What problem does it solve?
Classic quality control samples. An inspector checks a share of the units, technicians test a random selection of welds with ultrasound, a specialist walks round the finished product at the end of the line. Defects that fall between samples travel on to the next station, to the customer or into a warranty claim, further away from the station that caused them.
Rule based machine vision suits simple, stable parts, but every new variant, lighting change or defect type means writing and tuning new rules. The shift is to learning models that check every unit, from images, sound or the process data a machine already produces, and send people only the cases that look wrong.
How does it work?
- Capture every unit. Cameras, microphones or the machine controller record each part or vehicle at the station, for example images of an assembly step, driving noise from the seats or the process readings a machine logs for each joint.
- Score it against what good looks like. A model trained on labelled examples, or on normal production when defects are rare, classifies the unit or scores how far it deviates from normal. Synthetic defect images can fill the gap when real defects are too rare to train on.
- Check it against the order. The expected variant, parts list and assembly steps come from the production system, so the model knows what this specific unit should look like.
- Alert the right person at the right station. An anomaly goes straight to the worker or inspector on a smart device, with the image or clip, while the unit can still be fixed in line.
- Close the loop. Confirmed and rejected findings are logged, used to retrain the model and analysed for root causes such as a drifting machine setting, so the process improves rather than just the inspection.
- Audience
- Employee facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cost reduction | Too few to pool | 7% | 1 | 1 vendor |
| Error reduction | Too few to pool | 67% | 1 | 1 vendor |
| Interactions handled | Not pooled | about 1.5 million | 1 | 1 organization |
Value drivers: Risk and loss reduction, Lower cost to serve, Employee productivity, Speed and cycle time.
Indicative value
An assembly plant building 200,000 units a year
USD 120,000 to USD 1.6 million
Cost of late defects avoided per year
How this is calculated
Formula: units * defectRate * costPerDefect * reduction. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Units produced per year units, units per year | 200,000 | 200,000 | The reference plant. |
| Share of units with a defect found late (final inspection, customer or warranty) defectRate, fraction of units | 0.02 | 0.04 | Editorial assumption, replace with your own late defect and warranty rate. |
| Cost of a defect found late costPerDefect, USD per defect | 150 | 400 | Editorial assumption covering rework, scrap and warranty handling. Replace with your own cost of poor quality. |
| Share of late defects avoided by inspecting every unit in line reduction, fraction of late defects | 0.2 | 0.5 | Conservative against the benchmark on this page (NVIDIA reports a 67% decrease in defect rates on Pegatron assembly lines using its visual AI agent), because that figure is one vendor reported case in electronics assembly. The benchmark is not like for like: it measures fewer defects produced, while this input measures fewer defects escaping to late stages. |
What it leaves out: Counts avoided rework, scrap and warranty cost only. It leaves out cameras, sensors, edge computing and integration, the labelling effort, any reduction in inspection staff hours and the value of fewer recalls and a better brand reputation.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Audi
Germany · Automotive · 2023
After a successful pilot at its Neckarsulm site, Audi began rolling out an AI system for quality control of resistance spot welds in car body construction (project WPS Analytics). The AI analyses around 1.5 million spot welds on 300 vehicles each shift, where production staff used to check around 5,000 spot welds per vehicle with ultrasound based on random analyses, so employees can now focus on possible anomalies. Audi developed the process with the German Association for Quality (DGQ) and two Fraunhofer institutes so that it would hold up in audits and certification. It began installing the infrastructure at Audi Brussels, with Ingolstadt and the Volkswagen plant in Emden scheduled to follow, and is retraining the model for differences in weld settings.
- Interactions handled: about 1.5 million, spot welds analysed per shift at Neckarsulm, on 300 vehicles
"Using artificial intelligence, Audi analyzes around 1.5 million spot welds on 300 vehicles each shift at its Neckarsulm site."
Claimed by: organization
BMW Group
Germany · Automotive · 2023
BMW built AIQX (Artificial Intelligence Quality Next), an in house platform that places cameras and sensors along the assembly line, analyses their data in real time and sends line workers immediate feedback on smart devices. It checks variants and completeness and flags anomalies, and at Plant Dingolfing a sub area called Acoustic Analytics listens to driving noises through microphones on the seats as the final check before handover. BMW describes these AI quality tools as part of the production master plan for its plants worldwide. Separately, in 2025 Plant Regensburg piloted GenAI4Q, developed with Datagon AI, which builds an individual final inspection catalogue for each of the roughly 1,400 vehicles built there each day.
