MaintenanceNewPilot Pilot, prototype, early rollout or time-limited trial; results are limited or not yet measured.

AI reads technicians' free-text work orders and assigns the failure codes, so the CMMS shows what really fails (Janssen Pharmaceutica, oil and gas operator)

Added 2026-10-20 · Last updated 2026-10-20

Role

Maintenance (reliability engineering: CMMS data and failure analysis) · Typical users: reliability engineer, maintenance planner, maintenance manager, CMMS or SAP PM key user

The problem

Every failure notification in the CMMS asks for codes: what was defective, the cause, the action taken. Technicians pick from drop-down lists in a hurry and write the real story in short, abbreviated free text. Codes end up wrong, vague or missing, so bad-actor lists, failure-mode analysis, preventive maintenance plans and spares decisions rest on data nobody trusts. Recoding years of history by hand against a failure-mode framework is a months-long job that rarely gets done.

How AI helps

A model reads the free text of each work order or notification (telegraphic, full of abbreviations, sometimes in the local language) and proposes the codes: defect, cause and action, or component, failure mechanism and cause against a failure-mode taxonomy. Each proposal carries a confidence level; low-confidence records, records that fit no code and records that are not failures (planned work, testing) go to a reliability engineer instead of being forced into a code. Comparing the proposals with the codes already entered also surfaces likely mislabels, ambiguous codes and gaps in the code list.

Step by step

Export two to three years of work orders or notifications for one asset class: free text, current codes, asset, date and work type.
Clean the code list first: a definition and an example for each code, and a note on codes that overlap.
Have two reliability engineers hand-code a sample of 100 to 200 records to build a reference, and note where they disagree with each other.
Run the AI on the sample and sort the disagreements into AI wrong, code ambiguous, and text too poor to code.
Recode the history with a confidence level per record; engineers review the low-confidence and no-fit records; keep the original codes in a separate field.
For new work orders, show the proposed codes to the technician or planner at close-out to confirm, and fix the codes or writing rules behind recurring ambiguity.
Use the recoded history for bad-actor lists, FMEA updates and maintenance strategy, and re-check accuracy on a fresh sample every quarter.

Tools

Custom text-classification models built by Ghent University on the site's CMMS data: Janssen Pharmaceutica, Geel
Custom AI agents with SAP and Nexaan APM (Shivaan Asset Management): oil and gas proof of concept
Other tools to evaluate: AI features of your CMMS or asset performance management tool (SAP PM, IBM Maximo, Fiix, MaintainX), reliability consultancies that offer work-order text analytics
Any LLM (ChatGPT, Claude, Copilot, Gemini) with the prompt below, on an export of 100 to 200 work orders and your code list, for a first test

Result

Janssen Pharmaceutica (Johnson & Johnson) production site in Geel, Belgium, with Ghent University: tens of thousands of Dutch-language failure notifications from 2010 to 2020, each carrying three codes (one of 16 defect, 17 cause and 15 action codes). The best model predicted the codes with 0.95 average accuracy per code and got all three right on 83% of notifications. A manual review of the disagreements found that only about 4.2% of notifications were truly mislabelled by the model; about 7.4% had at least one ambiguous code and about 5.7% had text too poor for anyone to code, which points to codes and writing rules to fix. The authors also note room for improvement in how the cause codes are defined (peer-reviewed paper in Computers in Industry). Global oil and gas operator (anonymous), proof of concept by Shivaan Asset Management with custom AI agents on SAP data: 3,126 historical ball-valve work orders classified in one batch run against a failure-mode taxonomy; 99.2% classified (the 24 others had nothing written against them), 73.7% came back actionable with 76.4% average confidence, and 40.7% described more than one failure (consultant's project page).

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

The Janssen result is a research study on one site's history and the oil and gas result is a consultant's proof of concept: test on your own work orders before relying on it.
Keep the technicians' original codes; recoded fields go in a separate column until an engineer has reviewed them.
AI cannot recover what was never written: records with empty or one-word text need better writing rules at close-out.
Work-order history can contain names and confidential process details; use an approved enterprise tool.

Try it yourself

You are a reliability engineer cleaning up CMMS failure data.

I will paste: (1) our failure code list with definitions [code, type (defect / cause / action or component / mechanism), definition]; (2) a batch of work orders [work order number, asset, date, work type, free-text description and technician notes, codes currently entered].

Do this for each work order:
1. Propose one code of each type from my list only, quoting the words in the text that support it.
2. Give a confidence (high / medium / low) and, if a second failure is described, list it as a second set of codes.
3. Mark 'Not a failure' for planned work, inspections or tests, and 'No fit' if no code in my list matches, with a suggested new code name.
4. Compare with the codes currently entered and mark Agree / Disagree.
Then give a table of all work orders, and a short summary: the most frequent failure modes, codes that are often confused, and work orders whose text is too poor to code.

Never invent a code that is not in my list (except as a suggestion under 'No fit'), and never overwrite the original codes: this is a proposal for review.
Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter

Related cases

All 9 Maintenance use cases →