A technician remembers a previous fault, a colleague searches for the work order, and a supervisor reads several notes to establish what was actually replaced. For a small or medium-sized manufacturer, improving that search can be a more practical starting point for artificial intelligence than trying to predict every equipment failure. The useful question is whether maintenance records can become easier to use while the existing ERP and maintenance system stay in place.
A limited project can connect those systems to document search or assisted classification. It still needs a clear view of data quality, access permissions and the time currently spent on the task. The workflow below is a proposed pilot to evaluate in your plant. It is not a customer case study or a promise of reduced downtime.
What is changing in industrial AI in 2026?
The NIST roadmap published on July 3, 2026 discusses industrial AI opportunities alongside barriers involving data management, integration and reliability. It does not establish that a general-purpose model can independently diagnose a smaller manufacturer's equipment.
The earlier NIST Nestor project offers a more focused reference: helping domain specialists structure technical text from maintenance work orders. Its relevance here is assisted annotation. It should not be presented as a newly released predictive-maintenance product for 2026.
Our assessment is that manufacturers with fragmented records may benefit first from improving how information reaches people. Finding a comparable intervention, drafting a source-linked summary and suggesting consistent labels are testable functions. Technical judgment and authorization to perform work should stay with appropriately qualified staff.
Turn an old work order into a verifiable answer
Consider this hypothetical example. A team receives a new issue involving a conveyor. Its intake form records the equipment identifier, date, observed symptom and relevant operating conditions. The application searches earlier records for that equipment. If it also finds similar machines, it keeps those results visibly separate rather than treating them as the same asset.
- Collect: import completed work orders from the computerized maintenance management system, or CMMS, or a controlled ERP export.
- Normalize: reconcile asset identifiers, units and part names while retaining the original text and marking missing fields.
- Retrieve: show a short set of relevant interventions with dates and links back to their source records.
- Draft: summarize documented observations without presenting an earlier hypothesis as a confirmed cause.
- Review: let an accountable person accept, correct or reject the summary before it becomes part of the new record.
The application also needs a useful “insufficient information” outcome. A note saying “noise near motor” does not identify a failed component or establish a repair procedure. An old repair record cannot replace current manufacturer instructions or the plant's safety procedures.
Check the records before connecting a model
Start with a representative sample of work orders that the team already uses. Include routine interventions, incomplete records, similar equipment names and both French and English notes where those languages occur. The purpose is to find ambiguities before reproducing them throughout a larger collection.
- A stable equipment identifier, linked to location and version where those distinctions matter.
- Separate dates for the issue report, intervention and return to service.
- The symptom, observations, completed actions and parts actually used.
- A clear record status, such as suspected cause, completed intervention or confirmed cause.
- A durable source reference and permissions that allow the intended reader to open it.
Do not ask a model to invent missing durations or part numbers. Keep unknown values explicitly unknown, and distinguish a period without recorded failures from a period without records. If structured fields already answer the question, filtered search or a straightforward dashboard may be sufficient. AI should earn its place through the evaluation rather than appear in every step by default.
Keep the ERP and add one useful capability
A pilot can begin with a read-only export and a limited group of assets. That makes it possible to compare answers without altering production work orders. If the results justify further work, a documented connection can bring in permitted updates from the existing system.
Decide how synchronization handles frequency, deleted records, permission changes and duplicates. Show when the collection was last refreshed. A summary based on a three-week-old snapshot needs to be recognizable, particularly when a recent intervention has changed the equipment's condition.
If the application eventually writes back to the CMMS, require an explicit review step. Record who approved the proposed change and which source supported it. A custom application connected to existing business tools can provide this interface, but the scope should follow an assessment of the actual integration options.
This approach is also relevant when planning a staged ERP modernization. The maintenance pilot should not silently become a replacement program. Give it its own owner, a defined input, a defined output and a way to stop using it without losing access to the original records.
Set limits before the first trial
A convincing summary can still be wrong. Display source excerpts beside important statements and test cases in which two records contradict one another. Instructions appearing inside an imported document are material to read, not permission for the application to perform actions in connected systems.
Keep document access aligned with staff responsibilities. Before sharing records with a model provider, review the information categories, retention arrangements, potential training use, processing locations and deletion options. Work orders may include employee information or confidential production details. Being present in an export does not make every field necessary for the task.
Exclude machine commands, operating-setting changes and return-to-service decisions from this initial pilot. Those involve different engineering requirements. A document assistant should remain an aid; staff must still be able to consult original records when it is unavailable or its answer is uncertain.
Measure whether the pilot is worth extending
Before the trial, time comparable searches using the current process. Build a set of questions with answers checked by a maintenance specialist. Reserve some examples for the final assessment instead of using every example to tune prompts or retrieval rules.
- Time: median time to find a useful record, followed by the time needed to validate the summary.
- Accuracy: share of answers linked to the correct asset and a relevant source.
- Corrections: rejected suggestions and the reasons for rejection.
- Total cost: integration, ongoing usage, data upkeep and review time.
Compare the full task, including correction, rather than generation speed alone. Agree on extension criteria with the people doing the work. Faster record retrieval does not automatically demonstrate fewer breakdowns. That separate hypothesis needs its own observation period and a comparison that accounts for changes in equipment use and maintenance practices.
Review difficult examples as carefully as successful ones. A short note with no asset identifier, two nearly identical part references, or a French abbreviation with multiple meanings can reveal a weakness hidden by an average score. Keep those cases in a repeatable evaluation set for later updates.
Define a first project around a real question
Prepare an anonymized work-order example, a list of the systems involved and one question that is currently difficult to answer. Identify who owns the records and who can evaluate technical accuracy. These inputs support a concrete pilot discussion before deciding whether generative AI is needed at all.
Our article on meeting notes and client follow-up describes another workflow with human review. To discuss maintenance information or another operational process, bring your application and integration requirements to Hamdi Services. The aim is to define a measurable problem and a credible trial, not to promise an industrial result before the evidence exists.