The manufacturing AI conversation in 2025-2026 has been dominated by proof-of-concept announcements and vendor capability claims. What is less common is honest analysis of what AI monitoring systems actually produce when deployed at scale in production environments run by large manufacturers with high operational standards and low tolerance for systems that generate noise rather than insight.
Jidoka Tech’s Nagare deployments at Procter & Gamble and Maruti Suzuki provide reference data on AI production monitoring P&G Maruti in two distinct manufacturing contexts: high-speed FMCG packaging and automotive component assembly. The operational and organisational lessons from both are relevant to any manufacturer evaluating AI monitoring for their own environment.
What automotive and FMCG manufacturing have in common
Despite the surface differences between a 500-units-per-minute FMCG packaging line and an automotive sub-assembly cell with 4-minute cycle times, both manufacturing contexts share the operational characteristics that make AI monitoring most valuable:
High process complexity. Both environments involve multiple process steps that must be executed in correct sequence under time pressure. Deviation from sequence generates quality or efficiency losses that often do not appear immediately at the point of deviation.
Mixed skill workforce. Both environments rely on a mix of experienced operators and new hires, with skill level variation that affects process compliance and quality output. Consistent monitoring provides a level of oversight that supervisor observation cannot achieve across full shifts.
Multiple product variants. Both environments run multiple product variants with different process requirements. Correct monitoring requires the system to recognise which product is running and apply the correct process standards.
Shift-based production with handover risk. Both environments run three shifts with handover periods that are high-risk for process continuity. Monitoring data that transfers with the shift provides incoming supervisors with a structured view of current process state rather than a verbal summary.
Automotive deployment: what Nagare added to a mature quality system
The Maruti Suzuki deployment was not a greenfield monitoring implementation. The plant had existing in-process inspection, CMM measurement, and a quality management system that generated regular performance reports. The question was not whether to implement quality monitoring, but whether AI process monitoring could improve beyond what the existing system was detecting.
The answer was yes, in a specific way: the existing system detected outcomes (rejections at inspection gates). Nagare detected processes (the assembly sequence deviations that preceded rejections). This distinction changed the improvement response from reactive quality correction to proactive process stabilisation.
The operational lesson: AI process monitoring and traditional quality inspection systems are not substitutes. They observe different things. Combined, they provide outcome data (did it pass?) and process data (was it built correctly?), which together enable root cause analysis that neither provides alone.
FMCG deployment: what Nagare added to a data-rich environment
P&G’s FMCG facility was not lacking in data. Historian data, SCADA records, and production logs provided substantial historical information. The challenge was not data availability but data utilisation: the data existed but was not being converted into real-time operational decisions.
Nagare’s contribution was the translation layer between continuous machine observation and actionable alerts for production supervisors. The historian recorded that machine 7 stopped 23 times during a shift; Nagare told the supervisor when machine 7 stopped during the shift, with enough context to respond.
The operational lesson: data richness and operational intelligence are different things. A monitoring system that generates a report after the shift creates awareness; a monitoring system that generates an alert during the shift creates an opportunity to respond.
What scale deployments reveal that pilots do not
Small pilot deployments of AI monitoring systems consistently look better than scaled deployments because pilots operate in conditions that are not representative of full production: better camera positions, more attentive implementation teams, and a smaller surface area of edge cases. The Nagare deployments at P&G and Maruti Suzuki were full-scale productions, not controlled pilots.
At full scale, three issues emerge that pilots obscure: alert fatigue from misconfigured thresholds, model accuracy degradation in shift conditions with different lighting and personnel, and integration friction with existing systems that were designed for manual data entry.
All three issues are addressable, but they require post-deployment tuning time that pilots typically do not build in. Manufacturers planning AI monitoring deployments should budget 4-8 weeks of post-go-live tuning in addition to the deployment period.