Industrial AI pilots need an exit rule before they multiply
China’s new guidance pairs industrial AI experimentation with warnings against bubbles and duplication. Pilot owners should define evidence thresholds for scale, redesign or stop.
What happened
Chinese authorities issued guidance on 9 October promoting new productive forces, including industrial AI pilots, while cautioning against speculative bubbles and repetitive projects.
Why it matters
A large pilot portfolio can create activity without transferable capability. Clear exit rules protect scarce data, engineering time and frontline attention.
Chinese authorities issued guidance on 9 October promoting “new productive forces,” including industrial AI pilots, while warning against bubbles and repetitive projects. Reuters reported that the guidance seeks broader application while discouraging indiscriminate expansion.
Decide how a pilot ends
Give every pilot a named operational decision, baseline and expiry date. Choose one primary measure that reflects the work: defect escape, unplanned downtime, energy intensity, time to verified diagnosis or another outcome a frontline owner can audit. Add safety, quality and workforce guardrails so a faster process cannot pass by shifting risk elsewhere. Specify the minimum sample, operating conditions and comparison period before data collection begins.
At the expiry review, allow only three outcomes: scale with a documented transfer package, redesign with one unresolved hypothesis, or stop and archive the learning. The package should include data lineage, model and prompt versions, human exception rules, maintenance skills and the cost of running outside the demonstration site. A central portfolio view should detect teams solving the same problem with incompatible measures.
The guidance states policy direction; it does not establish that any individual pilot improves productivity or that one sector’s result transfers to another. Local implementation, incentives and reporting quality can alter outcomes, and anti-bubble language does not guarantee cancellation of weak projects. Use the assessment discipline in China’s AI literacy outcome assessment, adapted to operational outcomes. Track expired pilots without decisions, replicated gains, frontline override rates and engineering effort per scaled use case. The decision skill is portfolio pruning: converting experimentation into comparable evidence and ending work that cannot meet its threshold.