A maintenance technician on the floor types one sentence into their phone: "Pump 3 in the boiler room is making a grinding noise and dripping water from the bottom seal." Eight seconds later, OxMaint has parsed the message, identified the asset (Pump CT-03, Boiler Room A), classified the issue (mechanical failure with leak — Priority 2), routed the work order to the on-shift mechanical craft, attached the asset's last 6 service records, suggested the likely failed component (mechanical seal, $84 in stock), and notified the area supervisor — all without the technician opening a form, picking from a dropdown, or knowing the asset's official tag number. LLM-powered request systems reduce work order submission time by 73% and increase reporting frequency by 2.6x because the friction of "filing a ticket" is replaced with the simplicity of describing what is wrong. If you want to see how natural-language requests work against your real asset hierarchy, you can start a free trial and forward your first plain-English request in five minutes, or book a demo to see live request-to-work-order conversion across a sample plant.
LLM · Natural Language CMMS
LLMs in CMMS: Natural Language Maintenance Requests
A grounded look at how large language models turn plain-English maintenance descriptions into classified, routed, asset-linked work orders — and why this is the biggest workflow shift in CMMS in a decade.
Pump 3 in boiler room making grinding noise and leaking from bottom seal
Parsed
Asset: Pump CT-03 · Boiler Rm A
Issue: Mechanical failure + leak
Priority: P2 · Routed to: M. Singh
How long has this asset been running since last service?
Answered
2,840 hrs · Last seal replaced 6 mo. Likely cause: end of seal MTBF.
WO #84217 created · technician en route
73%
Reduction in average time to submit a maintenance request via LLM-powered interface vs traditional forms
2.6x
Increase in reporting frequency — technicians and operators flag issues earlier, more often
94%
LLM accuracy in correctly identifying asset and issue type from unstructured natural-language input
11 min
Average reduction in mean time to acknowledge a maintenance issue across a 250-asset facility
From Form to Conversation
Your Team Should Not Need a User Manual to Report a Broken Pump
Most CMMS request forms have 14-22 fields. Most maintenance issues can be described in two sentences. The gap between those two facts is why 40-60% of facility issues go unreported until they become emergencies. OxMaint's LLM layer takes plain-English descriptions and does the classification, routing, and asset matching automatically.
What an LLM-Powered Maintenance Request System Actually Does
A traditional CMMS request workflow starts with a 14-field form: select asset from dropdown, choose category, set priority, write description, attach photo, route to craft, set due date. An LLM-powered system replaces all of that with a single text input — typed, voice-dictated, or pasted from email or chat. The model reads the unstructured input, extracts the asset reference (even if the user calls it "the loud pump near the office"), classifies the failure mode against a library of known issues, infers the priority from severity language, and creates a fully populated work order with routing applied. The technician submitting the request does not see any form. They describe what is wrong in their own words. To see this in action against your asset library, you can book a demo and forward an actual request from your floor.
The 4-Stage NLP Processing Pipeline
01
Entity Extraction
The LLM identifies named entities in the request: assets, locations, components, symptoms, and time references. "The loud pump near the boiler" maps to an asset record using the plant's location hierarchy and historical aliases.
Named-entity recognition · asset graph lookup
02
Issue Classification
The model classifies the failure type using a fine-tuned taxonomy: mechanical, electrical, hydraulic, pneumatic, instrumentation, structural. It also detects severity language ("pouring water," "small drip," "intermittent") to set priority.
Multi-label classification · severity scoring
03
Context Enrichment
The system pulls the asset's recent maintenance history, last 3 work orders, runtime since last PM, and known failure modes. This context becomes part of the work order — saving the responding technician 8-15 minutes of pre-job research.
Retrieval-augmented generation (RAG)
04
Work Order Creation & Routing
A fully populated work order is created with asset, priority, craft, suggested parts, and a structured task list. Routing rules send it to the right craft based on shift, skill match, and current workload — with notification to the area supervisor.
Workflow engine · rule-based routing
6 Reasons Traditional Request Forms Fail in Practice
Form Field Fatigue
A 14-22 field request form takes 4-6 minutes to fill out. On a busy production line, that is enough friction to skip the report entirely — until the issue becomes a downtime event.
Asset Search Friction
An operator who calls the chiller "the green one near tank 4" cannot find it in a dropdown of 8,000 asset codes. They give up and call maintenance directly.
Wrong Priority, Wrong Routing
Users default to "high priority" for everything — making the priority field meaningless. P2 issues sit behind P1 self-classifications, and real urgency is invisible.
Language Barriers
Multilingual workforces struggle with English-only forms. An LLM accepts requests in Spanish, German, Hindi, or Arabic and produces a work order in the team's primary operating language.
