It is 3am. The drive on line A3 has just shut down. The on-call technician has known this machine for three months. The procedure attached to the equipment in the CMMS describes a generic cause linked to cooling, but it dates back to installation and does not account for the two board replacements that have happened since. The expert who could have settled the matter is on leave. The technician hesitates, takes measurements, tests at random. Six hours later, the line restarts. Nobody really knows why. The same fault will come back in three weeks.
This scene plays out every day in French factories. AI-guided diagnosis offers another path: structured reasoning that ranks hypotheses by probability, steers the technician towards the most discriminating tests, and captures every intervention to make the next one faster. Mimorian is an industrial intelligence platform that models equipment, structures fault diagnosis and captures the know-how of maintenance teams through a multi-agent AI architecture.
The stakes are enormous: unplanned downtime costs the world's manufacturers several hundred billion euros a year according to sector estimates, and 23% of this downtime is caused by avoidable human error Source: [Vanson Bourne / ServiceMax, 2017]. This guide sets out the definition, compares AI-guided diagnosis with classic approaches (attached procedure, FMEA), details how it works in practice, and gives the criteria for assessing it before adoption.
What is AI-guided diagnosis in industrial maintenance?
AI-guided diagnosis is structured reasoning that starts from the observed symptoms, puts forward several candidate causes ranked by probability, and recommends the most discriminating tests to confirm or rule out each one. It differs from a simple document search in that it offers a path, not a library. It differs from a general-purpose conversational chatbot in that it relies on a structured model of the equipment, not on a statistical association of words.
The three technical components
AI-guided diagnosis rests on three building blocks that work together.
The functional digital twin models the equipment as a relational graph: each component is linked to its neighbours, its functions and its parameters. It is not a parts index, it is a GPS map of the machine where the AI understands cause-and-effect relationships. When you understand how it works, you understand what goes wrong. To explore how this twin is built from your diagrams, without sensors, read our guide to the functional digital twin for maintenance.
The document RAG (Retrieval-Augmented Generation) consolidates manufacturer manuals, wiring diagrams, internal procedures and safety rules. The language model queries this base instead of inventing a plausible answer.
The specialised AI agents orchestrate the two. One agent extracts the precise technical references from PDFs, a second normalises intervention histories that are sometimes full of in-house jargon, a third checks that each hypothesis respects the laws of physics. They cross-check one another before putting a suggestion to the technician.
The technician's role in the loop
The technician stays in control at every stage. They can rule out a hypothesis by explaining why, and this feedback helps the system refine its reasoning. The interaction happens by voice, in their own words, even with their hands full of grease. The system supports, suggests and structures, but the human remains the decision-maker. This is exactly the "human in the loop" principle required by the EU AI Act for high-risk systems applied to sensitive industrial environments Source: [Commission européenne, AI Act Reg. UE 2024/1689].
To go further on the principles of trustworthy AI in industry, read our complete guide to trustworthy AI for industrial maintenance.
AI-guided diagnosis, attached procedure, FMEA: what are the differences?
Three approaches coexist on the ground today to help a technician facing a fault: the procedure attached to the equipment, the FMEA produced upstream, and AI-guided diagnosis. They are not opposed, they complement one another.
The procedure attached to the equipment
This is the standard CMMS approach. A fixed document, usually a PDF, is linked to a piece of equipment and describes the steps to follow for a given type of fault. Advantages: regulatory compliance, administrative traceability, predictability of the action. Limits: a single line of reasoning, poor adaptation to atypical cases, ageing of the procedure when the equipment evolves or when components are replaced with different references. The intervention report often fits on a single line ("sensor replaced, machine restarted"), which makes it impossible to enrich the procedure with field feedback.
FMEA as an upstream analytical foundation
FMEA (Failure Modes, Effects and Criticality Analysis) maps the failure modes of a piece of equipment and their criticality upstream. It is a valuable analytical foundation, but a static one: once the FMEA has been produced, it lives little. It is not updated at every field intervention, and the lessons learned in the field do not feed back into it.
AI makes it possible to turn an FMEA into a living system: every structured diagnosis becomes an input that adjusts the probabilities, identifies new failure modes and flags critical components as they evolve. In its analysis of the World Economic Forum Lighthouses, McKinsey describes cases of plants that use AI to automate the extraction and updating of their FMEAs from thousands of lines of existing data Source: [McKinsey, 2023: Lighthouses and AI in manufacturing].
