Manufacturing Needs Black Boxes, Not Just AI Models
Industrial operations still rely on manual data entry while aviation records every second—and that gap is crippling efforts to deploy effective physical AI.
Manufacturing's data problem
When a commercial flight encounters trouble, investigators reconstruct events second by second from flight data recorders. When a manufacturing line produces a defect, quality teams ask workers what they remember about Tuesday.
This gap represents the binding constraint on industrial AI, according to analysis first reported by Natan Linder and Gilad Langer at Forbes. The problem isn't model quality—it's that AI systems are learning from records that were never designed to capture operational reality.
Most manufacturing facilities run on what the authors call "systems of assertion": MES, QMS, LIMS, and warehouse management platforms that record what operators say happened, not how it happened. An operator confirms a step, a technician enters a result, a supervisor signs off. These systems capture the what—a quality check failed at 10:02, a line stopped for eleven minutes—but never the how.
Why it matters
The manufacturing industry faces a projected shortage of up to 1.9 million workers through 2033, according to Deloitte and The Manufacturing Institute. Capturing how experienced workers actually perform tasks—including all the undocumented workarounds and improvisations—becomes critical before that institutional knowledge walks out the door. More immediately, regulated manufacturers investigating batch deviations are required to determine root cause but are equipped only with after-the-fact assertions typed into predefined fields. Physical AI trained on incomplete operational records will produce incomplete, potentially dangerous recommendations.
The shift to observation
The alternative is what Linder and Langer term "systems of observation"—using sensors, cameras, and computer vision to continuously record what actually happens on the factory floor. This approach would generate verifiable, time-aligned operational data without requiring workers to stop and document every step.
The shift faces substantial obstacles. Existing MES vendors are growing roughly 30 percent annually, according to LNS Research data cited in the Forbes piece. These systems are deeply entangled with validated states and regulatory commitments. Tom Comstock at LNS Research noted that some vendors are attempting to make MES "composable," but breaking a monolith into vendor-controlled modules isn't true composability.
More fundamentally, LNS found that MES companies "do not see the need for or demand for" connected frontline worker applications. A category that doesn't prioritize the person doing the work won't be positioned to learn from them.
The data synthetic training can't replace
Jensen Huang acknowledged at CES 2026 that real-world data collection remains slow and costly, driving massive investment in synthetic training data. But simulation can only generate scenarios already modeled. The critical learning happens in deviations—the fixture modified in 2019 for forgotten reasons, the workaround an experienced operator developed, the process drift that predefined schemas never anticipated.
Real operators performing actual work represent the one dataset that cannot be synthesized, specifically in how their work departs from documented procedures. That makes frontline workers the signal, not the automation target.
Regulatory and ethical constraints
The EU AI Act, effective February 2025, prohibits inferring worker emotional states from biometric data, with penalties reaching seven percent of global revenue. GDPR pushes toward anonymization while GxP data integrity requirements demand attribution. In Germany, works councils hold co-determination rights over technical monitoring introduction.
The principle, according to the Forbes analysis: systems cannot only extract data from operations—they must return value to the workers being observed, or risk becoming surveillance rather than support.
Linder and Langer acknowledge the technology for fine-grained temporal reasoning on manual work isn't yet reliable, consent frameworks remain undeveloped, and no validated ROI data exists. They describe this as a decade-long transformation, not a quarterly initiative.
The analysis was co-written by Natan Linder and Gilad Langer and originally published in Forbes.
This is an original analysis by the Omega editorial team. Source reporting: AI Watch.
Want systems like this working for your business?
Book a Call
