Generative Leadership Part II: Turning AI-Generated FMEAs into Actual Risk Reduction
Most FMEAs are a waste of time. Not because the method is flawed, but because the output never gets used. Teams will spend days, weeks, or months building and reviewing them, check the box, and move on. The risks are still there and nothing changes.
So the real problem isn’t just how to build an FMEA, it’s how to turn it into something that actually drives action. This is where AI is useful but only if you use it correctly.
I’ve created a template (attached below) for FMEAs that has served me well in countless industries and projects. Whether you’re an ODM, OEM, entrepreneur, or an engineer in manufacturing, utilizing this method can powerfully improve risk anticipation and reduction.
Where AI Actually Helps (and Where It Doesn’t)
AI is not here to make decisions.
It’s here to do the part engineers are bad at: exhaustive, structured enumeration.
For:
Custom equipment (DFMEA)
Real processes (PFMEA)
You’re not limited by understanding, you’re limited by time and mental bandwidth and your human ability to stay true to formats.
AI removes that bottleneck.
It will:
Generate far more failure modes than a team ever would
Maintain structure across large tables
Apply scoring consistently
What it will not do:
Understand cost
Understand practicality
Understand tradeoffs
That part is still on you.
Step 1: Force the AI to Understand Your System
The quality and abundance of information that you train the AI to here will determine the quality of everything downstream.
You need to feed it real content, just make sure you’re using a chatbot that is approved by IT within your organization first.
DFMEA:
RFQ
Requirements
Concept design
Layouts, utilities, safety, maintenance, tooling
PFMEA:
Real process descriptions
Photos of equipment running
Time studies
Value stream maps
Layouts, utilities, safety, maintenance, tooling
Known issues and operator behavior
The goal is simple:
Stop the AI from guessing. Make it operate inside your system.
Example Prompt
Ex:
“You are assisting in building a detailed FMEA.
Use the following scoring criteria for Severity (S), Occurrence (O), and Detection (D): [insert rubric].I will provide system documentation. Your job is to internalize it and generate failure modes specific to this system, not generic ones.
Do not generate output yet. Confirm when ready.”
Step 2: Generate the FMEA in Blocks
If you’re not getting 100+ clean rows in one shot, don’t try to force it.
Most models cap out around 20–30 rows before quality drops.
So control it.
Generate in 10-row blocks.
Example Prompt
Ex:
“Generate 10 FMEA rows using the provided system context.Requirements:
One concise failure mode per row
Fill columns A–H and J
Effects and causes can be up to 5 lines
No generic filler
Label as Rows 1–10. Wait for confirmation before continuing.”
Yes, it’s manual.
However it does have the benefit of forcing you to actually look at what’s being produced.
Step 3: Quick Sanity Check
If you generated 100+ rows, don’t pretend you’re going to carefully read every line right away.
You won’t.
Start with sampling rows at random.
You’re checking:
Does this actually make sense?
Are the failure modes realistic?
Are S/O/D scores even remotely sane?
If the sample looks good, you move forward.
If not, iterate/fix it now before the results become embedded in the workflows we’ll be using the FMEA for.
Example Prompt
Ex:
“Review these FMEA rows and flag:
Unrealistic failure modes
Incorrect S/O/D scoring
Duplicates or overlap
Only return the problematic rows and why.”
Step 4: This Is the Part Everyone Screws Up
You now have a big FMEA.
Now what?
This is where most teams stop and why the whole exercise becomes useless.
You need to extract a clean, prioritized action list.
Not buried inside endless columns. Not scattered across tabs.
One simple list.
Example Prompt
Ex:
“From the full FMEA table:Extract all ‘Recommended Actions’ into a single list.
Requirements:
Deduplicate similar actions
Keep intent, simplify wording
Rank from highest RPN to lowest
Output as: Column A = Action, Column B = Source reference
Do not create new actions.”
Now you have a usable execution list.
Step 5: Bring Real Engineering Back Into the Room
This is where people get reckless with AI.
The outputs will look confident. That doesn’t mean they’re good.
You will see:
Over-engineered solutions
Redundant sensors everywhere
Unrealistic tolerances
Expensive nonsense
Good.
That’s what you want: raw material for decision-making.
Sit down with:
Design engineers
Process owners
Operators
Stakeholders
Flip between:
The action list
The original FMEA rows that generated the action(s)
And decide what actually makes sense and decide if there are follow-up actions that address the underlying failure mode than the emergent recommended action does.
Non-Negotiable Rule
Do not let AI decide:
Who owns the work
How it gets implemented
What it costs
What tradeoffs are acceptable
Step 6: Actually Execute (This Is the Whole Point)
Once actions are validated:
Move them into your project system
Assign owners
Track them
This is the step most teams skip.
And it’s why most FMEAs are exercises that don’t significantly reduce risk.
Step 7: Close the Loop
After implementation:
Record completion of tasks in the Recommended Actions Checklist tab
Re-score S, O, D in the columns K-M, O in the FMEA tab
See if there’s still HIGH or CRITICAL risk that needs to be accepted or converted into action.
Now you can answer:
Did we actually reduce risk or did we just talk about it?
Final Thoughts
AI doesn’t make FMEAs more important.
The real shift in creating FMEAs with AI is usability at scale:
You’re no longer building an FMEA to satisfy a requirement.
You’re building a risk reduction system that actually drives action.
And if your FMEA doesn’t lead to execution, it’s just paperwork.
See my template attached that draws on multiple industries to be as comprehensive and rigorous as possible.
And good luck on your next project!