No outcome disclosed.
Pegatron
Taiwan · Manufacturing · 2025
Pegatron, an electronics manufacturer with 24 sites, built an Assembly Guiding Agent on NVIDIA's video search and summarization blueprint that watches manual assembly in real time, spots missed steps such as a forgotten screw and alerts the worker, who can replay the clip and ask the agent questions. It also generates synthetic defect images from CAD drawings and historical data to train its visual inspection models, because real defect images are scarce on high yield lines. NVIDIA reports lower defect rates and labour cost per line for the assembly agent. The page is undated; the assembly agent dates from 2025, while the synthetic defect image work may be later.
- Error reduction: 67%, assembly lines using the Assembly Guiding Agent
"By augmenting the assembly process with this AI agent, Pegatron is seeing a 7% reduction in labor costs per assembly line and a 67% decrease in defect rates."
Claimed by: vendor - Cost reduction: 7%, labour cost per assembly line
"By augmenting the assembly process with this AI agent, Pegatron is seeing a 7% reduction in labor costs per assembly line and a 67% decrease in defect rates."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Images, sound or process signals captured consistently per unit and station
- Labelled examples of good units and of each defect class, or long runs of normal production
- The expected variant, parts list and work steps per unit from the production system
- Confirmed outcomes of flagged cases to retrain and to measure false alarms
Systems to integrate
- Manufacturing execution system for orders, variants and unit identity
- Cameras, microphones or machine controllers at the stations
- Edge computing near the line for low latency scoring
- Quality management system for defect records and corrective actions
- Worker devices or line displays for alerts
Complexity: High
The model is rarely the hard part. The work is in stable image or signal capture on a moving line, enough labelled defects, a link to the order so the model knows the expected variant, and a validation approach that quality auditors and certification bodies accept.
- 1
Pick one station with a costly, visible defect
Choose a defect that escapes today, is expensive later and can be seen, heard or measured at one station, for example a missing fastener or a weak weld. Record its current escape rate.
- 2
Fix the capture before the model
Stabilise lighting, camera angle, microphone placement or signal logging first. A better model cannot make up for inconsistent capture.
- 3
Build the defect library
Collect and label real defects with the quality team. When real defects are rare, train on normal production for anomaly detection or add synthetic defect images, and keep a real holdout set to test on.
- 4
Run in shadow mode next to the current inspection
Let the model score every unit while the existing sampling continues, and compare both on catches, misses and false alarms before anyone relies on it.
- 5
Agree the validation with quality and auditors
Document how a result is produced, the acceptance thresholds and the retraining rules, so the process survives audits and certification. Audi worked with DGQ and Fraunhofer for this.
- 6
Go live with a human on every alert, then widen
Route anomalies to the station, measure confirmation rates per defect class, and only then reduce manual sampling or add stations, plants and suppliers.
Guardrails
- A documented acceptance threshold per defect class, with a human decision on every flagged unit
- A fallback to the previous inspection method when the model or capture is unavailable
- Retraining only through change control, tested on a fixed holdout set before release
- Drift monitoring on input images or signals after changes to materials, lighting or machines
- Recording limited to the product and process, with worker privacy rules for any video of people
KPIs to instrument
- Escape rate of each defect class to later stations, the customer and warranty, before and after
- False alarm rate per defect class and station
- Share of flagged units confirmed as real defects
- Inspection coverage (share of units inspected) and time to detection
- Cost of poor quality per unit produced
Human in the loop
Quality inspectors confirm or reject every flagged unit and own the decision to release, rework or scrap. Quality engineers approve each new defect class, threshold and model version, and a sample of passed units is still inspected manually to measure what the model misses.
Common failure modes
- False alarms that train people to ignore alerts
- A model that flags too much gets overridden by habit. Tune thresholds per defect class and report confirmation rates to the line.
- Silent drift after a process change
- A new supplier, paint batch or light fitting changes the input and accuracy drops without anyone noticing. Monitor input drift and recheck on a holdout set after every change.
- A model trained on too few real defects
- Rare defects are exactly the ones that matter. Use anomaly detection or synthetic data, but always test on real defects.
- Inspection without root cause
- The model catches more defects but nobody fixes the cause. Feed findings into corrective actions and machine settings.
What are the risks and rules?
