An AI maintenance assistant is software that answers a technician's questions, such as "what does this fault code mean and how did we fix it last time?", from machine manuals, procedures and past breakdown records, citing the source document. It does not monitor machines or predict failures; it puts the right knowledge in front of the technician when something breaks.
Key takeaways
- An AI maintenance assistant answers technicians' questions from manuals, procedures and past breakdown records, and cites the source.
- A maintenance assistant differs from predictive maintenance: it needs no sensors and answers "how do I fix it?", not "when?".
- Atlassian's State of Teams 2025 survey found that teams spend 25% of their time just searching for answers.
- Rolling out a maintenance assistant starts with three to five critical machines; tidy breakdown records set answer quality.
- A maintenance assistant's success is measured by mean time to repair (MTTR) and outside support calls before and after a pilot.
What does a maintenance technician have to hand when a machine stops?
It is the night shift and a packaging machine stops with a fault code on the panel. The technician has three options: open the 400-page manual in the cabinet, leaf through the maintenance log for something similar, or ring the senior fitter who is asleep at home. Usually it is the phone call.
The problem is not a lack of knowledge but where it sits. The manufacturer's manual is a PDF, the in-house procedures live in a shared folder, and "we replaced the sensor cable last year" survives only in a free-text note on an old work order, or in one person's head, and leaves when that person retires.
This guide is not about systems that use sensor data to warn you before a failure. It covers the AI assistant that helps a technician once the failure has happened. We explain condition monitoring in our predictive maintenance guide, and work orders and spare parts in our CMMS software guide.
How much does searching for information add to downtime?
Unplanned downtime is expensive in every sector, and 2024 figures show it still happens often:
Siemens estimates that the world's 500 largest companies lose almost $1.4 trillion a year to unplanned downtime, equivalent to 11% of their revenues, with large plants averaging 25 downtime incidents and 27 lost hours a month. — Siemens, The True Cost of Downtime 2024
Some unplanned downtime goes on waiting for parts or on the repair itself, but some goes on finding the right information. Searching is a general burden: Atlassian's State of Teams 2025 survey of 12,000 knowledge workers found that teams spend 25% of their time just searching for answers. On the shop floor, every minute of that search is added straight to the downtime.
The real value of a maintenance assistant is that it narrows the experience gap. In the NBER study "Generative AI at Work", a generative AI assistant given to 5,179 customer support agents raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled agents. The study was not done with maintenance teams, but the mechanism it shows carries over: the assistant moves what experienced people know to those who are new.
What does an AI maintenance assistant do, and what does it not do?
Setting out the assistant's role early prevents the wrong expectations and the wrong investment. The table below separates the maintenance assistant from predictive maintenance and from a CMMS (computerised maintenance management system):
| Criterion | Maintenance assistant (knowledge) | Predictive maintenance (sensors) | CMMS (records) |
|---|---|---|---|
| Core question | "How do I fix this fault?" | "Which machine will fail, and when?" | "Which job was done, by whom, when?" |
| Data source | Manuals, procedures, past work orders | Vibration, temperature, current data | Work orders, schedules, stock |
| When it works | During a breakdown and in training | Continuously, before failure | Planning and closing jobs |
| Hardware needed? | No; a tablet or phone is enough | Yes; sensors and connectivity | No |
| Main indicator | Mean time to repair (MTTR) | Number of unplanned stops | Schedule compliance, record quality |
A maintenance assistant, predictive maintenance and a CMMS are complementary, not competing. In a sound set-up, closed CMMS work orders feed the assistant's knowledge base, and a predictive maintenance alert links to the assistant with the question "what do we do when this alert fires?".
What a maintenance assistant does not do should be equally explicit. It does not operate the machine, it does not shortcut lockout/tagout (LOTO), and it does not "estimate" a torque value or setting that is missing from the documents. If the answer is not in the sources it should say so; we explain how in our article on reducing AI hallucinations.
Which knowledge sources feed a maintenance assistant?
Manufacturer manuals. Service manuals, fault code tables and spare parts lists. Many are written in English, German or Italian, while technicians often think and ask in another language; the assistant can answer from the foreign-language section and cite it. Old scanned manuals are first made searchable with document OCR (optical character recognition).
In-house procedures. Work instructions, planned maintenance checklists, safety procedures and quality documents. Where they contradict the manufacturer, decide in advance which one wins.
