AI in Automotive Quality — Tool or Mentor? Why „Generate Me an 8D“ Is a Dangerous Sentence

A Wednesday afternoon email lands. „We need this ready by Friday — an 8D for a customer complaint, a PPAP submission for the new part, and an updated MSA Gage R&R because it's been a year since the last validation."

Three artefacts, one team, two and a half working days. You open ChatGPT or another generative tool and type: „Generate me an 8D for complaint type X. Generate me a PPAP package for an injection-moulded part for VW. Generate me an MSA report for Gage R&R."

You get three good-looking outputs. They look like work. They're properly formatted. They go into the customer report, the PPAP submission, the internal repository.

Three months later, an audit. The auditor looks at the 8D and asks: „What was the containment? How many parts were in quarantine? When did you start blocking production?" Then the PPAP: „Where is the customer-specific PSW for VW? Where are the VDA 6.3 P6 process evidences? Where is the IMDS reference?" Then the MSA: „How many operators, how many parts, how many measurements per part? What was the reference standard?"

And you stay quiet. Because the AI gave you an output — but nobody walked you to the evidence.

This is the difference between a tool and a mentor

Generative AI is a tool. You give it a request, it returns an output. Responsibility for whether that output makes sense stays with you. But if you don't know the subject intimately — or you're under pressure and want it done fast — the output looks finished.

A mentor is a different category. The senior quality engineer next to you over coffee doesn't ask what you want to generate. He asks why you need it and what you have for it. Containment? Scope? Customer scorecard for the last three months? Which process changed? What data do you have for the FMEA update?

That's the difference between someone making you paper, and someone walking you to a process.

AI in automotive quality has to be a mentor. Not because it's fashionable. Because in quality, paper without process is more expensive than no paper at all. No paper = no PPAP approved, no launch. Paper without process = approved PPAP, launch, and six months later a complaint, customer scorecard A → B, business lost.

Concretely — what this means for 8D

Generative AI will write an 8D in structure: D0 immediate, D1 team, D2 problem, D3 containment, D4 root cause, D5 corrective, D6 verification, D7 prevention, D8 recognition. It fills every line with text that sounds competent.

A mentor-AI asks at D0: Do you have containment? Parts already in production, in stock, on the way to the customer, at the customer — are they isolated? Without containment validation in D0 and D3, the whole 8D is paper without process. If defective product is leaving for the customer during your RCA, you don't have an 8D — you have an escalation.

The mentor-AI asks at D4: What evidence do you have? Correlations you tested. Data you collected. Hypotheses you ruled out. Without evidence, AI won't absolve you of the fact that root cause is speculation. And speculation comes back in three months.

The mentor-AI asks at D5: Who, what, by when? Without owner and deadline, D5 isn't an action plan. It's a wishlist. Generative AI will fill owners and deadlines with text like „[Owner] / [Date]" or invent names. A mentor-AI asks — give me the names, or I have no way to close this.

That's not nagging. That's methodology.

Concretely — what this means for PPAP

PPAP isn't a document archive. It's submission level + 18 elements + customer-specific requirements. AIAG-VDA PPAP 4th edition defines 5 levels and 18 elements. But customer-specific PPAP for VW (Formel Q New Parts Integral), Stellantis (PPAP through the SCAR portal), Ford (Q1), GM (BIQS) each have their own deviations.

Generative AI gives you the 18 elements in general. It knows customer-specific requirements only loosely — without an explicit input it doesn't know who you're submitting to.

A mentor-AI asks: Who is the customer? VW Group? Tier 1 into VW? Stellantis? What submission level (Level 3 default)? Do you have a PSW? An IMDS reference? A dimensional report from the lab? An MSA for critical characteristics? A process flow per VDA 6.3 P6?

If the AI doesn't know the customer and the requirements, it can't give you a PPAP package that gets approved. It can give you a template — but a template with the customer-specific fields blank is a failed submission.

Concretely — what this means for MSA

An MSA without process variables is mathematically correct and operationally useless. AIAG MSA 4th edition defines Bias, Linearity, Stability, Repeatability, Reproducibility. Gage R&R is typically 3 operators × 10 parts × 2–3 measurements. % GR&R < 10 % excellent, 10–30 % acceptable, > 30 % not acceptable.

