AI-Assisted PFMEA, Where the Agent Helps and the Engineer Decides
Pull-quote: “The agent can tell you that a failure mode has no detection control and that the control plan line it points to was deleted last March. It cannot tell you what a hydraulic leak does to the person standing next to the press. Severity is a human number.”
Why this matters
Process FMEA is the document automotive quality lives by and the document most plants quietly dread. Done properly, every process step is interrogated for failure modes, every mode scored for severity, occurrence, and detection, every risk priority number tied to a prevention or detection control, and every control carried through to the control plan that governs the line. Done typically, it is a two-hundred-row spreadsheet assembled in a launch-week war room, frozen at PPAP, and never opened again until the customer audit. The gap between those two states is exactly the kind of sustained clerical vigilance humans are bad at and software is good at, which is why PFMEA is one of the most productive places to put an AI agent, and one of the most important places to draw the line about what the agent owns.
Where the agent genuinely helps
An agent with access to the process flow, the historical FMEA library, and the quality-management corpus does four jobs well. It drafts: given a new process step, it proposes candidate failure modes and causes drawn from how similar operations have failed before, turning a blank-page exercise into an editing exercise. It cross-references: 8D reports, warranty returns, and internal reject codes that correspond to a failure mode the current PFMEA does not contain are exactly the needles retrieval is built for. It enforces structure: every failure mode carries a cause, every cause a prevention control, every detection ranking an actual named control, every special characteristic a control plan line. And it recomputes: when an occurrence score changes after a process improvement, the risk priority numbers, the rankings, and the linked control plan revision all update together instead of drifting apart in three files.
Process step
│
▼
Failure mode ──── effect ───────────► Severity (engineer)
│
├──── cause ────────────────────► Occurrence (engineer, with data)
│ └─ prevention control ─┐
└──── detection control ────────┼─► Detection (engineer, with MSA)
│
S × O × D = RPN
│
▼
Control plan line (agent keeps
the linkage; engineer owns it)Where the engineer decides
| Decision | Why it stays human |
|---|---|
| Severity ratings | Safety and regulatory consequence is a judgment about harm, not a pattern in past data |
| Occurrence ratings | Requires knowing what the process actually does, including what the data does not record |
| Whether a detection control detects | A gage that exists is not a gage that works; that is an MSA question and a floor question |
| Action closure | Deciding a risk is mitigated is an engineering commitment someone signs |
| Special characteristic designation | Contractual and safety weight; customer-facing consequence |
The pattern is consistent: the agent owns completeness and consistency, the engineer owns consequence. A risk priority number is arithmetic, and arithmetic can be delegated. The severity input to that arithmetic is an assertion about what happens to a customer or an operator when the failure escapes, and no retrieval hit legitimizes delegating it. The AIAG-VDA handbook’s move toward action priority tables makes this sharper, not softer, since severity dominates the prioritization logic.
In practice
The working arrangement fits inside quality workflows aligned to IATF 16949 and ISO 9001 expectations: risk-priority calculation stays live, and control plans stay linked to the failure modes they control. The agents draft, retrieve, link, and recompute. Every severity, occurrence, and detection score is entered or confirmed by a named engineer, and the linkage the agent maintains is what makes the annual review a review instead of an archaeology dig.
Closing
The PFMEA that protects a line is the one that is current, complete, and connected to the control plan the operators actually run. Agents are excellent at current, complete, and connected. They have no standing to decide what failure costs. Split the work that way, and the two-hundred-row spreadsheet becomes a living risk model with a human signature on every judgment that matters.