AI that prepares quotes, service answers and documentation for machine builders and manufacturers, while your data stays in the company

Most of the change happens in the office and in service. Quotes go out faster, orders reach the ERP without retyping, and technicians find the fix without calling the one colleague who knows. These processes run on expert knowledge that sits in past quotes, manuals, service reports and people's heads, and the skills shortage makes that knowledge harder to replace every year.

Much of the industry already uses AI. In the ifo business survey of May 2025, about 50% of mechanical engineering companies used AI, and manufacturing overall was above 40%. A VDMA survey reported by produktion.de in 2026 found 43% of machine builders using AI productively and 21% planning to. The gap is getting from trial to daily use: in the 2025 study by VDMA and Strategy& of 247 industrial companies in the DACH region, 79% use or plan generative AI, 29% have implemented at least one project, and only 7% have rolled it out systematically.

The same study puts numbers on the pressure. The gross result of machinery companies fell from 32.2% in 2018 to 29.4% in 2023, and sales and marketing carry the largest modelled margin potential of generative AI, at 2.4 percentage points. 53% of respondents expect generative AI to have an impact in sales and marketing. Florian Stürmer of Strategy& draws the lesson in the study: success with generative AI means not trying everything, but putting the right thing into practice quickly.

Look for processes where documents and expert knowledge feed a step that repeats. Language models handle that kind of work reliably. Each one has its own use-case page, linked below.

• Quotes and order entry. An assistant finds similar past quotes, reads current prices and parts from the ERP and drafts the quote for your sales engineer to check. Orders that arrive as an email with a PDF are read and prepared as sales orders in SAP or your ERP. The food manufacturer Zentis reported in the press in 2023 (Computerwoche) that, with a specialist product, about 80% of its order documents were processed automatically.
• Service knowledge for technicians and after-sales. Technicians and hotline staff ask in their own words and get an answer from manuals, service reports, tickets and parts lists, with the page it came from. Service reports are drafted from the technician's notes. The maintenance software maker WAKU Robotics described the problem in 2026: maintenance teams lose valuable time every day searching, documenting and asking colleagues.
• Technical documentation. Drafts of operating manuals, spare parts lists and variant documentation from product specifications and earlier versions, released by your technical writers. The VDMA and Strategy& study (2025) lists technical documentation from product specifications among the generative AI use cases for machinery.
• Company knowledge. One place to ask about internal standards, design guidelines, past projects and supplier terms, answered from your own documents with the permissions of the person asking. According to the VDMA study, SMS group runs its own "SMS-GPT" and prioritises its AI work by the requests of its employees.
• Incoming supplier documents. Order confirmations, delivery notes, certificates and test reports read and matched against the purchase order in the ERP, with deviations passed to a person.

The automotive industry has one of the highest AI usage rates: 70.4% in the ifo survey of May 2025. For suppliers under price pressure, the same commercial processes matter as in machine building, only with more volume: requests for quotes with long specifications, cost breakdowns, change requests and supplier correspondence.

What differs is the data rules. Contracts with carmakers often set strict conditions for drawings, specifications and prototype data. We place the system where those conditions allow, usually on your own servers, and your information security team decides whether it falls inside an existing assessment such as TISAX. The price itself comes from your calculation in the ERP, not from a language model: the assistant gathers the inputs and explains the differences to earlier quotes.

About every second chemical company uses AI, according to the ifo survey of May 2025. Chemicals, pharma and medical technology have their own industry page, which covers GMP, the MDR, the AI Act and how validation works. The document use cases on this page apply there too.

Some of this is already in standard products, and we say so. SAP S/4HANA Cloud can create sales order requests from a PDF or an image, suggest field values and simulate price or item deviations before conversion (PIKON, 2025), so if you run it, try that first. Siemens sells its Industrial Copilot for PLC programming and a Maintenance Copilot for maintenance teams.

Predictive maintenance on sensor data and visual inspection with cameras are specialist territory. They need sensor and image models, plant data and often hardware, and specialist vendors and your automation partners do them better. We do not build them. We build the language and document side around them: the assistant that explains an alarm from the manual, the drafted service report, the answers from your quality procedures.

