AI chat tools (such as ChatGPT, Claude and Gemini) have already quietly entered the daily work of maintenance teams. Some technicians type in a fault symptom and ask for possible causes; some supervisors ask for a draft of the monthly report. Used well, they save a lot of time; used badly, they can lead you astray with convincing but wrong information. In this post I describe where AI really helps in maintenance, sample prompt patterns, and the limits to watch out for.

Where does it really help?

TaskWhat AI contributesWhat you contribute
Writing a breakdown reportTurns a scattered note into a clean, readable reportProviding the real measurements and observations
Root cause analysisLists possible headings for 5 Whys or a fishbone diagramDeciding which one holds true on site
Preparing maintenance instructionsDrafts a step-by-step work instructionVerifying machine-specific torque, settings and safety values
Reading manuals in a foreign languageTranslates technical text into plain language and explains termsChecking critical values against the original
Excel formulas and reportsWrites the formulas for calculations such as MTBF and MTTRMaking sure the data is correct

Breakdown reports: from a scattered note to clean text

A note taken in the field usually looks like this: "cmp3 high temp alarm, fan belt loose, replaced, 2 hours downtime, oil level normal". The note is correct, but incomplete for anyone else reading it. Giving it to the AI like this works well:

"Turn the field note below into a maintenance breakdown report. Headings: Equipment, Symptom, Finding, Action taken, Parts used, Downtime, Recommendation. Do not add information that is not in the note; write 'not stated' where something is missing. Note: …"

The key sentence here is "do not add information that is not in the note". Without this instruction, the AI fills the gaps with its own guesses and may put a temperature value in the report that was never measured.

Root cause: use it as a sparring partner

With a recurring breakdown, the team often gets stuck on the same two or three causes. AI is good at reminding you of headings that were overlooked:

"The mechanical seal on a centrifugal pump leaks every 3 months. Flow and pressure are normal. List possible causes for a fishbone diagram under the headings People, Machine, Method, Material, Measurement, Environment. For each cause, write in one sentence how to verify it on site."

The last sentence matters: the question of "how to verify it" turns the list from a pile of guesses into a checking plan.

Preparing instructions and checklists

Writing a work instruction for a new maintenance item can take hours. AI produces a draft in minutes. However, these values must always be checked against the manufacturer's documentation:

  • Tightening torques and tightening sequence
  • Oil type, viscosity and quantity
  • Clearance, tension and alignment tolerances
  • Energy isolation (lockout-tagout) steps

AI can produce "plausible-looking" numbers for these values. Saying "usually 25 Nm" for a bolt's torque does not mean the bolt on your machine is 25 Nm.

Limits: where not to trust it

  • Fabrication (hallucination): The model can write about something it does not know in a confident tone. Standard numbers, model codes and numerical values in particular should be read with suspicion.
  • Out-of-date information: Models' knowledge stops at a certain date; they may not know about a newly issued regulation.
  • Lack of context: It does not know the environment your machine works in or its past breakdowns. The more context you give, the more accurate it gets.
  • Safety decisions: Decisions such as whether a machine may be run, or whether a piece of lifting equipment is safe, must never be left to AI.

Data privacy: what not to type

Text typed into free chat tools may, depending on the terms of service, be used to improve the model. So:

  • Do not type personal or commercial data such as customer names, contract values or staff details.
  • Equipment codes and technical symptoms are usually fine; removing the company name is enough.
  • For business use, prefer plans where data is not used for training.

The four parts of a good prompt

  1. Role and context: "I am a maintenance supervisor at a food factory; the equipment is…"
  2. Task: "Turn this note into a report" or "list possible causes"
  3. Format: Headings, table, number of items
  4. Limit: "Do not make up values you don't know; mark them 'to be checked'"

A prompt with these four parts gives a far more reliable result than a one-sentence question. Thinking of AI not as a source of information but as a fast-writing, tireless assistant is the shortest way to put it in the right place in maintenance.

Frequently asked questions

Can you trust a torque or setting value from an AI tool?

No. Language models can produce convincing but wrong numbers. Values such as torque, tolerance and pressure must always be checked against the maker's documents or a standard.

Is it safe to upload breakdown records to an AI tool?

Customer names, personal data, passwords and trade secrets should not be typed into publicly available tools. If there is no company-approved tool, anonymise the data first.

Where does it help most in maintenance?

Turning scattered notes into a tidy breakdown report, listing possibilities you had not thought of during root cause analysis, and drafting instructions and checklists. The final check always stays with someone who knows the job.

Sources

This article is a compilation based on the sources below and personal experience; its assessments are opinions. In practice, the manufacturer's official documentation and applicable regulations take precedence.

  • EN 13306 — Maintenance: Maintenance terminology
  • Law No. 6698 on the Protection of Personal Data (when entering personal/company data into artificial intelligence tools)