AI in manufacturing: factory and supply-chain benchmarks (with sources)

Manufacturing has the largest published AI gains and, at the same time, the widest gulf between pilot and scale. And the bottleneck has a name: data and legacy systems. This post shows both sides, sourced, and why legacy is the central issue.

About 90% of companies already use AI in at least one function, but two-thirds remain stuck in the pilot phase, capturing no financial return, per McKinsey. Only 6% become "high performers" with meaningful EBIT impact. Near-universal adoption, near-absent scaled ROI.

Where AI already delivers in manufacturing

Application	Result (sourced)	Source
Factory transformation (aggregate)	+50% productivity, −80% defects and −30% CO2 on average	WEF Global Lighthouse Network
AI-specific use cases	−41% defects, −28% energy and −44% cycle time	WEF Lighthouse (recent cohort)
Specific plant	OEE raised to 88% and unit cost −41%	EVE Energy Jingmen (WEF)
Demand forecasting	20 to 50% reduction in forecast error	McKinsey
Quality and quoting (Brazil)	Quote accuracy at 95% and related cost −65%	Gerdau (Azure AI)

The World Economic Forum's Global Lighthouse Network, run with McKinsey, is the most reliable benchmark set for AI at scale in factories. Among named cases, the Zhengzhou Coal Mining plant raised output per worker by 205% and cut lead time by 66%. This is not a lab: it is audited factory floor.

Where AI broke in manufacturing

Tesla: Elon Musk admitted "excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated." The over-robotized Model 3 line capped output below half the target and had to be torn out.
Pilots that never scale: McKinsey shows two-thirds of manufacturers stuck in pilots; RAND finds AI projects fail to reach production at a rate well above ordinary IT projects.
Low digital-transformation success: a BCG study of 850+ companies found only about a 35% success rate, and a 2024 survey has 65% of manufacturers calling themselves "laggards".

The real bottleneck: data and legacy

Here is the cause behind the failures. About 70% of manufacturers say data problems, such as quality, context and validation, are the single biggest obstacle to AI (Deloitte, 2025). And roughly 60% of AI leaders cite integration with legacy systems as their primary challenge in adopting agentic AI. McKinsey sums it up: scaling failures come from data readiness and organizational alignment, "rarely technical alone". Before plugging in AI, you must fix the legacy foundation, or the project dies in the pilot.

In Brazil

AI is considered essential by 69% of industries per the CNI, and Brazil's national AI plan earmarks R$9.4 billion for industry challenges. In concrete cases: Gerdau raised quote accuracy to 95% and cut related costs by 65% with Azure AI, and saved about $3 per ton of steel with machine learning; Embraer built a "Smart Planning" system to optimize parts inventory and supply-chain agility.

The lesson for anyone implementing

Industrial AI fails first at the data and legacy layer, not the algorithm. Fixing that foundation is the precondition for Lighthouse-grade gains. That is exactly where Reche's Legacy Code Diagnosis comes in: before investing in AI, map what is blocking it, so the project does not die in the pilot the way two-thirds of them do.