Breakdown history. For a maintenance assistant, the most valuable and most scattered source is breakdown history: the "symptom, cause, action" notes on closed work orders. The assistant answers "how many times has this fault code appeared on this machine in two years, and how was it fixed?" from these records. The quality of the notes sets the quality of the answers.
Voice notes. A technician wearing gloves, with oily hands, will not type long notes. Speaking a few sentences into a phone after the job, converted with speech-to-text, turns into searchable text on the work order for the next breakdown.
The architecture that grounds answers in manuals, procedures, breakdown records and voice notes is called RAG (retrieval-augmented generation): the assistant first retrieves the relevant passage, then answers from it. The details are in our RAG explainer.
How do you set up a maintenance assistant? Six steps
- Choose the critical machines. Start with the three to five machines that stop most often or hurt the line most when they do. List a year of breakdown records and the documents you hold for each.
- Collect documents and assign owners. Bring manuals, procedures and checklists into one place. Decide who approves the current version of each; two conflicting procedures mean two conflicting answers.
- Tidy the breakdown records. Split free-text notes into at least machine, symptom, cause and action fields. If there are no records, start collecting them today with a mobile service report or voice notes.
- Put it on the device technicians already use. The assistant belongs on the workshop tablet, the technician's phone or the work order screen. A QR code on each machine that opens a flow focused on that machine's documents shortens every search.
- Write safety limits into the rules. For safety, electrical and pressurised-system steps, the assistant should show the approved procedure verbatim, without commentary. Where authorisation is required, the authorised technician has the final word.
- Measure with a pilot. Compare mean time to repair, the number of calls for outside help and the share of "not found" answers before and after. Our guide to measuring AI project ROI sets out the calculation.
Bringing manuals and procedures into one place does not always require a new system. SmartFiles document management, part of Digital Bridge's Smart360 family, keeps manuals and procedures in department spaces; a document uploaded under the same name is never overwritten but stored as a new version. Each upload is summarised automatically by AI, and a technician can ask questions of a single document or a whole folder, with answers drawn only from files that person is allowed to see.
There is another reason to keep a maintenance assistant pilot small. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 because of poor data quality, inadequate risk controls, escalating costs or unclear business value. In a maintenance assistant, scattered breakdown records are the first of those risks.
What have companies gained from internal AI assistants?
Two public examples from outside maintenance show how scattered knowledge can be packaged and put to use. According to Microsoft's Eaton customer story, finance teams at the electrical equipment maker used Microsoft 365 Copilot to help produce 1,000 standard operating procedures. The company says time per procedure dropped from one hour to 10 minutes, saving more than 650 hours.
On the access side, OpenAI's Morgan Stanley customer story reports that over 98% of the firm's advisor teams actively use its internal assistant, and that access to documents rose from 20% to 80%. The lesson carries over: the assistant makes existing documents actually get used. Both figures are company or vendor statements. For uses of AI beyond maintenance on the factory floor, see our roundup of AI in manufacturing examples.
How does Digital Bridge build a maintenance assistant?
At Digital Bridge we do not sell a boxed maintenance assistant; we start by seeing your workshop and maintenance records on site. In discovery we map your critical machines, where documents live, what your breakdown records look like and the moments when technicians most need information. We then prepare a written proposal covering scope, phases and fees.
We build the core as an enterprise LLM assistant that answers from your own documents and cites its source, and it can run entirely on your own servers. We use voice recognition to make note-taking easy on the floor, and AI integration to connect the assistant to your existing work order and field service management screens.
The maintenance assistant pilot runs on a few critical machines with one maintenance team, and the metrics are written down first. Once the assistant is in place, if you want early warnings on the same machines we can run a predictive maintenance pilot on one to three critical assets; when an alert fires, the technician asks the assistant what to do. To see where this fits in your wider plan, read where to start with AI in business; for the office-side equivalent, see our agent assist guide.
Breakdown records may include technicians' names, and voice notes may contain personal data. We agree retention and access before the pilot starts. In Turkey this falls under KVKK, the Personal Data Protection Law No. 6698; our article on AI and KVKK summarises the framework.
What is the first step in preparing for a maintenance assistant?
Pick the five longest breakdowns of the past six months and, with your senior technician, roughly mark how much time went on searching, asking someone and finding a document in each. That list shows which machines and which questions the assistant should tackle first. Bring it to us via our contact page and we will scope the pilot together in a discovery call. For more articles, visit our AI topic page.