A note on this case study. This is a composite drawn from three engagements with Tier-1 automotive component suppliers in South India between 2025 and 2026. Specific names, product categories, and identifying timeline details have been withheld pending publication consent. The defect pattern, the diagnostic findings, and the coaching arc are reported accurately from across the three engagements; the firm described here is, in that specific sense, real — but is no single one of the three.
Client Overview
The composite firm is a Tier-1 automotive component supplier with three manufacturing plants in South India, a workforce of approximately 1,800, and a portfolio of safety-critical components supplied to global OEMs. Engineering excellence had been the firm's competitive identity for two decades. Quality metrics were strong. The firm's senior engineering leadership — six chief engineers and roughly twenty principal engineers — was widely regarded as one of the deepest technical benches in the regional supplier ecosystem.
In mid-2024, the firm rolled out a suite of AI-augmented engineering tools across its design, validation, and quality functions. The deployment was professionally executed, the change management was thoughtful, and adoption was rapid. By the end of the first six months, the firm was reporting measurable productivity gains in FMEA generation, control plan drafting, supplier evaluation cycles, and design review throughput. The dashboards looked good. Leadership was, justifiably, pleased.
It was at the eighteen-month mark that the problem surfaced.
The Incident
A customer-discovered defect on a safety-critical assembly was traced back, in the subsequent root-cause investigation, to a control plan that had been auto-generated by the firm's AI tooling and signed off by a principal engineer with twenty-two years of experience in the part family. The control plan had a subtle but consequential gap. The principal engineer, on independent review six weeks later, said he could not understand how he had approved it. "I would have caught that, easily, three years ago."
This was not a story about a single careless approval. The follow-up investigation surfaced four further near-misses across the prior nine months — design review approvals, supplier qualifications, and FMEA sign-offs — where senior engineers had accepted AI outputs that they would, in their own judgement, have caught in the pre-AI era. None had reached the customer. All had been intercepted downstream by chance — a junior engineer asking a question, a manufacturing-floor observation, a supplier flagging an inconsistency.
The CEO called us not because of the specific defect — the defect was contained, the customer relationship was managed, the corrective action was straightforward. He called because he had read the literature on AI-assisted work and recognized the pattern. "I do not think this was one bad approval. I think we have a problem with the senior layer that I cannot see from where I am sitting."
He was correct.
The Diagnostic
We ran a four-week diagnostic across the senior engineering cadre, combining structured interviews, a calibrated unassisted exercise (a representative engineering task performed without AI assistance, with results compared against the engineer's AI-assisted baseline), and a review of the prior twelve months of decision journals.
Three findings converged.
First, the senior engineers had stopped articulating their reasoning. In the pre-AI era, when a chief engineer signed off on an FMEA, the sign-off was preceded by an internal rehearsal — what would I challenge if I were reviewing this from the outside? In the post-AI era, that rehearsal had quietly stopped. The AI's analysis appeared coherent; the senior accepted it. The reasoning that they used to do, before sign-off, had migrated out of their workflow. They could not describe, in interview, why they had approved most of the prior quarter's outputs. They had simply approved them.
Second, their unassisted performance had degraded measurably. On the calibrated exercise, the senior engineers' independent quality of FMEA reasoning had dropped to roughly 78% of their pre-AI baseline. The gap was concentrated, exactly as the empirical literature predicts, in catching subtle errors. The engineers could still produce good FMEAs unaided; they were less able than they had been to recognize that an FMEA produced by someone else had a problem.
Third, none of them had noticed. Asked to self-assess their current unaided performance against their pre-AI baseline, the senior engineers' median answer was that it was either unchanged or better. The literature calls this meta-cognitive blindness to capacity erosion, and it is the single most dangerous property of the Oversight Paradox: the senior whose verification function is degrading is not in a position to detect that degradation in themselves.
The diagnostic ended with a finding that the CEO, on receiving it, called "the worst possible kind of finding — the one I cannot argue with." Across his senior engineering cadre, a slow erosion had occurred that the firm's quality metrics, performance reviews, and individual self-assessments had not detected. The defect that had triggered our engagement was not the failure. It was the first visible symptom of the failure.
The Intervention
The intervention was a six-month coaching engagement built on three principles, each drawn directly from the structural recommendations in the AI-augmented work literature.
Principle 1: Deliberate friction at consequential decisions. We worked with the firm's quality function to restructure the senior sign-off workflow on a defined class of high-consequence decisions. Before viewing the AI's analysis, the senior was required to articulate, in a structured journal entry, their own independent first-pass reasoning — what they would expect the failure modes to be, what the highest-risk supplier characteristic was, what would change their mind. Then they reviewed the AI output. The friction was small — adding five to seven minutes to a thirty-minute review — but it restored the rehearsal that had migrated out of the workflow.
