AI in maintenance is software that takes on work which used to need a person to read, look or spot patterns: going through an inspection report for findings, recognising a part in a photo, noticing an unusual pattern in sensor data. Right now, artificial intelligence mainly helps maintenance teams with reading: reports, manuals, drawings and other documents. AI produces proposals, not decisions. The gain comes when a skilled person can check each proposal quickly, with the source alongside.
AI is not one thing: four types in maintenance
The label AI covers techniques that do very different things. Four of them matter for maintenance:
| Type of AI | What it does | Maintenance example |
|---|---|---|
| Machine learning on sensor data | Learns a machine's normal behaviour and flags deviations | A flue gas fan with an unusual vibration pattern |
| Language model | Reads text, extracts data, summarises and answers questions | Extracting findings from an inspection report |
| Image recognition | Recognises objects or defects in photos | Identifying a part that has no number |
| Text recognition (OCR) | Turns text in a scan or photo into searchable text | Making a scanned P&ID from 1998 searchable |
They often work together. Text recognition makes a scanned report readable; a language model then extracts the findings.
Not everything clever is AI
An alert when vibration exceeds a limit is a threshold. Recognising a tag number on a drawing by a fixed pattern is a rule. A trend line through running hours or wall thickness readings is statistics. None of that is AI, and it does not need to be: such methods are predictable, can be checked by hand and can be explained to an auditor. For usage-based maintenance, a projection based on measured running hours is often sufficient. AI comes into its own where rules and statistics fall short: unstructured text, images and patterns that nobody can describe in advance.
Where AI already helps in maintenance
The diagram below shows the applications by type of input, with the type of AI involved. Along the bottom is what applies to every one of them: AI makes a proposal with a source, a person reviews it, and only then does it become work.
Documents and reports
This is where most of the gain lies today, because maintenance runs on documents that nobody reads from cover to cover. A language model and text recognition work together here.
- Extracting findings from reports. After the shutdown of a waste-to-energy plant, there are dozens of service and inspection reports. AI can propose the findings in each report, with severity, page and the verbatim quote, so you can check them in context during review. How to follow up findings after that is covered in managing maintenance contractors.
- Tasks from a manual. Somewhere in a manual of several hundred pages are the periodic tasks, with intervals, lubricants and quantities. AI can extract them as a proposal for a maintenance plan. The manufacturer's advice remains a starting point; you set the intervals.
- Search by meaning. Ordinary search finds only the words you type. Search by meaning also finds the passage on tightening torque when you ask "how tight should the bearing cap be?" That only works if your documents are digital and searchable, and it works best if they are linked to the right assets; see technical document control. An example is shown under technical document management.
Photos
Image recognition, often combined with text recognition, turns a photo into a search query.
- Identifying a part. A technician at a distribution centre is standing at the racking with a sensor that has no label. Image recognition compares the photo with the items in your own catalogue and shows the candidates, while text recognition reads any type numbers. Results depend on the photos and descriptions of your items. An example from a maintenance stores setting is shown under search by photo.
- Reading a nameplate. Taking the manufacturer, type and serial number from a photo saves typing and typos when you populate an asset register.
- Defects in inspection images. Image recognition can mark corrosion, cracks or leaks in camera or drone footage. It needs many labelled example images and pays off mainly with large numbers of similar inspections.
Sensor data
Machine learning on sensor data demands more than the other applications: continuous measurements and, for predictions, failure history.
- Flagging abnormal behaviour. A model learns from vibration, temperature or current draw how a machine normally behaves, and flags when the pattern changes, even if no single value crosses a limit. That is a reason to go and look, not yet a diagnosis.
- Estimating remaining useful life. A model that learns when a machine fails needs examples of failures. Critical assets have few, precisely because you try hard to keep them from failing. Where you have a measurable wear parameter with a known limit, such as wall thickness, a trend line often gives a usable projection. There is more on this in predictive maintenance.
Questions and analyses
A language model can also answer questions and prepare analyses.
