Multi-Agent Orchestration in Munich, Germany

Automotive Manufacturing / Tier-1 Supplier

A Munich-based Tier-1 automotive supplier (precision machined components for BMW, Audi, and Porsche, 850 employees, EUR 180M revenue, IATF 16949 certified) managed quality across 3 production facilities with 45 CNC machining centres and 12 measurement stations. The problems: (1) Quality data was fragmented across systems — CMM measurements in one system (Zeiss CALYPSO), vision inspection in another (Keyence), SPC data in spreadsheets, and non-conformance reports in a legacy QMS. No single view of quality status existed. (2) SPC monitoring was reactive. Process capability indices (Cpk) were calculated daily in spreadsheets — by the time an out-of-control condition was identified, hundreds of potentially non-conforming parts had been produced. Average time from process drift to detection: 4.2 hours. (3) Customer complaints were increasing. BMW's supplier scorecard showed the company's PPM (parts per million defective) rate trending from 12 PPM to 23 PPM over 18 months — approaching BMW's 25 PPM threshold for supplier development intervention. (4) 8D problem-solving was slow. When quality issues were identified, 8D investigation reports (the standard automotive problem-solving methodology) took an average of 14 business days to complete — BMW required completion within 10 days. Late 8D submissions resulted in scorecard demerits.

We built a 5-agent quality management orchestration system over 9 weeks. (1) Data Integration Agent: ingesting and normalising quality data from all sources in real-time — CMM measurements (Zeiss CALYPSO API), vision inspection (Keyence data export), in-process gauging (machine tool probing data via MTConnect), and environmental monitoring (temperature and humidity sensors affecting dimensional stability). Creating a unified quality data lake with standardised measurement records linked to part numbers, work orders, machine IDs, and operator IDs. (2) SPC Monitoring Agent: performing continuous statistical process control on all critical characteristics (Cpk/Ppk calculation, control chart analysis — X-bar/R, X-bar/S, and p-charts as appropriate). The agent detecting: out-of-control conditions (Western Electric rules — points beyond control limits, runs, trends, stratification), process capability degradation (Cpk dropping toward the 1.33 minimum required by automotive customers), and tool wear patterns (systematic dimensional drift indicating tool replacement needed). Alert generation: immediate notification to the machine operator and shift supervisor when SPC rules are violated — within seconds of measurement, not hours. (3) Non-Conformance Agent: managing the disposition of non-conforming material. The agent: classifying the non-conformance (dimensional, surface finish, material, contamination), determining severity (safety-critical characteristic, fit characteristic, cosmetic), recommending disposition (scrap, rework, concession — with data-driven justification based on the specific deviation and historical acceptance patterns), generating containment actions (sorting suspect inventory, notifying affected customers if shipped material may be affected), and creating the non-conformance record per IATF 16949 requirements with full traceability. (4) 8D Investigation Agent: orchestrating the 8D problem-solving process when a customer complaint or significant internal non-conformance is reported. The agent: D1 — assembling the problem-solving team (based on the type of issue and affected process area), D2 — generating the problem description from quality data (what, where, when, how big, trend), D3 — recommending interim containment actions (based on historical similar issues), D4 — performing root cause analysis (Ishikawa/fishbone generation from process data — identifying the most likely root causes based on the timing, magnitude, and pattern of the defect), D5 — proposing corrective actions (based on root cause and historical effectiveness data), D6 — generating the verification plan (how to confirm corrective action effectiveness), D7 — recommending preventive actions (PFMEA updates, control plan changes), D8 — drafting the closure report. Average 8D completion time: reduced from 14 business days to 5 business days (AI orchestration + engineer review and approval). (5) Supplier Quality Agent: monitoring incoming material quality from 120+ suppliers. The agent: tracking incoming inspection results by supplier, part number, and characteristic, calculating supplier PPM and lot acceptance rates, triggering supplier corrective action requests (SCAR) when performance thresholds are breached, and generating supplier scorecards for quarterly business reviews. The orchestrator managing the pipeline with human gates at: SPC alert response (operator confirming the issue and taking action), non-conformance disposition (quality engineer approving the AI-recommended disposition), 8D approval (quality manager reviewing and approving the complete 8D report before customer submission), and supplier corrective action (SQE approving SCAR content before submission).

SPC detection time: from 4.2 hours average to real-time (within seconds of measurement). Process drift caught before producing non-conforming parts — preventing an estimated 15,000 suspect parts per month from entering the inspection queue. PPM rate: reduced from 23 PPM to 8 PPM within 6 months — well below BMW's 25 PPM threshold. BMW supplier scorecard improving from "Development Required" to "Preferred Supplier" rating. 8D completion time: from 14 business days to 5 business days — comfortably within BMW's 10-day requirement. 8D quality: zero BMW rejections of 8D reports in 6 months (vs. 4 rejections in the previous 6 months). Scrap reduction: 18% reduction in scrap rate through earlier detection of process drift and more accurate disposition recommendations. EUR 340K annual scrap cost savings. Supplier quality: incoming inspection reject rate reduced by 22% through proactive SCAR management — suppliers responding faster to data-driven corrective action requests. Annual value: EUR 340K (scrap reduction) + EUR 280K (productivity — quality engineer time savings) + EUR 200K (customer satisfaction — avoided scorecard penalties and potential business loss) + EUR 150K (supplier quality improvement) = EUR 970K annual value. Development cost: EUR 175,000. Payback: under 11 weeks.