1
Data Infrastructure & Signal Quality
Before AI can inspect a single part, your data pipeline has to earn the right to be trusted. Audit the foundation first.
For: QA Leads, Plant IT, Data Engineers
Inventory all inspection data sourcesList every CMM, vision system, gauge, and PLC feeding QA records — including manual entry points.
Confirm timestamp synchronization across machines (≤100ms drift)Misaligned clocks break root-cause traceability faster than bad sensors.
Verify image resolution meets defect size ÷ 4 ruleIf your smallest defect is 200µm, capture at ≤50µm per pixel.
Document lighting conditions for every vision stationLux, angle, color temperature — and tolerance windows for each.
Quantify labeled defect samples per class (target: ≥300 per defect type)Below this threshold, model performance becomes statistically fragile.
Establish ground-truth labeling protocol with inter-rater agreement ≥85%Two inspectors should agree on a defect call 85% of the time — or your labels are noise.
Audit data retention policy against MES and ERP timelinesInspection data should outlive the part's warranty window.
Map data lineage from sensor → storage → model → decisionIf you can't trace a flagged part backward through the stack, you can't defend the call.
2
Edge Compute & Real-Time Readiness
Cloud round-trips don't fit cycle times measured in seconds. Inspect the silicon, network, and latency budget on the shop floor.
For: Automation Engineers, OT Architects
Measure end-to-end inference latency budget (target: <200ms per part)Includes capture, transfer, inference, and PLC handoff.
Confirm edge GPU/NPU availability at each inspection cellCloud inference is a non-starter for tact times under 5 seconds.
Stress-test network during peak shift (packet loss <0.1%)WiFi that works at 6am can collapse at 2pm shift change.
Define graceful degradation behavior when AI is unavailableDefault to traditional gauging? Stop the line? Document the answer.
Implement edge model version control and OTA rollbackOne bad model push shouldn't require a service truck.
Benchmark cell uptime contribution from AI hardware (target ≥99.5%)If AI adds three nines of downtime, the ROI math collapses.
Validate thermal envelope for edge devices in production conditionsDatasheet specs assume 25°C. Your enclosure doesn't.
3
Compliance, Traceability & Audit Trail
Aerospace, medical, and automotive customers will ask hard questions. Have the documentation ready before they do.
For: Quality Managers, Compliance Officers
1Map AI decisions to your existing ISO 9001 / IATF 16949 control planEvery AI-driven accept/reject must trace to a documented control point. No exceptions.
2Document model training data provenance (parts, dates, operators, machines)Required for AS9100 and increasingly expected by Tier-1 automotive customers.
3Establish model change-control board with QA, engineering, and IT sign-offA model update is a process change. Treat it like one.
4Define explainability requirements per customer contractSome OEMs now require heatmaps or feature attribution for every reject.
5Verify EU AI Act risk classification for your specific use caseIndustrial QA generally lands in "limited risk," but verify — fines start at €15M.
6Build retention archive linking each part serial to its inference outputRecall scenarios depend on this. Build it now, not after the call.
7Conduct annual bias audit on model performance across operators and shiftsModels can quietly favor day-shift lighting or one operator's loading style.
4
Use-Case Prioritization & ROI Framing
Not every defect deserves an AI model. Score the candidates, fund the winners.
For: Operations Directors, Plant Managers
Scoring template — apply to each candidate use case:
Annual scrap cost from this defect ($): __________
Current inspection time per part (sec): __________
False reject rate of current method (%): __________
Image/data availability today (low / med / high): __________
Customer escape risk if missed (1–10): __________
Cycle time headroom available (ms): __________
Score >70 = greenlight. 40–70 = pilot. <40 = wait.
Rank top 10 defect modes by annual cost of poor qualityMost plants discover the top 3 account for 60%+ of scrap dollars.
Identify which defects current vision/gauging genuinely missesIf a $40k vision system catches it, AI may not be the right spend.
Calculate payback period assuming 18-month model lifecycleModels drift. Budget for retraining, not just deployment.
Confirm executive sponsor and shop-floor champion for pilotPilots without both die in month four.
5
Operator Trust & Change Management
The best model in the world fails if inspectors override every call. Design for adoption, not just accuracy.
For: HR, Training Leads, Production Supervisors
Run baseline survey on operator confidence in current QA methodsYou need the before-picture to defend the after-picture.
Design HMI to show AI confidence score, not just pass/failOperators trust systems they understand. Show your work.
Build operator override workflow with mandatory reason codesEvery override is free training data — capture it structurally.
Schedule monthly model performance reviews with the floor teamThe people closest to the parts will spot drift before the dashboards do.
Define escalation path for disputed calls (operator vs. AI)Who breaks the tie at 2am on a Saturday? Decide before it happens.
Allocate training budget at 15% of total project costUnderspending here is the most common reason pilots stall at scale.
Pro Tip
Score your readiness on a 0–2 scale per item (no / partial / yes). Below 50/84 means you're not ready for production AI — but you are ready for a structured pilot. Above 65 means you're leaving money on the table by waiting.
Ready to run the full audit on your line?
Download the printable 42-point checklist, complete with scoring rubric, customer-facing audit template, and a worked example from a precision aerospace supplier.
Get the Checklist