AI in Maintenance: Agents for Breakdowns, Servicing and Spare Parts
In most maintenance departments, the knowledge that matters lives in three places: equipment manuals, the breakdown history in the maintenance system, and the heads of experienced technicians. When a line stops on the night shift, only part of that is within reach. AI agents are built to close that gap: they make existing knowledge easier to find and take over preparatory work that is still done by hand.
What does AI in maintenance actually mean?
AI in maintenance covers two quite different approaches. Predictive maintenance analyses sensor time series to spot developing failures. Knowledge-based agents work with documents, breakdown reports and work order data. The first tells you that something is about to fail; the second helps you decide what to do about it.
A useful frame comes from the German standard DIN 31051 (current edition June 2019), whose terminology is aligned with the European standard EN 13306. It splits maintenance into four basic activities: servicing, inspection, repair and improvement. Each of them produces and consumes knowledge, and each offers room for an agent:
- Repair: describe the symptom, find likely causes, show the relevant manual pages and wiring diagrams, and pull up similar past incidents.
- Servicing and inspection: turn manufacturer intervals into draft maintenance plans, prepare checklists, and capture inspection findings in a structured way.
- Improvement: spot recurring failures across machines and shifts, so weak points become visible.
- Spare parts: identify a part from a description or bill of materials, check stock and alternatives, and draft a purchase requisition.
This is different from a plain chatbot. An agent does not just answer questions; it uses tools. It reads from the CMMS (computerised maintenance management system), queries inventory and creates drafts. For a broader view of what agents can do in a department and how such systems are put together, see the overview of AI agents for business functions.
How does a maintenance agent work?
A maintenance agent combines a language model, a searchable knowledge base and connectors to the CMMS and inventory system. The model understands the question, retrieval finds the right passages and past cases, the connectors fetch live data, and every answer points back to its source.
The building blocks:
- Language model. It understands free text such as "Press 3 throws error 47 after a tool change, hydraulic pressure fluctuating" and writes the reply. Whether a cloud model or a self-hosted one fits depends on your data protection needs and infrastructure; the trade-offs are covered under sovereign AI.
- Knowledge base with retrieval. Manuals, servicing instructions, wiring diagrams and incident reports are split into chunks and indexed for search. The technique is called retrieval-augmented generation, described in 2020 by Lewis et al., and pairs a language model with a search over external documents. For maintenance, that means answers come from your own documentation rather than general knowledge. The guide to enterprise RAG architecture explains how to build such a knowledge base.
- Tools. Through tool use, the agent calls defined functions such as "read all breakdown records for asset X from the last 24 months", "check stock for part number Y" or "create a draft maintenance notification".
- Rules and approvals. What the agent may read, suggest and write is fixed in advance. For safety-critical steps such as isolation, lockout/tagout or work on guarding, it never issues instructions of its own; it points to the applicable procedure.
- Audit trail. Every query, every retrieved source and every suggestion is logged, so you can trace why the agent recommended something.
A typical breakdown flow: the technician describes the symptom on a tablet. The agent looks up the error code in the manual, finds three earlier reports with similar wording, shows the cause and fix recorded at the time, and lists the parts involved with their bin locations. The technician decides and does the work; the agent turns a short voice note into a clean completion text.
Where can agents help in maintenance?
Agents can help with troubleshooting, maintenance planning, spare parts, documentation and weak-point analysis. Troubleshooting is almost always the right place to start: the benefit shows up in daily work straight away, and the documents you already have are usually enough to begin.
Troubleshooting
The agent searches manuals and breakdown history and proposes likely causes, each with a source. Its value grows with the quality of past records. Where earlier jobs were documented with cause and remedy, it finds patterns no single person remembers. Where the record just says "fixed", it has only the manual to go on.
Maintenance planning
From manufacturer data on intervals, lubricants and checkpoints, the agent drafts servicing plans and work instructions. It can also flag assets that have no plan at all, or match next week's planned jobs against spare part availability. Approving the plans stays with the maintenance manager.
