AI Quality Control in Manufacturing: Cut COPQ Without Adding QC Headcount

The Cost of Poor Quality is the most under-reported line on a manufacturer's P&L. Most companies know their scrap cost. Few have accurately calculated rework labour, warranty returns, customer concessions, and the overhead buried in their quality department. When you add them up, the industry average COPQ runs 4–8% of revenue.

For a manufacturer turning £30M annually, 6% COPQ is £1.8M per year leaving through the back door in defects, rework, and customer complaints. The CFO sees some of it. Most of it is invisible — absorbed into overhead rates, booked as "miscellaneous", or simply not tracked because tracking it requires more headcount.

AI quality control does not require more headcount. It requires better data, and a system that acts on that data faster than any human inspector can.

The COPQ Benchmark That Should Be Your Target

COPQ consists of four cost categories:

Internal failure costs — scrap, rework, re-inspection, downgrading. These are caught before the product reaches the customer.

External failure costs — warranty claims, returns, customer complaints, product liability. These are caught after delivery and carry a multiplier: external failure typically costs 5–10x the equivalent internal failure.

Appraisal costs — inspection labour, test equipment, quality audits. These are the costs of finding defects after they are created.

Prevention costs — process design, SPC systems, training, supplier qualification. These are the costs of not creating defects.

Most manufacturers over-invest in appraisal (more inspectors) and under-invest in prevention. AI quality control shifts the balance by making prevention cost-effective at scale: monitoring every unit, every process parameter, every shift — without adding inspection headcount.

Techseria client outcomes: COPQ from the 4–8% industry average to 1.5–2.5% of revenue. For a £30M manufacturer, that is £450k–£1.35M per year in measurable, bottom-line value.

What AI Quality Control Actually Monitors

The quality control agent operates at the process level, not the product level. Rather than inspecting finished goods (appraisal), it monitors the process parameters that determine whether the product will be in-spec before the defect is created.

This distinction is the technical core of the value proposition. A human inspector finds defects after they exist. The AI agent finds process deviations before defects are produced.

Process parameters monitored via IoT/OPC-UA:

• Temperature — curing temperatures, mould temperatures, heat treatment cycles
• Pressure — injection moulding pressures, hydraulic system pressures, vacuum levels
• Speed — spindle speeds, conveyor speeds, extrusion rates
• Cycle time — actual vs. standard, deviation flags slow cycles (tool wear) and fast cycles (setup error)
• Vibration — typically used for machine condition but correlated with surface finish in machining
• Current draw — motor current is a proxy for load, useful for detecting tool breakage or process overload

These parameters are captured via OPC-UA (the industrial protocol standard for CNC machines, PLCs, and SCADA systems) or via IoT sensors with Azure IoT Hub as the ingestion layer. The agent reads time-series data at 1–10 second intervals depending on the process.

Statistical Process Control at Machine Speed

The agent applies SPC (Statistical Process Control) logic to incoming sensor data. SPC defines control limits based on the natural variation of a stable process — typically the mean ± 3 standard deviations (3-sigma limits). Data points outside control limits, or patterns within control limits (runs, trends, systematic bias), signal that the process is no longer in control.

A human SPC analyst reviews control charts periodically — shift-end, or when a batch is finished. The AI agent applies the same logic in real time, at every data reading.

The eight Western Electric rules for non-random patterns — single point beyond 3-sigma, two of three beyond 2-sigma on the same side, four of five beyond 1-sigma, eight consecutive points on one side of the mean — are all evaluated by the agent for each monitored parameter at each reading interval.

When a control limit violation or non-random pattern is detected, the agent does not wait for a human to notice. It acts immediately.

The Agent's Response Architecture (LangGraph.js State Machine)

The quality control agent is implemented as a LangGraph.js state machine with the following nodes:

Sensor Data Node — ingests real-time readings from OPC-UA/IoT sources via Azure IoT Hub. Validates data completeness. Handles sensor dropouts without false alarms.

SPC Evaluation Node — applies control limit checks and Western Electric pattern rules to each parameter. Computes a process stability score (0–100) for each monitored characteristic.

Anomaly Classification Node — classifies detected deviations by severity (warning vs. alarm), likely cause (based on deviation type and process context), and affected product range (which units produced during the deviation window are at risk).

ERPNext Action Node — based on classification:

• Warning: creates a Quality Inspection record in ERPNext (`/api/resource/Quality Inspection`) flagging the affected batch for in-process check.
• Alarm: creates Quality Inspection plus a Non-Conformance Report, triggers a material hold flag in the Batch record, notifies the shift supervisor via ERPNext notification.
• Critical: all of the above plus a work order hold, halting further production until a quality engineer reviews.

Corrective Action Node — links the Non-Conformance Report to a Corrective and Preventive Action (CAPA) workflow in ERPNext. Suggests probable cause based on the deviation pattern. Tracks open CAPAs and escalates if overdue.

