A general AI will write you half an FMEA. It won't tell you which half.
Ask a general-purpose AI model for a DFMEA on a simple part and you'll get a good one. Ask it for a DFMEA on a complex product and you'll get a document that is professionally formatted, structurally complete and internally consistent, and carries roughly half the analysis it should.
Nothing is visibly missing. There's no warning and no truncation. The engineer reviewing it has no signal that anything is wrong.
What we measured
We gave the same model the same instructions for two products, on the same day. One was a 13-part commercial flat mop. The other was a 48-part soap dispenser.
Failure modes per component fell from about 4 to 5 on the mop to about 2 on the dispenser. Output doesn't grow with the product. It grows much more slowly, roughly with the cube root of the product's size.
We then took ten of the same parts and analyzed them two ways. On their own, they produced 34 component-level failure modes. Inside the full product, they produced 14: a 59% loss, spread evenly, with no part singled out. The model can do better and doesn't. Analyzed alone, the same parts drew on 119 distinct mechanisms instead of 86.
It isn't running out of room
We tested five explanations and ruled out each one by measurement: the output being cut off, attention fading over a long document, pressure on the output budget, a length target in the instructions, and certain components being neglected. None held up. The model stops with a third of its budget unused. It stops because it believes it's finished.
Our best explanation is a hypothesis, not an established cause, though three separate observations support it. The model applies a document-shaped standard of completeness. It has learned what a finished FMEA looks like, and it produces one, spreading a roughly fixed amount of analysis across however many parts it's given.
If that's right, the behavior is recognisably human. An engineer analyzing a simple bracket knows a two-page FMEA won't pass review, so they dig deeper. On a complex assembly the document is already long, and "this looks pretty complete" arrives before the analysis is complete. The model may have inherited that instinct from the documents it learned from.
Why you can't see it
- It passes every automated check. Schema validation, coverage checks and content scans all pass cleanly on the thinned output.
- There's nothing to spot. No missing section and no error. The only way to notice is to already know what the complete answer should have been.
- It's worst where the risk is. A 200-part appliance is where the exposure lives, and that's where the shortfall is largest.
What we did about it
Telling the model to try harder doesn't work. The fix had to be in how the work is structured, so a large product gets the same depth of analysis as a small one, with a check that nothing is lost when the pieces come back together. On the 48-part dispenser, the same model went from about 93 rows to 335.
That doesn't make the result complete. No FMEA is, whoever writes it. What it does is remove one failure the reviewer had no way to see, and report where coverage ended, so the engineer's review can aim at the gaps instead of hunting for them blind.
The question to ask any tool
"Can a general AI produce an FMEA?" It can, and it will look good. The better question is whether it will tell you what it left out. A general model won't, because it doesn't know it left anything out. On the complex, high-consequence products where an FMEA earns its keep, what it leaves out can be about half.