Spare parts and stores
On the shop floor, parts are rarely called what the material master calls them. An agent translates between a description ("the small proximity sensor on the left of the gripper"), the bill of materials, the manufacturer's part number and the internal material number, checks stock and suggests a requisition if needed.
Documentation and shift handover
Writing completion notes is unpopular, so they are often thin. An agent can turn a quick voice memo into a structured record with symptom, cause and action taken, which in turn improves the data for future troubleshooting. At shift change, it summarises open breakdowns and unfinished jobs.
Weak-point analysis
Across many records, the agent can spot recurring causes, for example the same assembly failing on different machines. That is the "improvement" activity from DIN 31051 in practice: instead of repairing the same thing again, you see where a design change pays off.
How this relates to predictive maintenance
Predictive maintenance works on sensor data such as vibration, temperature or motor current and detects deviations from normal behaviour. A knowledge-based agent works on text and work order data. They fit together: the prediction flags an abnormal bearing on pump 2, and the agent supplies the replacement procedure, the part number and past experience. Industrial vendors now combine both. Siemens presented a generative AI maintenance offering in March 2025 and, according to its press release, first pilot use cases saved an average of 25% of reactive maintenance time. That figure is a vendor claim from pilots, not an independent measurement.
A closely related use case is field service, where equipment makers support their own service engineers at customer sites. That is covered on the page about the technical service agent.
What do you need before you start?
You do not need perfect data, but you do need three things: accessible technical documentation, a breakdown history with usable text, and read access to the maintenance system. Add one accountable person from maintenance who can judge whether the answers are right.
Data
- Manuals and instructions: ideally as searchable PDFs, not image-only scans. Older scanned documents can be run through OCR, but that costs time and some accuracy.
- Wiring diagrams and drawings: valuable, but harder for language models to interpret than text. At first it is often enough for the agent to find and display the right page.
- Breakdown history: notifications and work orders from recent years, with symptom, cause and action. The free text is what matters; codes alone say little.
- Master data: functional locations, equipment, bills of materials, material master.
Systems
In SAP environments, corrective maintenance typically runs through notification, order, confirmation and technical completion. According to SAP Learning, the notification describes the technical condition of an object, while the order is used to plan, execute and control the work. An agent plugs into exactly these objects: read-only for troubleshooting first, drafting later. Other CMMS products name these objects differently; depending on the system, the connection runs through an API or a data export.
People and organisation
- A subject-matter owner who writes test questions and reviews answers.
- Clear rules on who uses the agent and what it may do.
- Early involvement of employee representatives where they exist (see the risks section).
- A willingness to improve the quality of completion notes, because the system depends on them over time.
How do you introduce it step by step?
A five-step approach works well: pick a narrow use case, review the data, build a read-only prototype, test it against real incidents, and only then roll it out. Write access and further asset groups come last.
- Narrow the use case. One asset group with frequent breakdowns and decent documentation, and one clear question ("What can cause error X on machine type Y?").
- Review the data. Which manuals exist digitally? How good were the completion notes over the past two years? How will you get CMMS data: API or export?
- Build a read-only prototype. Set up the knowledge base and retrieval, and require a source for every answer. No write access yet.
- Test with real cases. Take twenty to thirty closed incidents with a known root cause and check whether the agent finds it. Experienced technicians grade every answer. This becomes your test set for measuring later changes.
- Put it where the work happens. A tablet at the machine, the control room, a phone. Collect user feedback and keep the knowledge base current. Then enable drafts for notifications, completion notes or requisitions step by step, always with human approval.
The test set from step 4 is the most important part of the process. Without it, you cannot tell whether a change to the model or the retrieval made things better or worse.
If you want to work through use case selection with your team, a structured AI workshop is a practical starting point.
What are the limits and risks?
