Robots in a Vietnamese EV plant don't replace welders — they need them: fieldwork shows workers routinely override and reinterpret AI prescriptions, meaning industrial AI's performance depends on embodied, tacit coordination rather than autonomous machines alone.

Main Finding

Industrial AI and digital twin systems in a Vietnamese electric-vehicle welding plant do not simply replace or fully automate human labor. Instead, their functioning depends crucially on human practices of "desynchronization" — workers routinely departing from digitally prescribed instructions to interpret machine outputs and adapt to material contingencies. The author frames this capacity as relational intelligence: the tacit, coordinative skill by which workers align computational plans with ever-changing workplace realities. Industrial AI’s apparent intelligence is therefore emergent from human–machine coordination, not machine autonomy alone.

Key Points

• Corporate narratives present digital twins and industrial AI as enabling synchronized, centrally managed, and predictive factory operations.
• Empirical observation and interviews with welding workers reveal a contrasting reality: workers frequently override, reinterpret, or supplement machine-directed processes to achieve working products and maintain flow.
• Desynchronization: a routine practice where workers move beyond machine prescriptions to handle variability in materials, tooling, fixturing, and other contingencies that models cannot perfectly anticipate.
• Relational intelligence: a concept introduced to capture workers’ ability to translate between computational outputs and practical adjustments — a mix of tacit knowledge, situational judgment, and coordinated action.
• Workers describe their role metaphorically as “robots that run on rice,” signaling both their centrality to production and the precarious, underpaid character of that labor.
• The paper challenges techno-optimistic automation narratives: rather than obviating human labor, industrial AI reorganizes it and often makes previously invisible, undervalued human skills essential to system performance.

Data & Methods

• Qualitative fieldwork in a high‑tech electric-vehicle welding factory in Vietnam.
• Semi-structured interviews with welding workers (focus on their practices, perceptions, and constraints).
• Observational data of on‑shop-floor interactions between workers, robots, and digital systems.
• Discourse analysis of corporate materials and industrial AI narratives (promotional texts, vendor claims, managerial presentations) to contrast official framing with on-the-ground practice.
• Theoretical development (conceptual framing of desynchronization and relational intelligence) grounded in ethnographic evidence.
• Limitations: single-site qualitative study, not intended for statistical generalization but for detailed processual and conceptual insight.

Implications for AI Economics

• Complementarity, not substitution: Industrial AI investments often create high complementarities with specific human tacit skills (relational intelligence). Models that assume routine task automation may overstate labor displacement and understate demand for on-site adaptive skills.
• Measurement and productivity accounting: Standard productivity metrics risk misattributing gains to machines while ignoring the human labor that produces, stabilizes, and maintains AI performance. Economists should measure and value the tacit, coordinating labor that sustains AI systems.
• Returns to capital and skill: Firms may obtain returns from AI investments only insofar as they harness and appropriate worker relational intelligence; this can create firm-specific rents and change the distribution of returns between capital and labor.
• Adoption barriers and diffusion: Successful diffusion of industrial AI depends on local tacit knowledge, training, and organizational practices. Adoption costs include investments in worker training, supervision regimes, and adjustments to production processes.
• Wage and inequality effects: Because relational intelligence is tacit and embodied, it may be undervalued in wage setting, producing precarious labor outcomes even as firms invest heavily in AI — potentially increasing within-firm inequality and limiting wage gains for adaptive shop‑floor skills.
• Policy and firm strategy: Policies aimed at automation-driven displacement (retraining, unemployment insurance) should be complemented by recognition of and support for tacit skill formation, worker bargaining rights, and regulations that prevent the exploitation of labor essential to maintaining AI systems. Firms should design AI as a socio-technical system, accounting for worker expertise in implementation and ROI estimates.
• Research implications: Quantitative models of automation should incorporate complementarities with unobserved human coordination-capacity, and empirical work should seek measures or proxies for relational intelligence (e.g., firm-specific training hours, incidence of on‑line human interventions, rework rates attributable to model limitations).