AI-guided diagnosis, the third way
AI-guided diagnosis takes the best of both: the structured rigour of the procedure and the systemic view of the FMEA, adding dynamic reasoning that adapts to the real case.
| Criterion | Attached procedure | FMEA | AI-guided diagnosis |
|---|---|---|---|
| Form | Fixed document | Analytical table | Dynamic reasoning |
| Updating | Manual, rare | Manual, project-based | Automatic at every intervention |
| Adaptation to atypical cases | Poor | Moderate | High |
| Trace produced | One line in the CMMS | None | Full structured report |
| Knowledge capture | None | None | Virtuous circle |
AI-guided diagnosis replaces neither the procedure nor the FMEA. It brings them to life by injecting structured field data into them.
How does AI-guided diagnosis work in practice?
The field scenario unfolds in five clear steps, from the moment the technician arrives in front of the faulty machine to the documented closure of the intervention.
Step 1: observation and context
The technician describes the scene by voice. No keyboard entry, no drop-down menu. They use their own words, their workshop jargon, sometimes in French mixed with manufacturer anglicisms. The system understands the context (equipment concerned, production line, ambient conditions) and captures the first symptoms.
Step 2: ranked hypotheses
The system cross-references the described symptoms with the equipment's digital twin and the history of interventions on comparable machines. It puts forward three to five hypotheses ordered by estimated probability. Each hypothesis comes with its rationale: "this hypothesis is likely because this component has already failed under similar conditions on other lines".
Step 3: tests in sequence
The system recommends the least invasive tests first: a voltage measurement before a full strip-down. Each test confirms or rules out a hypothesis and moves the reasoning forward. The technician can at any time say "that's not it, move on to something else, because I already tested X last week". This feedback becomes an input to the model.
Step 4: remedy decided
Once the root cause is identified, the system proposes the remedy with the exact part references, the estimated duration of the intervention, and the applicable safety procedure. The technician validates, carries it out, and confirms the resolution.
Step 5: knowledge capture
The structured report is generated automatically from the technician's dictation: symptoms observed, hypotheses tested, root cause identified, components involved, remedy applied. For a given intervention, the technician saves 30 to 45 minutes of admin a day on their usual tasks. This data immediately enriches the base and makes the next similar diagnosis faster.
To see these five steps applied to a real case, read our example of guided diagnosis on a Schneider ATV930 drive for the augmented technician, where the fault is resolved in 50 minutes instead of 6 to 8 hours of random exploration.
The multi-agent architecture behind the scenes
Behind this smooth experience, several AI agents work in parallel, like an operating theatre where each one has its role. One agent reads the technical PDFs, one agent interprets the historical tickets, one agent checks physical consistency against the digital twin. The orchestrator aggregates their outputs and presents a unified recommendation. This architecture is explained in detail in our article ChatGPT vs multi-agent AI in industrial maintenance.
What concrete gains can you expect from AI-guided diagnosis?
Four families of gains are measurable at sites that deploy AI-guided diagnosis, over a horizon of a few weeks to a few months.
Reduced MTTR (Mean Time To Repair)
MTTR measures the average time to resolve a fault. AI approaches applied to industrial maintenance significantly reduce this time, mainly by eliminating the random exploration phases. Fewer pointless reassemblies, fewer replacement errors, the right part ordered first time. The orders of magnitude published by analyst firms range from 10 to 50% depending on the site and the use case, with a more pronounced effect on complex and atypical faults.
Capture of field know-how
Every intervention enriches the base with no extra data-entry effort for the technician. The structured report is generated from the voice dictation. The loss of know-how when an expert leaves is one of the heaviest hidden costs in the sector: Panopto estimates the cost of this loss at around $47 million a year for a large company Source: [Panopto, 2018: Valuing Workplace Knowledge]. A guided diagnosis that captures every intervention turns this loss into an asset. To go further, read our complete guide to capturing know-how in industrial maintenance.
Upskilling of junior staff
A junior technician diagnoses with the same rigour as a senior from the very first weeks, because they benefit from the structured history of each piece of equipment and from the reasoning passed on by the system. "Study & Learn" sheets are generated automatically from real diagnoses and serve as continuous training material. The time it takes for a junior to get up to speed, traditionally counted in years, is now measured in months.
Field data usable by reliability engineers
Clean diagnosis closures become raw material that reliability engineers and methods teams can use directly. A living FMEA, RCM (Reliability-Centred Maintenance), preventive maintenance plans: all these tools gain in relevance when they are fed by structured field data rather than one-line reports. Recurring weak points become visible. Failure patterns map themselves automatically.
How do you assess AI-guided diagnosis before adoption?
Five concrete criteria make it possible to tell apart a platform designed for industrial rigour from a marketing product that prints "AI" on its packaging.