Photo & Voice Inputs Missed
Form-based systems treat photos and voice memos as attachments — never reading them. Critical visual context (cracked weld, dripping flange) gets archived but not actioned.
No Historical Context Surfaced
A traditional WO arrives at the technician with zero context. They spend 8-15 minutes pulling up asset history, recent PMs, and known issues before they pick up a wrench.
How OxMaint's LLM Layer Works in Production
The LLM layer in OxMaint is not a generic chatbot bolted onto a CMMS. It is a fine-tuned model grounded in the customer's own asset hierarchy, work order history, and failure mode library, surfaced through six concrete capabilities that maintenance, operations, and reliability teams use daily. To run a request from your real plant data through the model, you can start a free trial and submit a sample request in any phrasing your operators use today.
CAP 01
Plain-English Submission
Type, paste, voice-dictate, or forward email. The model handles incomplete sentences, jargon, slang, and even spelling errors common in mobile typing.
CAP 02
Asset Disambiguation
"The loud pump" or "boiler room compressor" maps to the right asset using location hierarchy, historical aliases, and confidence scoring — with confirmation when ambiguous.
CAP 03
Auto-Classification & Priority
Failure mode, severity, and craft assignment inferred from the description language. "Pouring water" triggers P1; "occasional drip" sets P3.
CAP 04
Multilingual Acceptance
Native support for English, Spanish, German, French, Arabic, Hindi, Mandarin, and Portuguese. Operator submits in their language, work order produced in plant operating language.
CAP 05
Context-Aware Work Order
Each WO arrives with last-3 service history, runtime since last PM, top-likely failure causes, and suggested spare parts — pre-staged for the responding craft.
CAP 06
Channel-Agnostic Intake
Requests arrive via mobile app, SMS, WhatsApp, Slack, Teams, email, or QR-code scan on the asset itself. All routes through the same LLM pipeline.
Traditional Form vs LLM-Powered Request
Traditional Form Submission
Time to submit4-6 min
Fields to fill14-22
Asset selectionDropdown of 8,000+
Priority accuracy~ 38%
History attachedNone
Mobile UXForm-zoom, dropdowns
Languages supported1
Voice inputNo
vs
OxMaint LLM-Powered Request
Time to submit20-40 sec
Fields to fill1 (free text)
Asset selectionAuto-matched from description
Priority accuracy~ 91%
History attachedLast 3 WOs + runtime
Mobile UXSingle text box, voice-enabled
Languages supported8+
Voice inputYes (whisper transcription)
Reported Outcomes from Plants Using LLM Requests
73%
Faster average request submission time
From 5.2 min average to 1.4 min
2.6x
More issues reported per shift
Lower friction surfaces small issues earlier
91%
Auto-classification accuracy
Vs 38% for self-prioritisation in forms
11 min
Mean time to acknowledge cut
Pre-staged context skips research lookup
Frequently Asked Questions
How accurate is the asset matching when an operator describes equipment informally?
Asset disambiguation accuracy averages 94% on first try across mid-market plant deployments. The model uses location hierarchy, historical asset aliases, and prior request patterns to score candidate assets — and confirms with a single-tap selector when confidence is below threshold. Over time, accuracy improves as the system learns each plant's local naming conventions.
Does the LLM layer require sending our maintenance data to a third-party AI service?
No. OxMaint runs the LLM layer in a tenant-isolated environment with no training on customer data and no cross-tenant data sharing. For regulated industries (defence, pharma, government), a private deployment option keeps the model entirely within the customer's cloud or on-prem infrastructure.
What happens when the LLM is uncertain or makes a wrong classification?
Every classification carries a confidence score. Below threshold, the system prompts a one-tap confirmation from the requester or routes to a human triage queue. Misclassifications are logged with feedback — the model improves on each plant's specific vocabulary over the first 60-90 days of use.
Can requests come from channels other than the OxMaint app?
Yes — intake channels include SMS, WhatsApp, Microsoft Teams, Slack, email-to-WO, QR-code scan on the asset itself, and direct mobile app voice. All channels feed the same LLM pipeline and produce the same structured work order output regardless of how the request arrived.
Natural Language · AI-Powered CMMS
Stop Building Forms. Start Reading What Your Team Actually Wrote.
Your operators and technicians know exactly what is wrong with the equipment. They describe it in their own words every shift — in chats, emails, voice memos, hallway conversations. The friction has always been the gap between what they said and what your CMMS would accept. OxMaint's LLM layer closes that gap completely. Plain-English in. Fully classified, asset-linked, context-rich work order out. The reporting friction is gone. The reporting volume goes up. The issues get caught earlier.