- Assistant for your own data. "Which pumps had more than three breakdowns this year?" or "Which document has the replacement procedure for the mechanical seal?" An assistant that answers such questions should show which work orders, assets or documents the answer rests on, and only show what the person asking is allowed to see.
- Proposing failure modes. In an FMECA, a language model can propose possible failure modes, causes and a maintenance strategy for each function. That works as a checklist during the session, so the team misses less. It does not replace what operators and technicians know about how your filling machine or compressor actually fails.
What you need to use AI well
- Your basic data in order. A finding about "pump 3, left" can only be linked to the right asset if your asset register has consistent codes and descriptions. Documents need to be digital and searchable.
- A person who reviews. Decide who checks the proposals, and give them the time to do it.
- A source with every proposal. A page and quote, or the records an answer rests on. Without a source you have to reread everything and the time saving is gone.
- Agreements about your data. Where is a document processed and stored, who can access it, and is it used to train models? Reports contain names, and drawings and contracts are confidential. For personal data, the GDPR applies.
- Users who understand the tool. Anyone working with AI needs to know what the tool does and where it can go wrong. The EU AI Act requires organisations to take measures for this.
- An owner. Someone who measures the results, collects problems and makes adjustments.
Limits and risks of AI in maintenance
- Made-up answers. A language model can state something false with complete confidence. In its risk profile for generative AI, the US National Institute of Standards and Technology (NIST) calls this confabulation, popularly known as hallucination, and notes that citations can be made up too. So a proposal can only be trusted once the quote is actually in the document. Check that, or use software that checks it.
- Misreads in scans and handwriting. Text recognition easily confuses 0 and O, or 1 and I. In a tag number, that means a finding ends up against the wrong asset, or none at all. Blurred scans and handwritten notes on drawings go wrong more often.
- What AI misses, you do not see. An incorrect proposal can be spotted during review; a missed finding cannot. On critical plant, always read the conclusions and rejections in a report yourself.
- Too much or too little trust. NIST also describes automation bias: excessive deference to automated systems. The reverse happens too: experts who distrust a tool from the outset and leave the time saving on the table. Measuring helps against both.
- Confidentiality. Staff who paste a report or drawing into a public chatbot are making a decision about company data that is not theirs to make. Provide an approved alternative and clear rules.
The EU AI Act in outline
The EU AI Act (Regulation (EU) 2024/1689) has been in force since 1 August 2024. It takes a risk-based approach: the greater the risk of an application, the stricter the rules. Which category applies depends on the intended use. The European Commission's overview sets out four levels:
- Prohibited practices, such as social scoring and emotion recognition in the workplace.
- High risk. For example, AI intended as a safety component in the management and operation of critical digital infrastructure, road traffic or the supply of water, gas, heating or electricity. This also covers AI that allocates tasks based on individual behaviour or personal traits, or that monitors and evaluates the performance and behaviour of workers. Both are listed in Annex III. AI as a safety component of products covered by EU product legislation, such as lifts and pressure equipment, can also be high risk. High-risk systems must meet requirements including risk management, logging, documentation and human oversight.
- Transparency obligations: people must know when they are talking to a chatbot, and AI-generated content must be recognisable.
- Minimal risk: the Act sets no requirements for the system itself, although the AI literacy obligation still applies. According to the Commission, the vast majority of AI systems currently used in the EU fall into this category.
Most obligations fall on the provider: the party that develops an AI system, or has one developed, and places it on the market or puts it into service under its own name. A maintenance department using an AI tool is a deployer. If you have a model built for your own use, for example for predictive maintenance, and put it into service under your own name, you may also count as a provider.
