40th FVI Expert Breakfast

Topic: "AI in Maintenance – Hype or Real Added Value?" – Practice meets reality in spare part search, knowledge management, and personnel deployment planning.

The session was an honest reality check after several years of AI pilot projects. The guest reported from the semiconductor industry and presented three application fields that have already been tested under real production conditions and partially rolled out internationally.

• Spare part recognition measurably saves search time: In a plant, around 15,000 spare parts were photographed. Maintenance personnel can capture an unknown component with a smartphone and receive suitable stock hits within seconds. The reported hit rate was around 99.5 percent. New colleagues particularly benefit from this, but experienced maintenance personnel also adopted the solution once the personal benefit became apparent.

• The virtual shift colleague makes existing knowledge usable: A Large Language Model was connected with SAP long texts, wikis, Excel files, manufacturer PDFs, and scanned paper manuals. In case of a malfunction, the maintenance personnel can, for example, ask for a circuit diagram or previous solutions and receive relevant references within a short time. During the night shift, this does not replace the expert but allows access to their documented knowledge. As a result, machines could sometimes be restarted in the same shift, saving about 20 percent of the previous search time.

• Acceptance does not arise from frontal training, but through co-creation: The first development attempts provided only limited usable answers. The breakthrough came when maintenance personnel, together with software developers, aligned the system with their language, questions, and work methods. In a pilot team, individual colleagues were built up as multipliers. It was crucial not only to train people in prompting but also to adapt the system to the people.

• AI needs processes, data, and human approvals: The discussion showed clear boundaries. A recognized component is not automatically ordered unchecked; disposition, purchasing, and approval rules remain relevant. A flexible personnel deployment planning also only works if planned activities predominate, qualifications are known, and a responsible person gives the final touch. AI can provide recommendations and reduce complexity – responsibility and plausibility checks remain with humans.

Classification: With ADAM, distributed AI functions become operational intelligence in the daily maintenance process

The episode fits perfectly with the Hahn-PRO strategy because it shows: The greatest benefit does not come from an isolated AI model, but from the connection of knowledge, data, language, and concrete shop floor processes.

• Spare parts disappear in the brownfield data chaos: Outdated material numbers, different manufacturer designations, second sources, and decades-old equipment make the classic search slow and uncertain. With ADAM, we do not assume that a company must first clean up all its master data. ADAM is brownfield-native and taps into existing information through hybrid search and vector search. Different spellings, old reports, documents, and spare part histories become findable in the plant context. The digital life cycle record also notes which part was used when, where, and for what reason.
• Experience knowledge is available but not accessible at the moment of the malfunction: In the episode, knowledge was distributed in heads, PDFs, SAP texts, wikis, and local files. Especially at night or in other locations, direct access to the right expert was missing. We make ADAM the digital colleague of maintenance. Reporting, solution, and feedback assistants connect existing knowledge with the specific plant and the current malfunction. Technicians receive relevant document excerpts, previous solutions, and experience knowledge directly in the work context – like a "YouTube for maintenance," but plant-centered and searchable.
• Poor feedback prevents learning and automation: Under time pressure, malfunctions are often not documented at all or only with abbreviations. As a result, the foundations for root cause analysis, skill planning, and intelligent recommendations are missing later. With the Captain Kirk approach, ADAM lets the maintenance personnel simply speak into their smartphone. The assistant asks for missing information about the damage pattern, cause, and measure in a dialogue and generates a structured feedback from it. Through SaaS integrations and no-code workflows, this information flows back into existing systems like SAP or a CMMS. Critical decisions remain consciously human-in-the-loop.

Conclusion: AI creates real added value in maintenance when it is not thought of as an autonomous replacement for humans, but as an operational amplifier – and ADAM connects data chaos, experience knowledge, and daily work into usable operational intelligence for exactly this purpose.