AI Agents for Machinery Manufacturers: Where They Pay Off First
Companies that build machines to order, one at a time or in small batches, compete on know-how: which customer requirement is risky, which existing design can be adapted, which part went into a specific machine a decade ago. That know-how sits in customer specifications, bills of materials, service reports and the memories of a few senior engineers. AI agents work on exactly this layer. They do not control the machine; they support the office and service work around it.
What are AI agents in machinery manufacturing used for?
An AI agent is software built on a large language model that completes a task in several steps: it reads documents, looks things up in ERP or PDM, compares results and hands a proposal to a person for approval. In machine building, agents are used mainly in quoting, design, spare parts and field service.
What separates an AI agent from a chatbot is that it uses tools. It opens the bill of materials for a serial number, searches the archive for similar orders and drafts a list of open questions. A person still makes the call.
The German engineering association VDMA gives a sense of where the industry stands. In its survey of German machinery and plant manufacturers (February and March 2025, 206 responses), around 43% of companies reported having AI solutions in use, and 61% said AI agents were among the trend topics they were already working on. Customer service, field service and commissioning were rated as an area where AI matters a lot by 26% of respondents.
Why machine builders are a special case
Three traits shape where agents make sense:
- Engineer-to-order and small batches. Almost every order has its own configuration. Knowledge from past projects is valuable but hard to find, because it was rarely stored in a structured way.
- High product variety. Engineering works with modular platforms, options and one-off solutions. Answering "have we built something like this before?" often takes longer than the design change itself.
- Long service lives. Machines often run for decades. The aftermarket depends on finding the right documentation, parts list and change status for one specific serial number, many years after delivery.
This mix favours agents. There is a lot of text, many sources and recurring decisions, but hardly two identical cases. Fixed rules break quickly in this setting; an agent that can search the history copes better.
How does an AI agent for machine builders work?
An agent combines a language model, connectors to your business systems, a knowledge base built from your documents and explicit rules on what it may do alone and what needs sign-off. For machine builders, the critical piece is the link between serial number, order and the as-built parts list.
The building blocks:
- Language model. Reads specifications, emails and reports, drafts answers and decides which tool to call next. Whether a hosted model or a self-hosted one fits depends on your rules for export control and protecting design data; see sovereign AI for the trade-offs.
- Tools. Read access to ERP, PDM or PLM, document management and the service ticket system. Write access only where mistakes are cheap, such as creating a draft. The underlying mechanism is called tool use or function calling.
- Knowledge base. Operating manuals, wiring documentation in text form, service reports, previous quotes. The agent searches by meaning rather than keywords, using retrieval-augmented generation.
- Linking layer. Serial number to order, order to the bill of materials as delivered, and on to every service report for that unit. Without this chain the agent finds documents, just not the right ones.
- Approval and audit trail. Every proposal cites its sources, every action is logged, and critical steps need a human in the loop.
A concrete flow: a customer emails a photo of a worn part and a picture of the nameplate. The agent reads the serial number, retrieves the as-built bill of materials, matches the photo against part descriptions, checks the ERP for a superseding part and prepares a reply for the parts desk with part number, stock status and links to the sources it used.
Which use cases pay off first?
Start where heavy reading meets repeated decisions and errors are caught before anything reaches the customer: reviewing specifications for quotes, finding reusable designs, identifying spare parts and supporting service diagnosis. Technical documentation usually comes next.
Specification review and quoting
Customer specifications for special-purpose machines can run to dozens of pages, full of references to standards, special requests and hidden obligations. An agent can check each one against your modular platform and past quotes and sort requirements into standard, option, new development and missing information. The output is a deviation and clarification list for the sales engineer. The general pattern is covered in our guide to AI quoting agents for RFQs, specifications and cost estimation, and the quotation agent overview shows the scope. What differs for machine builders: standards, safety requirements and interfaces to the customer's line matter more than price and lead time.
