AI in a maintenance shop is useful today for paperwork, search and summaries. It isn’t a mechanic, and it can’t see a machine it has no data on. Knowing which is which keeps a trial from turning into a disappointment.
What the tools actually do
The AI assistants now showing up in fleet software are built on language models. They read what they are given and write a likely response. Given a machine’s fault codes, hours, inspection notes and service manual, they summarize, search and draft well. Given nothing, they produce something that sounds right.
That leads to the one test for any AI feature: what data is it reading, and can it show you where its answer came from?
Four jobs it does well now
1. Drafting work orders. A fault code, an inspection note about a weeping hose and the machine’s hours go in. A draft work order comes out, with tasks and a parts list taken from the manual for that model. A shop lead reads it, fixes what is wrong and approves it. The typing is the part that goes away.
2. Summarizing fault history. Fault codes pile up faster than anyone reads them. Under SAE J1939-73, each diagnostic trouble code combines a Suspect Parameter Number, a Failure Mode Identifier and an Occurrence Count. An AI summary can turn 300 hours of those into a sentence a mechanic can use: this code logged six times, always in the first ten minutes after a cold start, only on this machine.
3. Searching manuals. “What is the torque for the bucket pin retaining bolt on this model?” A good tool answers with the value and the manual page it came from, and the mechanic checks the page. It can answer in Spanish from an English manual, which matters in a New Mexico shop.
4. Answering fleet questions. Which machines on the Rio Rancho job are due for a 500-hour service in the next two weeks? How many hours did rented machines run last month? Questions like these used to wait for someone in the office to build a report. When the data is structured and current, an assistant answers them in plain English or Spanish in seconds.
A fifth job sits close behind: cleaning up inspection notes typed with gloves on, so “hyd leak lft boom cyl” becomes a line a mechanic two shifts later can act on.
Where it falls short
Diagnosis. A model can’t hear a knock, feel a hot bearing or see a cracked weld. A fault code points to a circuit or a condition, not always a cause. A sensor code can mean the sensor, the harness or the thing the sensor measures. AI can lay out the troubleshooting steps from the manual; a mechanic still does them.
Missing data. If engine hours come from a timesheet typed in on Fridays, the AI summarizes bad hours very fluently. If service history lives in a binder, the AI doesn’t know it exists.
Guessing. Asked for a torque value or a part number it wasn’t given, a model may produce a plausible one. Treat any answer without a source as a guess, however confident it sounds.
Prediction. Predicting a failure needs many examples of that failure on that model. A 60-machine fleet may see a given failure only a few times a year, which is thin ground for a forecast. Schedules, fault codes and inspections still carry most of the load.
Acting alone. A tool that orders parts or closes work orders on its own will eventually get one wrong, with nobody watching.
Safety calls. Whether a machine with a defect runs, runs with a limit or gets tagged out is a decision for the shop lead or superintendent. AI can put the defect, the history and the manual’s guidance in front of them. It shouldn’t make the call.
Example: a morning in the shop
Example engagement. A 90-machine contractor in Albuquerque runs an AI agent on its own fleet data. At 6:30 a.m. the shop lead opens the draft queue. Overnight, fault codes and yesterday’s inspections produced 11 draft work orders. Eight are right as written and get approved. Two need the parts list changed, because the shop stocks a different filter number. One is a duplicate of an open work order and gets rejected.
At 7:15 a.m. a mechanic heading to a loader near Los Lunas asks, in Spanish, for the hydraulic filter change steps for that model. The answer comes back with the manual page. The mechanic checks the page on a phone before leaving the yard.
At 9 a.m. the equipment manager asks which machines on the Rio Rancho job have open defects older than a week. The list has three machines on it, each with a link to the record.
Nothing in that morning replaced a mechanic. It removed typing, searching and report building, and every action still had a person’s name on it.
A person approves every action
The rule that makes AI safe in a shop is simple: the AI drafts, a person decides.
| Action | AI may | A person must |
|---|---|---|
| Work order | Draft tasks and parts | Approve, edit or reject |
| Parts order | Suggest a list | Place the order |
| Closing a work order | Summarize what was done | Close it |
| Warranty claim | Assemble the history | Review and submit it |
| Fleet question | Answer with sources | Check before acting |
Two more controls matter. The AI should have the same permissions as the person using it, and no more. And every draft, edit and approval should be logged, so you can see who approved what and when.
Decision rule. If an action spends money, changes a machine’s status or reaches someone outside the company, a named person approves it. No exceptions for “small” orders.
Get the data ready first
An AI tool is only as useful as what it can read. Before a trial, check:
- engine hours come from the machine, not a timesheet
- fault codes arrive with SPN and FMI, by machine
- service history is imported with dates and hours, not left in binders
- manuals for each model are loaded as searchable documents
- inspection notes are text, not photos of paper
- machine numbers match across every system
- part numbers are cross-referenced, so one filter has one record
If three or more of these are missing, fix them first. The same work makes your reports better whether or not you ever turn on an AI feature.
Test it on your own questions
Demos use questions that work. Use yours.
- Write 20 real questions from your shop and office, with the right answers, before the trial starts.
- Score every answer: correct with a source, correct without one, or wrong.
- Draft real work orders for two weeks and count what the shop lead changed.
- Check the Spanish with a bilingual mechanic, not a translation tool.
- Measure time saved on the paperwork, with a stopwatch rather than a feeling.
Weekly hours saved = work orders drafted per week × minutes saved per draft ÷ 60
Worked example. A shop drafts 40 work orders a week. If AI drafts save 6 minutes each after the shop lead’s review, that is 240 minutes, or 4 hours a week. Measure your own minutes, and don’t accept anyone’s estimate, including ours.
Ask any vendor three questions. Does it cite its sources? Can it act without a person’s approval? Can you export everything it touched? The buyer’s guide lists these alongside the other questions worth asking.
Mistakes in early AI rollouts
- Starting with prediction. Paperwork pays sooner and fails safer.
- Write access without approval. One wrong parts order or a work order closed unchecked undoes the time saved.
- Answers without sources. Accepting a torque value nobody checked.
- Messy data in. Expecting good summaries from timesheet hours.
- Judging by the demo. Test with your questions, your machines and your manuals.
- Leaving out the mechanics. If the people turning wrenches don’t trust the drafts, they won’t use them.
An agent that drafts, not decides
Gytix’s AI maintenance agent reads fault codes, hours, inspection notes and your service manuals. It drafts work orders and parts lists for a person to approve, answers questions about the fleet in plain English or Spanish, and never closes a work order or orders parts without a human approving it. It is part of the Command plan.
If your team already uses an AI assistant, API and integrations includes an MCP server that lets it query fleet data and create draft work orders, with the same permissions as the user who connects it.
