The future of FMECA.
By 2026, AI has become impossible to ignore - and for an FMECA it's a genuine help. It takes on what used to mean weeks of reading manuals and filling spreadsheets, building the same analysis in days, sometimes hours. And the real surprise isn't the speed: it's that, done right, the work often comes out more thorough than it would by hand.
I say done right on purpose because... garbage in = garbage out. Treat the AI like a smarter Google and you'll get a thin answer and walk away disappointed. Set it up properly - method, context, sources - and it's genuinely high quality, for three reasons. It stays consistent from first line to last, it reads your whole failure history (no single engineer holds that in their head), and every line is grounded in real data from your own site (at Levels 2 and 3).
Rather than promise, I'll show you: we'll follow one use case the whole way - a 200 V DC electric motor - across the three main ways to do this, from a quick chat to a grounded, fleet-scale analysis.
The three levels at a glance
|
1
The Amateur
Level 1 · in the chat
|
2
The Intermediate
Level 2 · a Project
|
3
The Expert
Level 3 · Claude Code
|
|
|---|---|---|---|
| What it is | A normal chat, sharpened with meta-prompting. | A Project with your /fmeca method + the asset's manual & CMMS history. |
Claude Code agents running /fmeca across folders, with a catalogue, read-across & an auditor. |
| What it's for | Ideation & a first draft - beating the blank page. | A sourced, decision-grade FMECA for one asset. | Recurring programs at fleet scale, straight into your CMMS. |
| Documents | None. | A handful - the manual + its history. | Many, across the whole fleet. |
| Assets | One, in theory. | 1 to 10. | Dozens or more - a fleet. |
None of this is meant to replace your SME's final review - that call stays with your experts. The point is the time you save. Who wouldn't want to cut the time it takes to build an FMECA by 80 %, when everyone knows it's usually a slow, tedious grind? The AI takes the heavy, repetitive 80 %, so your specialists spend their hours where they count - judgment, the tricky edge cases, and signing off on the last 10 to 15 %. The AI is the leverage; the expert still owns the call.
The Librarian.
Before you start
AI can feel like a black box - sometimes brilliant, sometimes confidently wrong. A librarian is the simplest way to see why, and how we keep it honest.
Think of the AI as a librarian. Your question is the request. Your manuals and maintenance history are the books on the shelves. A failure mode, a wear limit, a past repair - each one is a real page, somewhere in those archives.
The desk matters too. A librarian can only read what's open on the desk at once - that's the AI's context window, its working space. The shelves hold far more than fits there, so part of the job is carrying the right pages over. A bigger desk - a longer context - lets it keep more open and cross-reference it.
Here's the catch: left alone, an AI hates saying I don't know. Ask for a page it can't find and it will often make one up - a confident answer with no book behind it. That's a hallucination: an invented failure mode, a number from nowhere. A bad librarian handing you a fake.
The whole method turns it into a good librarian. You hand it the real documents and set one rule: use only what's written, and cite where you found it. Now every line traces back to a real page. That discipline has a name - grounding - and it's what makes the result trustworthy.
From there, only one thing changes between the three levels: how big a library the librarian can reach.
The desk
= the chat
Just a desk: the AI works from whatever you put on it. You can drop a few papers, but there's no real collection behind it - so it leans on memory, and it's the most likely to invent.
One shelf
= a skill, in a Project
One shelf for your asset: its manual and its history. Now the librarian fetches the real page - and cites it.
The whole library
= skills + subagents, in Claude Code
A whole library, plus assistants. They search across the fleet at once, and a second reader checks every citation.
The golden rule The highest level is not always the best one. Pick the lowest level that solves your pain. For most single-asset work, Level 2 is the sweet spot - one asset, its documents, a sourced result. Level 3 only earns its keep when you have a fleet and a recurring program.
Level 1 - no data yet
The Amateur
You ask the AI in a normal chat. The twist that already sets you apart: don't write the prompt yourself - have the AI write it for you (it's named meta-prompting). You get a far better prompt than you'd have typed, and you learn what a good one looks like.
The move - meta-prompting
Free tips
- Ask it to interview you first. "Before you answer, ask me the 8 questions you most need answered to make this FMECA specific." Suddenly it's working with more context, not a stereotype.
Trade-off
- Zero setup, instant, perfect to beat the blank page and get a structure going.
- Generic modes; it invents the S/O/D ratings; it doesn't know your equipment or its history.
How to
Go to claude.ai (or any other LLM - ChatGPT, Gemini, Mistral...), open a new chat, and type this:
Animated demo: the meta-prompt is typed into claude.ai, then run back to generate the first generic FMECA.
Animated demo: the sharpened prompt is pasted back and run to generate the first generic FMECA.
Level 2 - one asset, its documents
The Intermediate
This is where it starts to get genuinely interesting and relevant. The idea is simple: we give the AI two things - a method and some data. The method is the recipe: it fixes how every analysis is done. The data is the ingredients: what that recipe is applied to.