Within six weeks, the senior engineers were independently identifying gaps in AI outputs that the pre-intervention version of themselves had been accepting. The capacity had not been destroyed. It had been under-exercised. The friction restored the exercise.
Principle 2: Periodic unassisted calibration. We instituted a defined cadence — one FMEA per quarter and one supplier qualification per month — performed entirely without AI assistance, with results then compared against the AI-assisted baseline. The framing matters: this was not framed as a productivity loss or as a trust exercise. It was framed as calibration maintenance, on the model of commercial pilots required to fly without autopilot for currency. The framing made the practice acceptable to engineers who would otherwise have experienced it as a backward step.
Principle 3: Reasoning capture as the substantive deliverable. The most consequential cultural shift was the slowest. We worked with the quality function and the senior leadership to redefine what the senior's deliverable was. In the pre-AI era, the senior's deliverable had been the approved document. In the post-AI era, with the AI producing the document, the senior's deliverable had become — operationally, if not formally — the approval itself, divorced from the reasoning that had once accompanied it. We rebuilt the role around a different deliverable: the senior's recorded rationale for accepting, modifying, or rejecting each AI output. The output became, in effect, the AI's responsibility. The reasoning became the senior's.
This shift was not easy. The senior engineers, all of whom had built their careers as producers of engineering artifacts, experienced the redefinition as ambivalent — partly as a relief that their judgment was now the centre of their role, partly as a grief for the craftsmanship-identity that had organized their professional life. The grief was not avoidable. It was, however, navigable, and the EQ Catalyst coaching work over the six months gave the cadre the space to do that navigation deliberately rather than reactively.
The Outcomes at Six Months
At the six-month review, three outcomes were measurable.
Unassisted performance recovery. On a re-administered version of the calibrated exercise, the senior cadre's independent quality had recovered from 78% of pre-AI baseline to 96%. The capability had been preserved. The intervention had cost the firm an estimated 3-4% of senior engineering time over the six months and had restored a capability whose loss had been on track to produce, eventually, a customer-visible failure.
Near-miss detection rate. Internal near-misses — issues caught upstream of the customer — had risen sharply, from an average of 1.2 per month in the year prior to roughly 4.8 per month in the six months of the engagement. We initially worried that this was a measurement artefact. It was not. The senior engineers were now catching issues that they had previously been approving through. The throughput cost was negligible. The risk cost was substantial.
Cultural shift at the senior layer. Perhaps the most consequential outcome was the one we had not specifically targeted. The senior engineers, having gone through the experience, became advocates internally for the diagnostic. Two of the chief engineers asked us to extend the engagement to the principal engineer layer below them, on the explicit reasoning that "if we did not notice this happening to ourselves, the engineers we are supposed to be developing definitely will not notice it happening to them either." The intervention propagated downward. The firm now treats verification capacity as a tracked metric, on a quarterly cadence, alongside its other quality indicators.
Reflections
Three observations from this composite engagement that I would offer to any leadership team in the same position.
The defect is not the problem. When the defect that triggered the engagement was discussed in the senior leadership review, the strong instinct was to focus on the specific failure — the specific control plan, the specific supplier, the specific corrective action. The actual problem was not the defect. The defect was the first visible expression of an organizational capacity erosion that had been building for eighteen months. The corrective action that mattered was at the organizational layer, not the document layer.
The senior layer cannot self-diagnose. This is the property of the Oversight Paradox that makes it organizationally dangerous. The verification capacity that is eroding is also the capacity required to notice the erosion. By the time the senior cadre realizes there is a problem, the problem has already produced an incident. The diagnostic must come from outside.
The intervention is not technological. None of the three principles we deployed required any change to the firm's AI tools. The tools remained in place. The tools remained heavily used. What changed was the workflow around the tools — the friction, the calibration cadence, the redefinition of the senior's deliverable. The capability preservation work is human work. It cannot be solved by the next model release.
The firm is now eighteen months past the original incident. The senior cadre is performing at or above pre-AI baseline on independent assessment. The AI tools are producing more output than ever. The two are, deliberately, no longer in tension. That is the outcome the engagement was for.
If your senior engineering layer is operating without a verification-capacity diagnostic, that is itself diagnostic. Our EQ Catalyst coaching method is built for this work. For the broader synthesis across all three organizational layers, see our pillar piece on the AI Anxiety Stack.