EU AI Act
Depends on design
Inspecting products is not an Annex III use, so a system that only judges parts, welds or assemblies is usually minimal risk. Two designs change that. Under Article 6(1) it is high risk when both conditions hold: it is a safety component of a product (or itself a product) covered by the Union harmonisation legislation in Annex I, and that law requires a third party conformity assessment of the product. For a production line the relevant product laws are the Machinery Regulation (EU) 2023/1230 and, for cars, the vehicle type approval regulations. Since the Digital Omnibus on AI, Regulation (EU) 2026/1744, moved the Machinery Regulation into Annex I Section B, where the vehicle type approval regulations already sat. Article 6(1) still classifies such a safety component as high risk, but under Article 2(2) only Article 6(1), Article 60a and Articles 102 to 112 of the AI Act apply directly. The requirements reach the system through the sectoral law instead: delegated acts amending Annex III of the Machinery Regulation, and type approval for vehicles. The AI Act rules for Article 6(1) high risk systems apply from 2 August 2028. An inspection system on the assembly line is usually not a safety component of the product it inspects. If it monitors and evaluates the performance and behaviour of individual workers, for example by scoring who made an assembly error, it falls under Annex III point 4(b) and is high risk. Keep the output about the unit, not the person.
Rules that apply
Guidance
- Article 2, scope (European Union, Europe). Article 2(2), as amended by Regulation (EU) 2026/1744, says that for high risk AI systems related to products under the Annex I Section B laws, such as machinery and vehicle type approval, only Article 6(1), Article 60a and Articles 102 to 112 apply.
- Annex I, list of Union harmonisation legislation (European Union, Europe). The product laws behind Article 6(1). Since Regulation (EU) 2026/1744, Section B lists the Machinery Regulation (EU) 2023/1230 as point 21, next to motor vehicle type approval (Regulations 2018/858 and 2019/2144), and the Machinery Directive 2006/42/EC is deleted from Section A.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 4(b) lists AI used to monitor and evaluate the performance and behaviour of workers, which matters when inspection video also shows people.
- Regulation (EU) 2023/1230 on machinery (European Union, Europe). The EU Machinery Regulation replaces Directive 2006/42/EC and applies from 14 January 2027. Since Regulation (EU) 2026/1744 it is listed in Annex I Section B of the AI Act. It is the product law that can make an AI safety component of a machine high risk under Article 6(1), and the AI requirements for such machines are to be added to its Annex III by delegated acts that apply by 2 August 2028.
- IATF 16949 automotive quality management (International Automotive Task Force, Global). The body behind the IATF 16949 automotive quality management standard. At certified plants, plan for an AI based inspection process to be reviewed as part of the quality management system in certification audits.
- Digital Omnibus on AI, Regulation (EU) 2026/1744 (European Union, Europe). Entered into force on 27 July 2026. It moves machinery from Annex I Section A to Section B, limits the AI Act's direct application to machinery to the provisions in Article 2(2), and sets 2 August 2028 as the application date for Article 6(1) high risk AI systems.
Controls to put in place
- Validation record per model version with holdout results per defect class
- Change control for thresholds, training data and model releases
- Traceability from every flagged or passed unit to the model version and input data
- Data protection impact assessment where cameras capture workers
- Periodic manual audit of passed units to estimate the miss rate
Frequently asked questions
- How much can AI inspection reduce defects?
- Published results vary widely by defect and line. NVIDIA reports that Pegatron saw a 67% decrease in defect rates on assembly lines using its visual AI agent, a vendor figure for electronics assembly. Plan for a smaller effect at first and measure escapes per defect class against your current sampling.
- Does AI inspection need cameras?
- Not always. At Plant Dingolfing, BMW's Acoustic Analytics listens to driving noises through microphones on the seats as a final check before handover. Choose the signal that shows the defect most reliably at the lowest cost.
- Is AI quality inspection high risk under the EU AI Act?
- Usually not when it only judges the product. It is classified as high risk if it is a safety component of a product under Annex I legislation, such as machinery or vehicles, and that law requires a third party conformity assessment. Since the Digital Omnibus on AI, machinery is in Annex I Section B, like vehicle type approval already was, so the requirements for those systems come through the Machinery Regulation (delegated acts that apply by 2 August 2028) and vehicle type approval rather than the AI Act directly. It is also high risk if it scores the performance of individual workers, which Annex III point 4(b) covers.
How to cite this page
Blits.ai AI Use Case Library, "AI quality inspection on the production line", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/production-line-quality-inspection. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published