Vision Inspection Node (optional) — where camera inspection infrastructure exists, the agent integrates computer vision outputs (pass/fail classification from a separately trained vision model) as an additional input to the anomaly classification step.

The Three Data Sources That Must Be in Place

The agent requires three data sources to function effectively. Missing any one of them degrades accuracy significantly.

1. Process parameter history — at minimum 90 days of historical sensor data for the monitored parameters, at the same sampling rate as the live feed. This history is used to calculate control limits (process mean and standard deviation under stable conditions). If historical data is not available from existing sensors or SCADA, a 30-day baseline capture period is required before the agent can go live.

2. Quality inspection results — historical Quality Inspection records in ERPNext linking batch numbers to pass/fail outcomes. These are used to validate that the process parameters the agent monitors actually correlate with quality outcomes. If your ERPNext Quality Inspection records are sparse or inconsistent, Techseria's implementation includes a 4-week data collection protocol before agent training.

3. BOM and routing — accurate bill of materials and routing records (as for production planning) to map which work centres and process steps are associated with which quality characteristics. This enables the agent to know which parameters to monitor for which product-operation combinations.

Sensor Infrastructure: What You May Already Have

Many manufacturers have more sensor data than they use. CNC machines built after 2015 typically have OPC-UA capability built in. PLCs controlling conveyor systems, ovens, and hydraulic presses are often connected to SCADA systems that log data — but nobody analyses it in real time.

The implementation scope begins with a sensor audit to identify what is already connected and readable. In Techseria's experience:

• 60–70% of required parameters are already captured in existing SCADA/PLC systems
• 20–30% require IoT sensor addition (temperature, vibration, or pressure sensors on equipment not currently monitored)
• 5–10% require OPC-UA enablement on equipment that has the capability but it has not been configured

Sensor retrofit cost: £5k–£20k depending on the number of unmonitored parameters and whether Azure IoT Hub is already deployed. This is separate from the software cost.

ROI Example: £2.1M Scrap Cost to £820k

A precision engineering manufacturer in the Midlands, £18M revenue, presented with the following baseline:

• COPQ: 7.2% of revenue = £1.3M
• Of which scrap and rework: £2.1M gross (some recovered as swarf)
• Net scrap cost (after recovery): £820k was the target

After 14 months of AI quality agent operation:

• COPQ: 2.1% of revenue = £378k
• Scrap and rework reduction: 61%
• QC headcount: unchanged (2 quality engineers) but role shifted from inspection to CAPA management and supplier qualification
• Annual value delivered: £922k
• Implementation cost: £44k (including sensor addition for three previously unmonitored work centres)
• Payback: 17 weeks

Implementation Timeline: 10–14 Weeks

The timeline is longer than other AI manufacturing agents because sensor integration is a physical, on-site activity that cannot be parallelised with software development.

Weeks 1–2: Site survey and data audit. Sensor coverage assessment, OPC-UA connectivity test, ERPNext Quality Inspection data review, baseline COPQ calculation.

Weeks 3–4: Sensor installation and connectivity. IoT sensor addition for uncovered parameters, OPC-UA configuration, Azure IoT Hub ingestion pipeline established and tested.

Weeks 5–7: Baseline data capture and model calibration. Live data flowing to the agent. Control limits calculated from stable-process data. SPC models calibrated per work centre and product family.

Weeks 8–10: Agent development and ERPNext integration. LangGraph.js state machine built, Quality Inspection and NCR creation tested, CAPA workflow integrated, alert routing configured.

Weeks 11–12: Parallel run. Agent flags anomalies in shadow mode. Quality engineers review and validate: is the agent catching real issues? Are false alarm rates acceptable? Parameters tuned.

Weeks 13–14: Live deployment. Agent operates autonomously. Real Quality Inspections and holds created. Quality team transitions from inspection mode to exception management mode.

Investment: £30,000–£65,000 Fixed Fee

• £30k–£42k: Fewer than 8 monitored work centres, sensor infrastructure largely in place, ERPNext Quality Inspection records available.
• £42k–£65k: 8–20 work centres, significant sensor addition required, vision inspection integration, complex multi-site configuration.

Sensor hardware is additional (typically £5k–£20k, procured separately or through Techseria's hardware partner network).

Annual software maintenance: 15–18% of build cost, covering model recalibration as products and processes change.

Start with a COPQ baseline. Techseria will calculate your current Cost of Poor Quality from your ERPNext data before you commit to anything. If the number is less than £300k annually, AI quality control may not yet be the highest-ROI investment. If it is above £400k, the business case is almost certainly compelling. Book a scoping call and we will show you the number.

[Book a Strategy Session] | [Get a Fixed-Fee Quote]