Technicians find the fix in your manuals and service history, on your own servers

Technicians look up the fix in your own documentation instead of searching folders or guessing. Maintenance teams lose valuable time every day searching, documenting and asking colleagues, as Victor Splittgerber of WAKU Robotics put it in 2026 (silicon-saxony.de). That is the time this assistant is built to give back.

Machine builders already see service as a main field for AI. The German machine tool builders' association VDW names chatbots in service for maintenance, fault diagnosis and machine status, and VDMA names predictive maintenance and knowledge management as main areas, both as reported by produktion.de in February 2026. In the VDMA survey reported there, 43% of machinery companies use AI productively and 21% plan to; the article does not state the sample.

We measure the effect on your own jobs, before and after: time to an answer, calls to experts, reports that need rework and first-time fixes.

• Less time per job spent searching manuals, old reports and parts lists.
• Fewer calls to the few experts, who get time back for the hard cases.
• The knowledge of experienced technicians kept in the service history before they retire.
• Complete service reports at the end of each job, so the job can be invoiced and the next technician sees what was done.
• New technicians and service partners working from the same knowledge as your best people.

The assistant uses retrieval, often called RAG: answers drawn from your own documents, with the source shown. It covers the questions a technician has before, during and after a job.

Vendors and the press report similar systems. WAKU Robotics from Radeberg describes a maintenance agent with voice-guided repair, checklists and automatic documentation, built on maintenance histories, technical documentation and SOPs, earlier tickets, asset data and IoT data, without published measured outcomes (silicon-saxony.de, March 2026). For its Maintenance Copilot, Siemens reports an average saving of 25% of reactive maintenance time in first pilot use cases, a vendor claim without sample or method (ingenieur.de, March 2025). SMS group runs its own SMS-GPT and prioritises its AI work by what its employees ask for (VDMA study with Strategy&, 2025). Cloudsail built none of these.

Sorting incoming service tickets is a proven step too: the BTC example on our IT service desk page shows ticket classification by work group and priority working in production, even before today's language models.

• Fault finding: likely causes and steps for a fault code or symptom, from manuals, service bulletins and earlier reports on the same machine type, with the page cited.
• Machine history: what was done on this serial number before, which parts were replaced and which settings were changed.
• Spare parts: the right part number from the parts list or exploded drawing, matched with the SAP material master and, if you allow it, current stock.
• Procedures and checklists: maintenance plans and work instructions step by step, with the manual's safety notes shown as written.
• Service reports: drafted in your template from the technician's notes or voice memo, with work done and parts used, for the technician to confirm.
• Hotline and second level: the same knowledge helps the service desk answer customer calls and prepare a visit with the right parts.

The same build works on the shop floor. Work instructions, setup sheets, deviation reports, 8D reports (structured problem-solving reports common with automotive customers), shift handovers and quality records hold the answers to questions such as how a fault was solved last time or which parameter change caused a deviation. An assistant over these documents answers with the source, drafts the shift handover from the shift log and helps write the deviation report in your format.

Demand is there: in Bitkom's 2026 survey of companies with 20 or more employees, 22% of the companies that use AI use it in production processes. VDW also names chatbots integrated into CNC controls for programming, parameters and diagnostics (produktion.de, 2026). We found no public German example of language models used on shift logs or 8D reports, so we start by testing on your own past reports before anyone relies on the answers.

Predicting failures from sensor data and inspecting parts with cameras are specialist fields with their own sensors, data pipelines and models. We do not build them, and we say so. If you need them, work with a specialist provider or your machine supplier. According to KfW's 2026 SME analysis, 5% of SMEs use image recognition, which is used for quality control.

Where the two meet, we connect them. When a partner's model raises an alert for a machine, the technician assistant brings up the matching procedure, the machine's history and the parts likely needed, so the alert becomes a prepared job instead of a red light on a dashboard.

The technician decides what to do on the machine. A wrong repair instruction can hurt people, so the assistant is built to stay inside your approved procedures and to say so when it cannot answer.