Concretely, the recipe is a reusable skill - the /fmeca skill - that fixes your columns, scales and rules. The ingredients are the documents only you have about the asset - for our example motor, that's its manual and its CMMS history: the work orders, downtime and cost. With both in place, the AI knows how you work and what it's working on. Here's how to set it up.
Not sure what a skill is? Read the quick guide →
Download the /fmeca skill
This is the recipe - one short text file that fixes your columns (function, mode, cause, local & final effect, controls, source, S, O, D, RPN, action), your scales, and the rules - every line cites a source, no invented numbers, mode ≠ effect. One click, no sign-up.
Download the free /fmeca skill
On Claude Code? Drop it straight into your skills folder and skip steps 2-4.
Create a Claude Project
Open claude.ai → Projects → New project, e.g. Super FMECA. A Project keeps your method and your documents in one place, reused on every chat inside it.
Add the skill to your project
Add the /fmeca skill you just downloaded to the project knowledge. That's your method in place - reused on every chat in the Project, so every analysis is done the same way.
Upload your documents
Add what only you have - mainly the equipment manual (maintenance sections, wear limits) and the CMMS history (work orders, parts, downtime, cost), but anything relevant helps: a criticality matrix, P&IDs, past reports and - above all - any RCAs you have. Raw exports are fine - let Claude make sense of them.
Generate your FMECA
Everything else lives in the skill - so the ask stays short. Name the skill, point it at the asset, and let it use the documents you just added.
Trade-off
- Realistic, sourced modes (grounding); Occurrence calculated from real history when the data's there - not guessed.
- Reusable method - the same skill gives you a consistent analysis on every asset you point it at.
- You gather and tidy the documents first - and decide what's safe to upload.
- One asset per chat - fine for a handful, but it doesn't scale to a whole fleet.
Level 3 - a whole fleet, recurring
The Expert
Now this is where it gets really interesting. The method stays broadly the same - with a few differences - only now it's run across a whole fleet, with the AI doing more of the work. You move from the chat window to a different interface: Claude Code.
Don't let the word "code" put you off - it's not really harder, just a knack to pick up, and the payoff is big. The practical change is this: nothing has to be moved or uploaded by hand anymore. Your documents stay where they already live - in folders - and Claude Code goes and finds them itself: it browses your folders, opens the right files, picks up the /fmeca skill on its own, and genuinely thinks further to push the analysis deeper. Reading straight from your folders, it can also take on far more at once - a whole fleet, not one asset at a time.
Same method, far more power. On top of that base, three extra pieces take it further - each one delivered either as a skill Claude Code runs, or as a folder of documents it reads:
- Catalogue (a data asset). One file per equipment type listing every confirmed mode, cause, control and source. A new DC motor starts at ~80 % before reading a single document - and it grows with every study you validate. This is exactly the kind of library we've spent time building at Rob Reliability, so your study can start most of the way there.
- Read-across (a skill). A fleet-analysis skill that reads the history of other similar motors to surface modes yours hasn't shown yet - e.g. a flashover seen on M-101, a bearing seizure on M-301 - flagged "to confirm", with the source asset cited. Nothing is silently invented.
- A built-in auditor (a second skill). Launch all 50 motors in one prompt; the
/fmecaskill keeps every study identical, while a separate auditor skill - not the one that built the study - reviews the output for missing sources, mode/cause/effect confusion, or numbers without a basis, before any human opens it.
And the result lands as a clean Excel file - ready to open, or to import straight into your CMMS or FMECA tool, with zero copy-paste. Prefer a PDF, a PowerPoint, something else? It can produce just about any format you ask for.
Best for: recurring programs · many sources · fleet scaleHow to
Open Claude Code
Best is to download the Claude Code desktop app - the simplest route, since it gets direct access to your local folder. Open it and pick the Code tab up top.
You can also run it from a code editor like VS Code or Antigravity - a bit more involved, so we'll leave that aside here.
✵ What's up next, Rob?
What's newYou've used ~24× more tokens than War and Peace.
Pick your folder
Point Claude Code at the one folder holding everything - your data (manuals, CMMS history, RCAs, catalogue) and your skills (the /fmeca skill and the auditor). No uploading: the files already live there, so it finds them itself.
✵ What's up next, Rob?
What's newRun it
Now just ask. One prompt runs the /fmeca skill on every asset in the folder - starting from the catalogue, reading each manual, CMMS history and RCA, using read-across to flag sibling modes - then writes one grounded Excel file per asset.
Audit & export
Before any human opens a thing, a separate auditor skill reviews all 50 FMECAs - catching missing sources, mode/cause mix-ups or numbers without a basis. What's left is a clean Excel file per asset, ready to open or import straight into your CMMS - zero copy-paste.
And that's the whole thing. In a morning - a day at the very most - you've got your 50 professional-grade FMECAs: consistent with one another, each one cross-referencing your full equipment history. All that's left is a final read-through by your SME. The time saved is enormous - that's really all there is to it.
What keeps it trustworthy
AI that drafts an FMECA is easy. AI you'd sign your name under is the point.
