Shadow AI isn't a governance gap. It's an uncalibrated input.

The Smarsh study landed last week confirming what most of us already suspected: shadow AI is outpacing enterprise governance. The default reading is data exfiltration — IP leaking out through consumer LLMs, trade secrets pasted into chat windows. That's the IT-security frame, and it's not wrong. But it's incomplete. I've spent two decades in plants where the real damage isn't what leaves the building. It's what enters the quality system untraceably. An engineer pastes a PFMEA worksheet into an unvetted LLM, then pastes the polished output back into an 8D report. That's not a data-loss event. It's an uncalibrated input contaminating a controlled document. Same failure mode as using an uncalibrated micrometer. Different reading. Different fix.

What shadow AI actually looks like on the shop floor

It doesn't look dramatic. It looks efficient. A quality engineer under pressure to close an 8D within 72 hours drafts the root-cause narrative in ChatGPT because it writes faster than they do. A maintenance technician feeds vibration data and error codes into a free LLM to narrow down a fault on a stamping press. I've traced both in audits. Then there's the supplier who generates their PPAP narrative — dimensional results, capability indices, the lot — using a model nobody has vetted, and submits it as though it were authoritative. The outputs land in 8D reports, control plans, supplier correspondence, and occasionally in risk assessments that drive engineering changes. Untraceably.

My CEH training taught me to find shadow systems — rogue access points, unauthorised VPNs, services running on ports nobody documented. Shadow AI is shadow IT with higher stakes, because unlike a rogue wireless bridge, the outputs don't just create a network vulnerability. They enter quality records. They become part of the evidentiary chain. When an auditor asks how you arrived at a particular severity rating in a PFMEA, and the honest answer is "an LLM suggested it and we didn't verify the reasoning," you have a problem that no firewall rule will fix.

The AEI piece on America improvising its way through AI governance captured the regulatory void well, but the plants I work in can't wait for regulation. The TraceGains readiness survey circulating this week shows food-sector quality teams hitting the same pattern. Supplier quality crises don't pause for governance frameworks.

Why blocking makes it worse

Here is an uncomfortable observation. Every CISO who blocks ChatGPT at the firewall believes the problem is solved. It isn't. Usage migrates to personal devices on cellular networks, and now you've lost even the residual visibility you had. I've watched this happen with shadow IT for two decades. The pattern is invariant: when the sanctioned path is slower or less capable than the shadow path, people take the shadow path. They're not being malicious. They're being practical. The QRQC is due. The customer escalation is live. The 8D clock is ticking.

The governed channel has to be faster than the shadow alternative. Not slower with a badge that says "approved." Faster. That's the design constraint. This is why I built MultiPS the way I did — not as a single-model gateway with a polite logo, but as a multi-model orchestration platform running 63-plus models in parallel with a consensus synthesis layer. The consensus layer was not a feature. It was a control mechanism. When three independent models converge on a root-cause hypothesis and a fourth diverges, that divergence is flagged, logged, and routed for human review. When they agree, you get a defensible chain of reasoning — not a single stochastic guess that nobody can reproduce.

Every query is logged. Model versions are captured. Outputs get hashed. That's the audit trail. Without it, you're asking an auditor to trust a memory of what prompt was entered into which version of which model on what date. Good luck defending that under AS9100 or IATF 16949.

The instrument you can't calibrate is the one you can't trust — and AI you can't trace is the instrument you've already lost.

The calibrated-tool model

Operators don't bring their own micrometers to the line. You provide calibrated instruments with traceability certificates, regular verification intervals, and a documented chain from national standards to the shop floor. If a technician showed up with a micrometer from a website of unknown origin and started measuring critical characteristics with it, you'd stop production. Not because the micrometer is necessarily wrong — it might be fine. But because you have no basis to trust it, no calibration record, no traceability, no control over its drift.

AI inputs to quality decisions need the same regime. Model version. Prompt log. Output hash. Human sign-off. Consensus validation where the decision is non-trivial. This isn't bureaucratic overhead — it's the difference between an input you can defend in a recall investigation and one you can't. The cost of poor quality data has been visible for years; the Grattan Institute's work on pharmacy data quality this week is just the latest reminder that bad inputs propagate silently until they surface as failures. In aerospace, the surface area for that propagation is larger and the consequences are graver.

Safran's emphasis this week on the supplier performance manager role tells you something: supplier-generated content is already a vector. When suppliers use shadow AI to generate PPAP narratives or capability studies, the contamination enters your quality system through the front door, with a cover letter and a signature. The governed channel has to extend outward — not just internally, but into the supplier base.

Shadow AI won't be solved by policy documents or firewall rules. It gets solved when governed AI is demonstrably better than the alternative — faster, more capable, more useful — and when it follows the same discipline as every other calibrated input to your quality system. The engineers using shadow tools aren't the problem. The absence of a governed tool worth using is. Build the channel. Calibrate it. Log everything. Then the shadow path stops being the practical choice and starts being what it actually is: an uncalibrated instrument with no place on the floor.