For every organisation using AI, the AI literacy obligation has applied since 2 February 2025 (Article 4). Since the July 2026 amendment, it amounts to this: providers and deployers take measures to support the AI literacy of their staff and of others operating AI systems on their behalf, without having to guarantee any particular level for each individual. In practice, that means explaining what the tool does and where it can go wrong.
| Date | What applies |
|---|---|
| 1 August 2024 | The Act enters into force |
| 2 February 2025 | Prohibited practices and AI literacy |
| 2 August 2025 | Rules for general-purpose AI models |
| 2 August 2026 | Most of the Act, including the transparency obligations |
| 2 December 2027 | High-risk AI listed in Annex III, such as critical infrastructure |
| 2 August 2028 | High-risk AI in products, such as lifts |
The high-risk dates were pushed back in July 2026 by an amendment to the Act, the Digital Omnibus on AI (Regulation (EU) 2026/1744). For a maintenance department it usually comes down to this: take care of AI literacy, make sound agreements with the provider, and have the application assessed if AI has a say in a safety function or in how technicians are deployed. If you do use high-risk AI, use it according to the instructions for use, assign human oversight to people with the training and authority for it, keep the logs for at least six months and inform workers and their representatives before putting the system into use at the workplace.
How to get started with AI in maintenance
- Choose one reading problem with a clear payoff. For example, the reports after the shutdown, the manuals for a new installation or searching a drawing archive.
- Measure how things go now. How many hours does processing a report take, how long until a finding becomes a work order, and what gets left undone? Without a baseline you cannot prove anything later.
- Test on your own documents. Not just on a demo, but on your worst scans, your longest reports and your manuals in other languages.
- Measure time saved and quality. How many proposals were right, how many findings were missed, and how many quotes could not be found?
- Decide who checks and what may leave the building. Who reviews the proposals, what happens to rejected ones, and which documents may go to an external service?
- Expand step by step. Only when the first reading problem is demonstrably better do you tackle the next one.
Common mistakes with AI in maintenance
- AI as a goal. "We need to do something with AI" is not a problem statement. Start with the work that is left undone.
- Starting without basic data. Without a consistent asset register and digital documents, AI links findings to nothing.
- Accepting proposals blindly. A proposal in a confident tone is not a checked proposal.
- Starting big. A predictive model for every asset at once needs data you often do not have. Start with a few critical machines with good measurements, or with reading work.
- No owner. A trial without an owner and without measurement fizzles out after the first enthusiastic weeks.
- Expecting AI to prevent failures. AI delivers information faster. You reduce the likelihood of failures with the work that follows.
Frequently asked questions
Can AI predict failures?
Sometimes, if there is enough sensor data and failure history. A model that flags abnormal behaviour only needs data from normal operation and gives you a reason to go and look. A model that predicts remaining useful life needs examples of failures, and those are scarce for critical assets. For many assets, a trend or threshold on measured values already gives enough time to act. AI only reduces the likelihood of a failure when work follows.
Will AI replace technicians or planners?
No. AI takes over part of the reading and searching: going through reports, finding the right passage in a manual, identifying a part. Judgement stays with skilled people: whether a finding is correct, how urgent it is and what work follows. The physical work on the plant does not change. What does change is where a planner's time goes: less searching, more assessing.
Can you put company documents into an AI tool?
That depends on your agreements with the provider. Find out where documents are processed and stored, who can access them and whether they are used to train models. If they contain personal data, such as names in a service report, the GDPR applies. Do not paste confidential drawings or contracts into a public chatbot; use a tool your organisation has approved.
Does AI in maintenance fall under the EU AI Act?
Yes, but which requirements apply depends on the risk. The Act does not class reading reports, searching documents or identifying parts as high risk; the AI literacy obligation does apply. Stricter requirements apply to AI used as a safety component, for example in energy supply or in products under EU product legislation, and to AI that allocates work to technicians based on their behaviour or personal traits, or monitors and evaluates their performance. Since July 2026, AI that only supports efficiency or quality control does not count as a safety component, unless its failure would endanger safety.
How reliable is AI at reading reports?
It varies by document, and you can only establish it on your own reports. Clean digital reports usually go better than scans with handwritten notes. Two kinds of error count: an incorrect proposal, which you can spot during review, and a missed finding, which you do not. So measure both in a trial, and on critical plant always read the report's conclusions yourself.