Industries / Industrial equipment and services / Industrial services / CMMS and asset management software
Work orders
A work order is the record of one maintenance job: what was asked for, who did it, what they found, what they replaced and how long it took.
4
Milestones achieved, from 1964 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1964
One form for every job
Under the Navy's 3-M System each maintenance action was "documented (OPNAV FORM 4700.2B) by the persons performing the maintenance". The forms were collected on board and sent in batches to the Maintenance Support Office in Mechanicsburg, Pennsylvania, set up on 1 August 1964 as the system's central data processing facility, by Harvey Oelkers's thesis.
Credit
US Navy, Chief of Naval Operations
1985
Work orders on a PC
Project Software & Development Inc. released Maximo 2.0 in 1985, a single-user program for the IBM PC AT that handled work orders and inventory. By IBM's account it moved to Windows in a client/server form in the late 1980s, was "fully Java/Internet based" with Maximo 6 by 2005, and was bought by IBM in 2006.
Credit
Project Software & Development Inc., later MRO Software; IBM
2019
Tagging technicians' notes
Technicians write work orders in short, misspelt jargon, such as "Hyd leak at saw attachment". In November 2019 Rachael Sexton and Michael Brundage of the US National Institute of Standards and Technology (NIST) described Nestor, a public-domain tool that helps an analyst tag such text with items, problems and solutions so that it can be counted. NIST's own page calls the build "very early alpha".
Credit
Rachael T. B. Sexton, Michael P. Brundage, NIST
2023
AI picks failure codes
Michael Stewart, Melinda Hodkiewicz and Sirui Li tested GPT-3.5 at labelling maintenance observations with one of 22 failure-mode codes from ISO 14224, the standard list for oil and gas equipment. Tuned on 502 expert-labelled examples, it scored an F1 of 0.80, against 0.46 without tuning and 0.60 for a conventional text classifier. The test set held 62 observations.
LessonThe authors wrote that the result "reinforces the need for high quality fine-tuning data sets".
Credit
Stewart, Hodkiewicz and Li
Watch for
Failure codes filled by AI
IBM Expert Labs offers to set up "recommended missing failure codes" in Maximo, "analyzing completed work orders and recommending accurate failure codes", along with detection of duplicate work orders. The announcement gives no accuracy figures.
Achieved whena customer or outside group publishes how often the suggested codes match codes assigned independently by reliability engineers.
Claimed by
IBM, February 2026
Agents inside the CMMS
IBM says Maximo Application Suite 9.2 brings "agentic workflows designed to help guide decisions and move work forward", scheduling by plain-language conversation, and a server through which organizations "can bring their own agents and integrate them directly with Maximo Manage APIs". It gives no figures.
Achieved whena named customer publishes measured results, such as planner hours or schedule compliance before and after, checked by someone outside IBM.
Claimed by
IBM, June 2026
Open
01
Score the failure codes
An open test of AI failure coding. Take at least 1,000 real work orders, have two reliability engineers each code them separately with ISO 14224 failure mode, mechanism and cause, then have several AI systems code them. Publish the AI's agreement with the engineers beside the engineers' agreement with each other.
The briefCredit · unclaimed02
A shared work order set
A public set of at least 50,000 anonymized work orders from at least five organizations in different trades, such as buildings, water, manufacturing and mining, with names, places and costs masked and 2,000 records labelled with ISO 14224 codes.
The briefCredit · unclaimed
Open 01 · the brief
Score the failure codes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open test of AI failure coding. Take at least 1,000 real work orders, have two reliability engineers each code them separately with ISO 14224 failure mode, mechanism and cause, then have several AI systems code them. Publish the AI's agreement with the engineers beside the engineers' agreement with each other.
Why it is still open
Vendors now sell failure-code suggestions without accuracy figures. The 2023 academic test covered failure mode only, on 62 test observations labelled by one expert.
Built with what exists
the MaintIE set of maintenance texts, released under the MIT license; the ISO 14224 code lists; Gemini Decision 1 for code probabilities (see the Note "A thousand questions, one pass"); a small model tuned on one graphics chip, such as Liquid LFM3 (see the Note "One GPU to fine-tune"); Cohen's kappa, the usual measure of agreement between two coders.
First step
Take 100 texts from MaintIE and have two engineers code them separately, to measure how often people agree before testing any model.
Done looks like
A published table of human-to-human and model-to-human agreement for each of the three fields, on at least 1,000 work orders from at least two industries.
Open 02 · the brief
A shared work order set
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 50,000 anonymized work orders from at least five organizations in different trades, such as buildings, water, manufacturing and mining, with names, places and costs masked and 2,000 records labelled with ISO 14224 codes.
Why it is still open
The MaintIE authors write that work orders "are often confidential, and organisations are hesitant to release their data". Before MaintIE only one labelled set had been released, and MaintIE's own texts come from mining.
Built with what exists
the masking approach of MaintNorm, the same research group's 2024 dataset for cleaning and masking maintenance text; Python and pandas; a data-sharing agreement; an open license such as CC BY 4.0.
First step
Ask three public bodies, such as a university facilities office, a city water utility and a school district, for one year of closed work orders under a data-sharing agreement.
Done looks like
The set released with a description of each source, a check by two people that no names or addresses remain in a random 1,000 records, and a baseline coding score.