An agent can be wrong, quote outdated documents or make a wrong cause sound convincing. In maintenance, safety is also at stake. The rule is simple: suggestions with sources, decisions by the technician, and no safety-critical instructions generated by the model.
Technical risks
- Hallucinations. Language models can make things up (see AI hallucination). Countermeasures: answer only from the knowledge base, always cite the source, and say "I cannot find anything on this" when the documents are silent.
- Outdated documentation. If a machine was modified but the manual was not, the agent will repeat the old state. Version information in the knowledge base and a visible document date help.
- Poor breakdown history. Records like "OK" or "replaced" carry no diagnosis. In that case the agent's value is initially limited to the manuals.
Safety
The agent does not replace a risk assessment or a safe system of work. For electrical, hydraulic or pneumatic work, it must refer to your isolation and lockout procedures rather than invent its own sequence.
Team acceptance
Experienced technicians are often sceptical, and with good reason if a system seems designed to extract their knowledge and replace them later. Acceptance grows when the agent takes over tedious tasks such as writing completion notes or hunting for parts, and visibly builds on their expertise. Involve them as reviewers of the test set and make their corrections visible.
Employee consultation
If the agent logs who asked what and when, it may be able to monitor behaviour or performance. In Germany, Section 87(1) No. 6 of the Works Constitution Act gives works councils a co-determination right over introducing such technical systems, and other jurisdictions have their own consultation and privacy rules. An agreement covering purpose, reporting and retention periods creates clarity. This is guidance, not legal advice.
When the approach does not fit
- When there is barely any documentation and knowledge is passed on only by word of mouth. Documentation work comes first.
- When machines are simple and breakdowns are fixed in minutes.
- When the real problem is missing parts or missing staff, not missing knowledge.
How do I get started?
Start with a quick assessment: which asset causes the most downtime, how good is its documentation, and how many breakdown records contain useful text? The answers define your first use case. A read-only troubleshooting prototype is usually the fastest route to solid evidence.
A short checklist for the decision:
- Is there an asset group where breakdowns regularly take longer because knowledge is missing or hard to find?
- Are the manuals available digitally and searchable?
- Do breakdown records from the past two years include cause and action in the text?
- Is there read access to the CMMS, via API or export?
- Is an experienced technician available to grade twenty to thirty test cases?
- Is it settled whether data may leave your premises, or whether a self-hosted model is required?
If most answers are yes, a prototype is worth building. How we design department-level agents and connect them to existing systems is described on the AI agents page; terms such as retrieval or tool use are explained in the glossary. Further guides on troubleshooting and on combining agents with predictive maintenance will appear in the guides section.
Last reviewed: November 2026.
Frequently asked questions
Is a maintenance agent the same as predictive maintenance?
No. Predictive maintenance analyses sensor data to flag that a failure is developing. A knowledge-based agent helps with what to do next by searching documentation and past breakdowns.
Do I need sensor data to get started?
Not necessarily. For troubleshooting, manuals, wiring diagrams and well-written breakdown records are often enough. Sensor data matters once you move towards condition-based or predictive strategies.
Should the agent create work orders on its own?
It can, but at the start it should only produce drafts that a technician or planner reviews. Grant write access later, narrowly scoped, once the system has proven reliable.
Do employee representatives need to be involved?
Often, yes. In Germany, for example, works councils have a co-determination right over systems capable of monitoring staff behaviour or performance; many other countries have their own consultation rules. This is not legal advice, so check your local requirements early.
Sources
- DIN 31051:2019-06 – Grundlagen der Instandhaltung (DIN Media) dinmedia.de
- Siemens: Industrial Copilot – generative-AI-powered maintenance offering (24 March 2025) press.siemens.com
- Lewis et al. (2020): Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks arxiv.org
- SAP Learning: Outlining the Standard Corrective Maintenance Process learning.sap.com
- German Works Constitution Act, Section 87 – co-determination rights gesetze-im-internet.de