. The technician makes the call. That is a sensible pattern for safety-critical equipment.
Caveats
This is a Microsoft-published customer feature, so it presents the partnership favorably. The reporting leaves several questions open:
- Which fault types the app covers.
- How the briefing is assembled from source systems.
- How accurate it is.
- How voice records are reviewed and retained.
The 12-year technician's enthusiasm is genuine, but it is one voice. The company-reported results come from earlier pilots. Technology cannot fully capture an experienced technician's sensory judgment. Even so, TKE's approach shows how connected data, Azure services and generative AI can work as a practical field tool. Teichmann's own view is that as long as coins fall and shoes are dirty, he will still have elevators to fix.