A custom build pays off when the standard feature does not fit: several ERPs or an older SAP release, quotes that need engineering judgement and past projects, or documents that must not leave your own servers.

Design and pricing data are the core of a machine builder's value. Drawings, calculations and customer prices should not go to a public AI service, and often not to any outside one, which is why we build on your own servers when the data demands it. In a survey of 106 IT leaders of small and mid-sized companies in the DACH region by CIO and Computerwoche with Dell (2026), 20% had started building local server infrastructure for AI. The survey is vendor-sponsored and small, but it shows the direction.

The AI Act. Assistants for quotes, orders, service knowledge and documentation are not among the AI Act's high-risk uses, as long as they do not control the machine or evaluate staff. Evaluating employees' performance with AI is a high-risk use under Annex III and brings duties such as human oversight for the company using it. We do not build it. After the Digital Omnibus, in force since 27 July 2026, AI embedded in machinery is taken out of the AI Act's direct high-risk rules and handled under the Machinery Regulation, and obligations for AI in regulated products move to 2 August 2028 (Lewis Silkin, 2026). From 2 August 2026, people must be told when they interact with an AI system. That matters for any assistant that answers your customers directly. The duty to take AI literacy measures for staff remains.

Safety. Repair instructions and technical documentation are safety-relevant: a wrong torque value or a skipped lockout step can hurt someone. Answers therefore cite the manual page they come from, stay inside approved procedures, and say so when the documents do not contain the answer. Documentation drafts are released by your technical writers, never published by the system.

Works council. An assistant that logs who asked what can in principle monitor performance, which makes it subject to co-determination under § 87 (1) no. 6 BetrVG. We design the logging so it serves quality checks without tracking individuals: pseudonymised IDs, agreed retention and clear access rules, ready for the works agreement. Export-controlled drawings and technical data follow your export control rules and are excluded from the index or restricted to the people cleared for them. That is the technical side. You decide, advised by your data protection officer and lawyers.

We build where your data may go. On your own servers, an open-weight model, a model whose weights you run yourself, answers without any data leaving the building. In your own cloud account, the models run in an EU region under your contracts. Either way, an LLM gateway, one controlled access point to all models with keys, budgets and logs, sits in front of every model, and the logs stay in your systems.

• ERP: prices, parts, customers and orders from SAP or your ERP, read live through approved tools, without a copy in an index.
• Documents: past quotes, manuals, service reports and standards from SharePoint, file shares, d.velop or ELO, indexed with the permissions they already have.
• Engineering data: specifications and parts lists from PLM or PDM where the use case needs them, restricted to the people who may see them.
• Service and mail: tickets and service reports from your service system, and Outlook mailboxes for incoming orders and enquiries.
• Writing back: an agent, a program that carries out steps in your systems and presents the results for approval, creates a quote or order draft only through a tool that does one business action, with approval and a log entry.

Before go-live, we measure the system on your own past cases. For order entry, that is a set of real past orders with the correct ERP entries, compared field by field. For quotes, past enquiries with the quote your team sent at the time. For service, real technician questions with the right answer and the manual page it comes from.

You set the acceptance threshold with us, for example the share of fields that must be right before an order may enter the approval queue, and the acceptance test shows whether it is met. We also test that the system holds back what a person may not see, such as prices for a service technician or drawings for an external partner.

Before it takes over any step, it can run in shadow mode: it prepares quotes or orders alongside your team without sending anything, and we compare. We measure hours per quote or per order before and after, on your own cases, so management sees the effect in its own numbers. After go-live the same tests run after every change to models, prompts or sources.

Start with one process that costs your experts time every week and has clear past examples: quotes for standard machines or spare parts, incoming orders by email, or technicians' questions to the hotline. Describe it in a free 30-minute first call with an engineer. We need no access to your systems or data for that.

We agree the first step and its acceptance threshold with the person who owns the process, such as your head of sales or service. Each step has a fixed price, agreed in writing after the first call. Where design and sales teams already use ChatGPT or Copilot, the free AI usage review puts on one page what those licences cost and where they fall short, from exports you choose to send.