• Answers come only from released documents, with the page or report cited, so the technician can check the source.
• Safety notes, isolation steps and warnings are shown as written in the manual, never rewritten.
• If the documents do not cover the case, the assistant says so and names the expert or team to ask, instead of guessing.
• The assistant never changes machine parameters or controls. It reads documents and systems and does not act on the machine.
• Service reports and parts orders are saved only after the technician confirms them.
• Tips from technicians become searchable only after an expert has reviewed them.
• Statements to customers on warranty, liability or root cause stay with the service manager.

We connect the sources your service team already uses: PDF manuals, service bulletins and work instructions in SharePoint, file shares or a document management system such as d.velop or ELO, plus service reports, notifications and orders in SAP PM or your maintenance system. Tickets come from ServiceNow, Jira or Salesforce, and parts data from the SAP material master.

Documents are indexed so that tables, fault-code lists and parts lists stay intact, and superseded manual versions drop out. Manuals in German and English are searched together. Live data such as stock levels or a machine's order history is read directly through approved tools with their own permissions.

Permissions follow the source systems and your service structure: an internal technician sees what their role allows, a service partner sees only the machines and customers they serve, and one customer's reports never appear in an answer about another customer.

Drawings, manuals and service histories are among a machine builder's most guarded data. The system therefore runs on your own servers with an open-weight model, a model whose weights you can run yourself, or in your own cloud account. Technicians reach it through your company network or VPN on a laptop, tablet or phone. All model calls go through one LLM gateway, a controlled access point with keys, budgets and logs in your own storage.

Build or buy: if you already run Siemens' copilots, your machine supplier's service portal or the service functions in SAP, use them where they fit. A custom build is justified when your knowledge sits in your own manuals and service reports spread over several systems, must stay on your own servers, or has to serve service partners with separate permissions.

Before go-live, the assistant passes an acceptance test built from real jobs: questions your technicians asked, with the correct answer and source set by your experts. We measure how often the answer and the cited source are both right, how often the assistant correctly says the documents do not cover a case, and whether safety notes appear where they must. The pass mark is agreed in writing before the test.

Some cases must never fail: an answer that skips a safety step counts as a failure whatever else it gets right. After go-live, technicians mark wrong or missing answers in the app, experts review a sample, and the test set is rerun after every document update, model change or new source. Your team gets the runbook and can operate the system or have us run it.

A technician assistant answers from documents and is not part of a machine's control. If AI becomes part of a machine you sell, the Machinery Regulation route applies: after the Digital Omnibus, in force since 27 July 2026, AI embedded in products under the Machinery Regulation is taken out of the AI Act's direct high-risk rules (Lewis Silkin, 2026). The AI Act's AI literacy duty (Article 4) still asks you to take measures so staff can use AI competently, so technicians learn how to check the cited source. How your systems are classified you decide with your lawyers. This page is not legal advice.

Every question and report is logged, and those logs could show how each technician works. Under § 87 (1) no. 6 BetrVG, the works council has a say when a system is objectively capable of monitoring performance or behaviour, as Bitkom's 2026 guide on AI and co-determination explains. We design the logging with pseudonymised user IDs, agreed retention and no per-technician performance reports unless a works agreement allows them, and we document it so the works council can review it.

Service reports contain customer contact data and sometimes the names of the customer's staff. Processing stays on your servers or in your cloud account, access follows the permissions above, and your data protection officer decides on retention and deletion.

Start with one machine family or product line that has good manuals, many service reports and a few experts who are always being called. In the free 30-minute first call we look at the documents and systems for that line and how your technicians search today. If it is worth doing, the first step is a test set from real questions and a baseline, with a fixed price for each step agreed in writing after the call.

• Which machine family, and roughly how many service jobs it generates.
• Where manuals, service reports and parts data are stored today.
• Who your go-to experts are and what they are asked most.
• Whether service partners or customers should use the assistant too.