Lean-First AI Slashes Waste and Costs in Medium-Scale Injection Molding

Injection molding plants that sit in an awkward middle ground—too complex to run on whiteboards and clipboards, too cash-strapped for full Industry 4.0 automation—may have a new playbook. A study published in Heliyon reports that a deliberately simple, lean-governed artificial intelligence system deployed at a US-based molding facility cut scrap by 67 percent, unplanned downtime by 48 percent, and changeover times by 41 percent, all without a single new sensor or piece of capital equipment. The annualized savings reached roughly half a million dollars, with a payback period of under three months.

The research, led by Chirag Thummar, tackles a problem that has long frustrated mid-sized manufacturers operating between 10 and 50 presses. At this scale, the study argues, the dominant source of waste is not poor molding engineering but decision latency: the gap between when a decision is needed and when it is actually made with adequate information. Slow quotes, fuzzy scheduling, and unranked maintenance priorities force managers to compensate with excess inventory, overtime, expedited freight, and redundant inspection—each buffer adding cost and instability across the value stream.

Lean manufacturing tools such as 5S, standardized work, value stream mapping, Kanban, SMED, and total productive maintenance have long delivered gains in molding environments, including changeover reductions of 39 to 54 percent reported in prior studies. But in high-mix plants, lean implementations tend to plateau once decision complexity exceeds what standard work and visual management can absorb. The new study’s central claim is that this is precisely where AI belongs—not as an autonomous optimizer, but as a constrained decision accelerator operating inside a lean-defined architecture.

The system was tested at a facility running 33 presses between 55 and 720 tons of clamp force, supplying aerospace, automotive, and medical customers under ISO-certified quality systems. The plant juggled roughly 450 active stock-keeping units, frequent mold changes, and volatile order volumes across two to three shifts. At baseline, quote turnaround averaged 48 hours, schedule adherence sat at 78 percent, overall equipment effectiveness was 58 percent, and work-in-process inventory stood at 12 days.

The architecture has three layers. The first, a lean foundation, establishes operational stability through 5S, standardized work at each press, visual controls, and daily tiered management routines—while simultaneously creating a disciplined data-capturing environment. The second, data enablement, builds lightweight pipelines from existing ERP and MRP exports, downtime logs, maintenance tickets, and quality records, cleaning and time-stamping everything without sensors or IoT devices. The third layer provides interpretable AI outputs in three domains: quote generation via similarity-based retrieval across part family, material, and complexity; finite-capacity scheduling with setup-aware dispatch; and loss prioritization through Pareto-ranked downtime, scrap, and delivery signals.

Notably, the AI modules are rule-based and weighted rather than machine-learning models, a choice the author defends on three grounds. The facility’s roughly 450 SKUs and inconsistently coded records could not satisfy the data-to-feature ratios that gradient-boosted trees or sequence models demand without overfitting. Interpretability was a hard governance requirement, because every AI recommendation had to be accepted or overridden by a human at two formal gates, and overrides must be traceable to specific features. And when upstream lean layers already shrink the decision space, the marginal value of algorithmic sophistication drops sharply—a view consistent with prior work on the cultural tensions of AI adoption in lean quality management.

Two governance gates controlled the flow of orders. Gate 1, Quote Review, required documented routing assumptions, cycle-time estimates, and tooling requirements before any quote was released, with the AI supplying an analog-based recommendation that a quote engineer confirmed or overrode. Gate 2, Schedule Release, verified data completeness, downtime coding consistency, training readiness, and system validation before an order reached a press, with the scheduler explicitly confirming or overriding the AI’s ranked dispatch advice. Overrides were logged and used to recalibrate the model weights monthly, while a feedback loop routed scrap, downtime, and delivery signals back into daily improvement routines.

The results, drawn from 25 paired production lines each with a distinct press-mold-part-family combination, were striking. Cycle time fell from 47.95 seconds to 34.68 seconds, a 27.7 percent reduction. Scrap dropped from 6.46 percent to 2.11 percent, bringing the facility below its 3 percent internal target. Unplanned downtime fell from 2.65 hours per shift to 1.37 hours, and changeover time dropped from 78.48 minutes to 46.36 minutes. All four improvements were statistically significant after Bonferroni correction, with Cohen’s d effect sizes between 0.85 and 6.07—every one exceeding the conventional large-effect threshold. The scrap effect was so large that the author ran sensitivity checks: excluding the most extreme lines still yielded effect sizes above 5, and a decomposition of nearly 11,000 scrapped parts showed that transient-state events such as startups and changeovers accounted for 67.4 percent of scrap and fell by 56.6 percent during deployment, consistent with the standardized startup checklists and first-piece approval gates introduced under the lean foundation.

Principal component analysis reinforced the picture of a system-level shift rather than isolated gains. The first two components explained 78.7 percent of the variance, and pre- and post-implementation observations separated cleanly along the first axis, with post-implementation points clustering tightly. The correlation structure itself changed: the pre-implementation negative correlation between cycle time and downtime—evidence of frantic, compensatory schedule recovery—weakened from −0.51 to −0.09, suggesting that AI-supported dispatching absorbed disruptions at the scheduling level instead. Process variability fell by 15 to 26 percent across all four key performance indicators, and a target-based capability index for scrap improved from a negative value, meaning the process average sat above the internal target, to 1.24, indicating genuine capability.

Sustainability was tracked over 26 weeks after the deployment ended, and the gains held. Overall equipment effectiveness climbed steadily from 65.6 percent in week one to 72.0 percent by week 26, scrap converged to about 2.2 percent, and schedule adherence stayed within a narrow band. All four stability indices remained below the facility’s 0.006 target for mature, well-governed processes. Downtime coding accuracy, a proxy for data quality, improved from 88.2 percent in month one to 96.2 percent by month 12, creating what the author describes as a virtuous cycle in which lean discipline strengthens the data infrastructure, which in turn reinforces the reliability of AI recommendations. The line-level cost-benefit analysis yielded annualized savings of USD 563,250 under a full-impact scenario and USD 487,250 under a conservative estimate counting only directly metered categories, against an implementation cost of roughly USD 15,000.

The study is candid about its limits. It is a single-facility investigation, so the results cannot be extrapolated without multi-site replication, and the overlapping deployment phases prevent clean causal isolation of lean versus AI contributions. Facility-level trajectories show no distinct acceleration after AI activation, suggesting the lean foundation drove much of the early improvement. Still, the practical message for the thousands of mid-sized molding plants worldwide is clear: deploy lean as a system-level architecture first, constrain AI within governance gates as decision input rather than authority, and recognize that existing ERP data combined with lean-disciplined capture is enough—sensors and IoT are not prerequisites. Future work will pursue multi-site replication, closed-loop scheduling, and longer tracking windows to confirm that these durable gains generalize.

Subject of Research: Lean-AI integration for waste reduction and cost optimization in medium-scale injection molding

Article Title: A lean–AI integration for decision-centric waste reduction and operational cost optimization in medium-scale injection molding

Article References: Thummar, C. (2026). A lean–AI integration for decision-centric waste reduction and operational cost optimization in medium-scale injection molding. Heliyon, 12(15), Article e45517. https://doi.org/10.1016/j.heliyon.2026.e45517

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45517

Keywords: injection molding, lean manufacturing, artificial intelligence, waste reduction, governance gates, SMED, overall equipment effectiveness, scrap reduction, decision latency, ERP data, process capability, cost-benefit analysis