Configure-to-order and variant design
In design, the agent does not draw; it finds. A question like "which orders had a lift height above a given value and a stainless frame?" usually needs answers pulled from PDM, ERP and old project folders. An agent with access to parts lists, configuration data and order descriptions returns a shortlist with reasons. The engineer decides whether an existing design can be adapted.
Spare parts
The aftermarket hinges on one question: which part is this, exactly? Customers send photos, half-legible nameplates or numbers from an old manual. An agent connects serial number, as-built parts list and change history, proposes the matching item and warns when a part has been discontinued or superseded.
Field service and remote diagnosis
Experienced technicians recognise fault patterns because they have seen them before. A service assistant makes that experience available to everyone: it searches service reports, fault logs and manuals for similar cases and suggests what to check next. The technical service agent page describes the tasks such an agent can take on.
Technical documentation
Manuals, risk assessments and technical files are laborious and vary by configuration. Agents can assemble drafts from approved text modules, flag contradictions between parts list and manual and prepare translations. The new EU Machinery Regulation makes this more relevant (see the regulation section below).
Keeping knowledge when people retire
When long-serving staff leave, knowledge about older machine generations leaves with them. An agent that indexes service reports, design notes and email threads preserves part of it. More on this under knowledge management agents.
What data and systems do the agents need?
Above all, linked data: bills of materials as delivered, serial-number history, service reports, manuals and previous quotes. Volume matters less than the chain from a serial number to the correct document version. Without that chain, the agent produces answers that sound right and are wrong.
Typical sources and how they tend to look in practice:
| Source | What the agent uses it for | Common gap |
|---|---|---|
| ERP (orders, items, serial numbers) | Linking machine to order to parts | Older orders only partly migrated |
| PDM/PLM (BOMs, drawings, changes) | Correct revision for each serial number | As-built state not frozen at delivery |
| Document archive (manuals, specs) | Knowledge base for search and comparison | Scanned PDFs without text recognition |
| Service tickets and reports | Fault patterns, fixes, parts demand | Free text, inconsistent terms |
| Machine data (alarms, states) | Context for diagnosis | Often only available on the customer's site |
Two issues are regularly underestimated:
- Revision control. An agent that reads today's parts list instead of the as-built one will propose the wrong spare part. Check whether your PDM can return the delivered state per serial number.
- Scanned legacy documents. Much of the paperwork for older machines exists only as image PDFs. Without text recognition, the agent cannot see it.
The EU Data Act changes access to machine data
For connected machines sold in the EU, the Data Act has applied since 12 September 2025. Users of connected products are entitled to readily available product data, including the metadata needed to interpret it, free of charge and in a machine-readable format (Article 4). The obligation to design products so that such data is directly accessible by default (Article 3(1)) applies, under Article 50, to products placed on the market after 12 September 2026. For service agents this means data access has to be defined anyway, so plan agent use into the same work. This is orientation, not legal advice; outside the EU, check your local rules.
Regulation: what the EU Machinery Regulation means for AI
The EU Machinery Regulation 2023/1230 applies from 14 January 2027 (Article 54) and replaces the Machinery Directive. Two points matter for office and service agents: digital instructions, and stricter rules for safety functions that rely on machine learning.
Digital instructions. Under Article 10(7), instructions for use may be supplied in digital form. The manufacturer must then keep them online for the expected lifetime of the machine and at least ten years after it is placed on the market, and, if the user asks at the time of purchase, supply a paper copy free of charge within one month. Versioned, machine-specific documentation becomes an obligation, which also makes it a solid data source for service and spare-parts agents.
Machine learning in safety functions. Annex I lists safety components, and machines with embedded systems, that have fully or partially self-evolving behaviour using machine learning approaches and ensure safety functions. These fall into the higher-risk category that requires third-party conformity assessment. An agent that reviews specifications or searches service reports is not affected as long as it performs no safety function on the machine. Put that boundary in writing for every project.
Technical files and the EU declaration of conformity must also be kept for at least ten years. Agents can help compile them, but responsibility for the content remains with the manufacturer. If you sell into markets outside the EU, their machinery rules apply in addition. None of this is legal advice.
