AI equipment maintenance logs.

Why farm maintenance records go missing

Farm machinery maintenance usually lives in a workshop notebook, a mechanic’s invoice, or nowhere. The work gets done; the write-up, with greasy hands at the end of a wet day, is what slips. An AI maintenance log removes the writing. Staff speak a 20-second note into the farm phone, and the system turns it into a dated entry against the right machine, with the hour reading, the work done, the parts used and any defect found.

The log then tracks service intervals by hours or kilometres and makes sure a reported defect cannot sit unread. Safety decisions stay with the manager.

How the maintenance log runs

Inputs

Voice notes and photos from staff after pre-start checks, servicing or repairs; hour meter and odometer readings; the machine register with fleet numbers, serials and manufacturer service intervals; mechanic and parts invoices; and previous log entries.

Systems

A capture app or chat channel on the farm phone, a transcription model, a small language model to structure entries, a maintenance register (a spreadsheet or simple database), and invoices from email or Xero.

Steps

Transcribe the note; identify the machine; extract the reading, work done, parts and defects; check the reading against the last entry; classify any defect; update the register; calculate the next service in code; draft reminders; send safety defects to the manager at once.

Human review

The manager reads every defect marked as safety-related and decides whether the machine is used, tagged out or sent to a mechanic. A weekly review covers all new entries. Mechanics sign off their own repairs.

Outputs

A log per machine, a service-due list, a defect register with open and closed status, and a short monthly summary.

Measures

Services completed by their due reading, days a safety defect stays open, entries needing correction, and machines with no entry in 30 days.

Risks

Misheard readings, entries filed to the wrong machine, softened defect wording, and reminders that everyone learns to ignore.

Illustrative example: a Southland sheep and beef station

Illustrative example, synthetic numbers: a Southland sheep and beef station runs 14 machines: four quad bikes, two side-by-sides, two farm motorbikes, three utes and three tractors with attachments. Staff record about 90 maintenance events and pre-start findings a month in a shed notebook, and the manager spends about 4 hours a month deciphering it and chasing gaps.

Assumptions, at OpenAI prices listed on 19 October 2026: one minute of audio per entry transcribed with gpt-transcribe at US$0.0045 a minute (about 41 US cents for 90 entries), then structured with gpt-6-luna at US$0.10 per million input and US$0.50 per million output tokens, assuming 2,500 input tokens (transcript, register extract, last entry and instructions) and 300 output tokens, about 4 US cents for the month. Service-due arithmetic runs in ordinary code, not in the model.

The NZ health and safety context

• Safe plant. Section 36(3)(b) of the Health and Safety at Work Act 2015 requires a business to ensure, so far as is reasonably practicable, the provision and maintenance of safe plant and structures. A complete log is evidence of maintenance; it is not the maintenance.
• Farm vehicle code of practice. WorkSafe’s approved code of practice for safe farm vehicle operation comes into effect on 1 April 2027, alongside the Health and Safety at Work Amendment Act. It covers safe operation and maintenance of high-risk vehicles and attached machinery, and names quad bikes, side-by-sides, two-wheel motorbikes, utes, four-wheel drives and tractors. WorkSafe notes an ACOP is not mandatory but sets a standard for compliance, and that this one will need reapproval after 1 April 2027 before it gains safe harbour status.
• Staff voices are personal information. Voice recordings are personal information under the Privacy Act 2020. Delete the audio once the entry is confirmed, unless there is a reason to keep it.
• Invoices are tax records. IRD expects business records such as repair invoices to be kept for at least 7 tax years. Link them to the machine; do not copy them.

What the manager decides

The manager decides whether a machine stays in use, what gets repaired and by whom, whether a service interval changes, and who is competent to ride or drive what. No model output should mark a safety defect as resolved; only a person who has checked the machine can do that. Our guide to human review queues shows how to keep safety items separate from routine reminders.

How maintenance logging fails

• Misheard readings. “Twelve fifty” becomes 2,150 hours, and the next service appears to be overdue by 900 hours. Reject any reading that goes backwards or jumps more than a plausible amount, and ask the speaker to confirm.
• Two red quads. Staff call machines by colour or nickname. Put a fleet number on every machine, store nicknames in the register, and ask when the match is unclear.
• Softened defects. “Brakes are a bit soft” summarised as “minor brake adjustment” hides a hazard. Keep the speaker’s exact words for any safety item and flag brakes, steering, tyres, rollover protection and lights automatically.
• Attachments forgotten. A spray tank or trailer changes how a vehicle handles, and the ACOP’s scope includes attached machinery. Give attachments their own entries.
• Reminder fatigue. Daily nagging gets muted. Bundle routine reminders weekly; keep safety alerts separate.

Maintenance entries become searchable through farm record retrieval and form part of the pack in compliance evidence collection. See the agriculture sector guide and our AI workflow automation service for the broader build.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.