The five-criteria grid
1. Explainability. Every hypothesis the system puts forward is traced back to its source: manufacturer documentation, intervention history, technical diagram. The technician sees why this hypothesis is advanced rather than another. The whole chain of reasoning stays visible.
2. Adaptation to atypical cases. The system handles faults that are not in the documentation. A part reference replaced by a manufacturer equivalent, a never-before-seen failure mode, a combination of symptoms never encountered: the guided diagnosis must reason, not freeze.
3. Data production. Does the tool enrich the base at every intervention with no extra data-entry effort? If the input data has to be entered manually by the technician in a form, adoption will stay low. Voice dictation and automatic report generation are the markers of a tool designed for the field.
4. Field adoption. Voice interaction, one-click access to diagrams, minimal friction. A technician who has to open three applications to reach the diagnostic aid will not use it.
5. Underlying architecture. A relational graph of the equipment and multi-agent reasoning, not a simple RAG over PDFs. This architectural distinction determines the quality of the diagnosis: a system that does not understand the machine cannot offer solid causal reasoning.
The questions to ask the vendor
Before any adoption, five questions to ask:
- How does the tool handle a fault that is not in the manufacturer documentation?
- What happens when the technician rules out a hypothesis put forward by the system?
- Can the generated report be edited by the technician, signed, archived in the CMMS?
- Is the EU AI Act taken into account in the architecture (transparency, human oversight, traceability of reasoning)?
- What is the time between a team starting the pilot and the first measurable value?
The "industrial ChatGPT" trap
A general-purpose conversational language model, even fed with plant PDFs, does not amount to a guided diagnosis. It generates plausible text word by word, with no model of the equipment, no physical verification, no traceability of reasoning. On a complex fault, the risk of factual hallucination (a non-existent part reference, a procedure not applicable to the exact model, an invented threshold value) stays high. Industrial rigour requires an architecture designed for it. Our article ChatGPT vs multi-agent AI in industrial maintenance explains why.
AI-guided diagnosis and the existing ecosystem: CMMS, MES, ERP
One question comes up in every exchange with an industrial management team: do you have to replace the CMMS, the MES or the ERP to adopt AI-guided diagnosis? The answer is clear: no.
Complementarity, not substitution
The CMMS handles the scheduling of interventions, administrative traceability, the spare-parts stock. It remains the reference tool for these functions. The MES handles production execution, work orders, OEE indicators. The ERP handles financial flows, purchasing and global resource planning.
AI-guided diagnosis fits into this ecosystem without heavy integration. It comes alongside, on the technician's station, and feeds the CMMS on the output side with structured reports that the CMMS can archive and use. Mimorian complements the existing CMMS, it does not replace it.
Bottom-up deployment
The typical deployment of AI-guided diagnosis follows a pragmatic logic. First value in two weeks on a limited scope: a few problem pieces of equipment, one sensitive line. Bottom-up adoption (the technician saves 30 to 45 minutes a day on admin) rather than a top-down mandate. Once value is demonstrated on the pilot, gradual extension to the other lines and sites.
This approach avoids the pitfalls of the large IT programme that ties up several teams for six months before producing any measurable value at all. It rests on the maxim of pragmatic industrial deployment: prove the value on a real case before talking about extension.
Conclusion
AI-guided diagnosis is a third way between the fixed attached procedure and human expertise alone. It preserves the structured rigour of the CMMS procedure, captures the reasoning of the expert who is not always available, and enriches the base at every intervention to make the next one faster.
This third way rests on three layers that work together. The first layer is structural intelligence: a functional digital twin of the equipment, a document RAG, and the back-and-forth between the two that produces a deep understanding of the machine. The second layer is agentic orchestration: several specialised agents that coordinate to put structured reasoning to the technician while keeping the human in the loop. The third layer is the virtuous circle: every intervention enriches the base, identifies recurring weak points, and brings juniors and seniors alike up to speed.
The operational result is an immune system for the plant: every fault resolved creates antibodies against future failures. Knowledge stops being trapped in the head of a single expert, MTTR drops, and field data finally becomes usable for reliability engineers and methods teams.
📚 Sources :
- Vanson Bourne / ServiceMax, 2017 : Unplanned Downtime
- Panopto, 2018 : Valuing Workplace Knowledge
- Commission européenne, 2024 : AI Act (Reg. UE 2024/1689)
- McKinsey, 2023 : How manufacturing's lighthouses are capturing the full value of AI
- ISO 17359:2018 : Condition monitoring and diagnostics of machines, iso.org/standard/71194.html