How to go from pilot to production
Pick one narrowly defined process, measure one metric before and after the pilot, and only then expand. A sequence that works: choose the process, test the data, build a pilot with human sign-off, measure, then hand it over to day-to-day operations.
- Choose the process. Look for high frequency, lots of reading, a clear approver and errors that are caught before the customer sees them. Specification review and spare-part requests usually qualify.
- Test the data. Take 20 to 30 real cases and check whether a person could find the answer with the systems you have. If not, the problem is the data, not the model.
- Define the metric. For example, time to the clarification list, share of correctly identified parts or number of questions escalated to engineering. Measure the baseline first.
- Pilot with sign-off. The agent drafts, a person approves. Every correction is recorded and fed back into rules and knowledge base.
- Operate. Settle permissions, logging, ownership of the knowledge base and who handles errors. Then move to the next process.
In our own operations, step 2 has consistently taken the most time and told us the most: if the data test fails, a stronger model will not rescue the use case.
A common order of rollout
Many machine builders start with an internal process such as specification review or design search, because mistakes stay in-house. Spare-parts and service agents follow once the link between serial number and parts list is reliable. Customer-facing agents that answer directly come last, usually still with approval.
Limits and risks
Agents can produce plausible but false details, stumble over missing revision data and must never make safety decisions on their own. Protecting design know-how, export control and acceptance by the team add further risk. All of these can be reduced; none disappears completely.
- Hallucination. A model can invent a part number that looks real. Countermeasure: the agent may only output numbers it has retrieved from the ERP through a tool call, with the source attached.
- Wrong revision. See the data section; without the as-built state, spare-part proposals are unreliable.
- Know-how protection and export control. Design data and specifications contain trade secrets and sometimes export-controlled information. Decide upfront which data each model may see and whether you need a self-hosted model.
- Safety functions. Agents do not belong in the safety controller. Once a system with learning behaviour performs safety functions, the stricter conformity assessment of the Machinery Regulation applies.
- Acceptance. Senior staff may see an agent as a threat. It works better to involve them as reviewers and knowledge owners; their corrections are what make the agent useful.
- When this approach does not fit. Very few, very different cases per year; data that exists only on paper; or nobody with time to review the agent's output.
How do I get started?
Begin with an inventory: which three processes take up most of your senior people's time today, and is the data for them digital? Pick one, test it against real cases and build a pilot with human approval.
A practical first week might look like this:
- Days 1–2: List candidate processes and rate them by frequency, share of reading work and data quality.
- Day 3: Pull 20 real cases from the chosen process and note which systems a person needs to resolve each one.
- Day 4: Clarify access: interfaces, read permissions, data protection and protection of design data.
- Day 5: Fix the metric and measure the current value.
At the end of that week you will know whether a pilot makes sense and which data needs cleaning first. For an overview of agents by department, see AI agents for mid-sized companies. In-depth articles on specification review, spare parts, field service and the Machinery Regulation will follow in this section, and the guides hub collects the basics.
Frequently asked questions
Do we need a data platform before we can use AI agents?
No. A first agent only needs read access to your ERP, PDM or PLM system and your document archive. A shared platform becomes worthwhile once several agents rely on the same data.
Should an agent make engineering or commercial decisions on its own?
Not for quotes, part releases or fault diagnoses. The agent prepares the case, cites its sources and flags uncertainty; an engineer signs off.
Is a service assistant covered by the EU Machinery Regulation?
An assistant that reads documents and does not perform safety functions on the machine is not a safety component. Have the classification of your specific system checked by a legal adviser.
What is a sensible first metric for a pilot?
Pick one that you can measure before and after, such as time from receiving a specification to sending the clarification list, or the share of spare-part requests answered with the correct part number at first attempt.
Sources
- Regulation (EU) 2023/1230 on machinery (Official Journal L 165, 29 June 2023) publications.europa.eu
- Regulation (EU) 2023/2854, the Data Act (Official Journal, 22 December 2023) publications.europa.eu
- VDMA Software and Digitalisation: survey results on AI and machine learning in mechanical and plant engineering (10/2025, German) vdma.eu