FMEA stands for Failure Mode and Effects Analysis. ASQ describes it as a step-by-step way to find and rank possible failures in a design, a process, a product or a service. It goes back to the U.S. military in the 1940s.
A failure mode is the way something could fail. The effect is what happens when it does.
Two flavors
- DFMEA (design FMEA) looks at failures that come from the design.
- PFMEA (process FMEA) looks at failures that come from how it's made or done.
Three scores
Each failure gets three ratings, usually 1 to 10:
- Severity (S). How bad is the effect? 10 is catastrophic.
- Occurrence (O). How likely is the cause?
- Detection (D). How hard is it to catch before it reaches the customer? Higher means harder.
Two ways to rank
RPN, the risk priority number, is S × O × D. It ranges from 1 to 1000. It's simple and widely used.
Action priority (AP) comes from the 2019 AIAG-VDA FMEA Handbook. Instead of multiplying, you look up the combination of S, O and D in a table. It returns High, Medium or Low priority for action, and it weighs severity first, then occurrence, then detection. The handbook says an RPN threshold is not a recommended way to decide whether action is needed, because very different combinations can give the same number.
Whichever you use, one rule holds: a severe safety item gets attention whatever its score.
How a bot uses it
Say a bot drafts and sends a weekly email to customers. Here's the start of a PFMEA for that process:
| Step | Failure mode | Effect | S | Cause | O | How it's caught | D | RPN |
|---|---|---|---|---|---|---|---|---|
| Pull customer list | Wrong list | Email to the wrong people | 8 | Old file used | 3 | Person reviews recipients | 4 | 96 |
| Write body | Made-up figure | Customer acts on a false number | 7 | No source checked | 4 | Figures need a source link | 5 | 140 |
| Send | Sent with no review | Any error reaches customers | 8 | Routine set to send | 2 | Drafts only, a person sends | 2 | 32 |
Now the bot can sort by score, propose an action for the top rows, and rescore after the action.
Why bots are good at this
FMEA is tedious. It wants every step, every failure, every score. People get tired and paste boilerplate. A bot can walk every step with the same care, and redo the table after each change.
The catch
A bot can write rows that sound right and don't belong to your product. A generic failure that could be pasted onto anything isn't analysis. Our own practice is blunt about it: every row names the real part, feature or step it belongs to, or it doesn't count.
And the scores are judgments. A bot's "occurrence 3" is a guess unless it's tied to real data. Say so. A person reviews the table, checks the high-severity rows and decides what's acceptable.
Where bots differ
FMEA is meant to be a team exercise, with design, quality, manufacturing and the customer in the room. One bot is not that team. Use it to draft the table fast and completely. Then put it in front of the people who know.
Next: The 80 percent line. Back to Engineering tools for bots.