Generative AI will compute results from the values you give it. But which values? Without context you can pick any 3 operators, any 10 parts — the answer is mathematically correct and statistically meaningless.

A mentor-AI asks: What characteristic are we measuring? Critical (CC) or significant (SC) per the customer? What is the tolerance range (USL − LSL)? Which gauge? Which measurement principle (contact, optical, automated)? Operators — all trained the same way? Parts — do they span the tolerance range?

Without those questions, the MSA report doesn't hold. Bias can be masked by a systematic offset in the instrument. Reproducibility can be falsely low because two operators measured in different shifts under different lighting. Linearity isn't meaningful on a narrow sample range.

You can't guess this. It has to be answered before the AI runs the first calculation.

A strategy for AI in automotive quality

From what I've seen across 12 years in automotive — Forvia, TRW, HELLA — AI in quality has to honour three principles.

First: domain specialisation. A generic LLM knows IATF 16949 as one standard among thousands. For a quality engineer that means a question like „what does IATF 8.5.1 ask for" returns a general answer without customer-specific context. AI for automotive has to know at least IATF 16949, VDA 6.3, AIAG-VDA FMEA, AIAG MSA, AIAG PPAP, and the major customer-specific requirements (VW Group, Stellantis, BMW, Mercedes-Benz, Ford, GM, Toyota). Without that it's just a faster Google.

Second: structured dialogue. A senior quality engineer doesn't answer immediately. He asks containment before root cause, scope before hypothesis, evidence before conclusion. AI needs the same discipline. „One question at a time" isn't aesthetics. It's protection against skipping a critical step.

Third: incompleteness detection. The single most important property of AI for quality isn't speed of output. It's the ability to stop the output when context is missing. „Without containment status I can't build you an 8D" is a more valuable answer than a nicely formatted page with a speculative D3.

The same logic in the other automotive processes

Quality isn't the exception. The same rule holds in the parallel processes.

Project management (APQP). Generative AI gives you a gateway review with 23 deliverables. A mentor-AI asks: do you have the customer drawing? The cost target? The timing chart with critical path? Without those inputs, the gateway review isn't a decision — it's a ceremony.

Procurement / Supplier Quality. Generative AI gives you a supplier scorecard. A mentor-AI asks: what's your supplier qualification process? Do you have IMDS reference? Conflict minerals declaration? VDA 6.3 audit score? Without those inputs, the scorecard reports the past without preventing the future.

Manufacturing Engineering. Generative AI gives you a DFM review. A mentor-AI asks: what does PFMEA say about this process? What's the takt time, OEE, FTQ from similar lines? Without those inputs, DFM stays a designer's exercise without manufacturing reality.

Service / Field. Generative AI gives you a failure analysis report. A mentor-AI asks: do you have field returns data by batch / VIN? A Pareto chart of top 5 failure modes? Correlation with production date? Without those inputs, failure analysis is a story without data.

In every case the difference is the same. A tool speeds up the output. A mentor improves the process. Long-term, the difference for the company is dramatic.

The strategic thesis for the next 12 months

The future of AI in automotive — not just in quality — is not about faster outputs. It's about better questions.

AI that lets you be lazy does you a disservice. It makes you a paper that looks like work, and breaks your customer scorecard in six months.

AI that forces you to be disciplined makes you a better quality engineer. And, more broadly, a better engineer overall — because the same disciplinary questions apply across PM, supply, ME, service.

QualityOS was built on this principle. An AI assistant for quality engineers that asks containment before root cause, scope before hypothesis, evidence before conclusion, owner before action, deadline before closure. That has no motivation to compliment you on your question or add bonus tips beyond your situation. That can stop the output when context is missing.

That's the difference between „generate me an 8D" and „let's prepare together an 8D that survives audit and complaint". The first is a tool's sentence. The second is a mentor's.

In automotive, that difference decides whether your paper holds up under a complaint — or your processes do. And I'd rather the processes did.