Applying AI to the FMECA - who does what
- Reads every document in full
- Drafts the modes, causes and effects
- Cites a source on every line
- Calculates Occurrence from your history
- Proposes a Detection rating
- Audits its own output
- Set Severity from your site matrix
- Confirm or correct the Detection
- Own the risk appetite
- Sign it off
That's the whole point of AI applied to reliability: it drafts and grounds, you judge and decide. It never invents a number - every value traces back to your data or your call.
The three levels at a glance
Worried about data security, quality or hallucinations? I answer the most common worries in Questions you're probably asking, a little further down.
Going further
What RAG is
You'll hear a lot about RAG when AI meets your documents. It's just one way to give the AI access to what you have - not the only one, not a magic one. Here's how it works, and the one catch.
RAG (Retrieval-Augmented Generation) chops your documents into little cards and indexes them. For each question, it pulls the handful that look closest and hands just those to the AI - so it never reads everything. Clever and efficient.
But that's also its catch: reading only a few cards, it can miss what sits outside them. For "find the page about X" it's excellent. But ask something that needs the whole history - "which failure mode caused the most downtime?" - and it can answer confidently from the cards it pulled, while the real answer sat in one it never opened.
"Which failure mode caused the most downtime?"
It retrieves the 5 cards closest to your question and sums those - so it can confidently name the wrong mode, because the real maximum sat in a card it never pulled. The error is invisible.
Asset by asset, read everything, then score.
Reading the whole file sees the entire history, so totals, maxima and frequencies come out right. For a single asset this is both simpler and more correct.
Why RAG can miss the answer
RAG · retrieves a few cards
Pulls the 5 cards nearest the question. The real maximum ★ sits in a card it never reads - wrong answer.
Whole file · reads everything
Reads all 24 cards, so it actually finds the maximum ★ - right answer.
Each square is a slice of the failure history. "Which mode caused the most downtime?" needs all of them, not the nearest few.
And agentic RAG?
Agentic RAG is a smarter version of the same idea. Instead of one blind grab, the AI reads actively: it sees what came back, notices what's missing, and goes again until it has enough to answer. Less five random cards, more a librarian who keeps walking back to the shelves.
You keep RAG's efficiency on a large library but close most of the gap on questions that need the whole picture - which is why, at fleet scale, it's often the right tool. And if your data is sensitive, the same setup runs on your own local servers or a private environment, so nothing leaves your perimeter.
Just keep in mind RAG isn't bad - it's different, built for scale, and plain retrieval can miss what it didn't pull. For a single asset, reading the whole file is simpler and safer; RAG and agentic RAG earn their place once the library is too big to read in one go.
Not tied to one tool. None of this is Claude-specific. We run it here with Claude - a Project for one asset, Claude Code for a fleet - but the same whole-file, RAG and agentic-RAG ideas apply just as well with OpenAI or Google Gemini.
Before you start
Questions you're probably asking yourself
Before you point AI at your plant data, a few worries always come up. Here's how I'd answer them - straight.
How do I keep my data safe?
Fair question - it's the first one I get. A few honest options, from simplest to most locked-down; pick the one your team is comfortable with.
The method here is the same whichever you choose - only the tool changes.
Isn't the quality shaky - what about hallucinations?
Honestly? No more than a normal engineer on an off day - and we design against it. (A "hallucination" is when the AI says something that sounds right but is made up.)
Not magic, not perfect - which is why you keep the final call. It gets you most of the way, fast; you bring the judgment. Put those three guardrails in place once - sourced, computed, audited - and you stop re-reading every line just to trust it; I promise that's where the hours come back.
How long does it actually take?
Far less than by hand. One asset (Level 2) is an afternoon the first time, then minutes once your skill and documents are set up. A whole fleet of 50 (Level 3) lands in a morning - a day at most. The slow part was always the reading and the typing; that's exactly what goes away.
Will it work with my data and my CMMS?
Yes - no clean-up ritual first. Raw exports are fine, whether it's SAP, Maximo, a scanned PDF or a messy spreadsheet; let Claude make sense of them. The result comes back as an Excel file (or any format you ask for), ready to import straight back into your CMMS.
Do I need to be technical?
Not for Levels 1 and 2 - if you can write an email, you can do this. Level 3 needs a little more setup, and that's the kind of thing you'd hand to someone like us.
Take it with you
Same method throughout - the only difference is how much of your context the AI gets to read: a chat, a Project, or a whole fleet in Claude Code.
If you take one thing away: start at Level 2. One asset, its manual and history, the /fmeca skill - that's where most of the value is, and it's an afternoon's work the first time.
One more step
Your own FMECA app
Want to go even further? Your own FMECA app - every asset, plugged into your CMMS, all in one place:
In the AI era you don't have to settle for off-the-shelf software - you can build the exact tool you need, fit to your own process.
That's exactly the kind of thing we do - if you need it, we can help you build it.
Build it yourself, or have us do it
/fmeca skill and follow this guide. Download the skill →Hope this helps 🙂 - Rob Reliability