Automotive factories already use many forms of advanced technology. Industrial robots move and weld parts, optical systems inspect components, production software coordinates materials, and engineers and operators use measurements to keep closely connected processes within their limits. Statistical methods, mathematical optimisation and specialist machine learning have been part of this work for years. Newer AI capabilities are therefore entering factories that have considerable experience in automation and its practical limitations.
The earlier article in this series examined how AI is changing the alternatives automotive engineers can investigate, the experiments they choose and the methods they use to assess designs. Manufacturing presents a related but different problem. A prediction or recommendation may lead to an adjustment on a running line, an inspection, a decision about scarce parts or a movement by a robot. Whether it helps depends on what the factory can do with the result and what existing equipment and methods could have accomplished instead.
AI itself covers different techniques in these settings. A trained model may predict a process condition, a statistical method may identify an abnormal measurement, an optimiser may calculate a feasible schedule without learning from examples, and a newer robot policy may learn how to manipulate an object. The examples that follow were selected from published studies, industrial reports and manufacturer disclosures to compare different production tasks, technical approaches and levels of practical use. They are not a representative survey of AI adoption across the automotive industry. Their significance depends on which production problem a method changes, what it adds to established practice and how that contribution holds up under operating conditions.
Monitoring equipment and improving production processes
At Magna's vehicle plant in Graz, several models pass through body-in-white production using transportable positioning devices that keep underbodies aligned during fabrication. A device can wear or shift between scheduled inspections, while removing equipment merely to check it creates maintenance work and potential disruption. The problem is not a lack of automation, but how to decide which devices need attention and when.
A 2026 study involving Magna engineers describes a system combining existing optical measurements, geometric relationships and statistical change-point detection to identify positioning devices that may require maintenance. In the research text, section 3.1 and Table 11, the authors report approximately 50,000 underbody measurements between January 2023 and March 2024. Relative to a fixed 180-day inspection strategy, they calculate 25% fewer maintenance removals for one vehicle platform and 51% fewer for another.
The figures describe positioning-device removals, rather than fewer defective cars or an independently audited financial return. The operating system relied on geometry-informed statistical change-point detection within the plant's production-monitoring process. The study is also notable for what was not implemented. Two other proposed factory applications were halted because the available data or validation arrangements did not meet their requirements, although work to address those gaps continued. For the operating positioning-device system, the investigators judged that a faster, more fully integrated response would cost more than its expected additional reduction in risk justified.
An automotive injection-moulding study published in 2026 offers a different example of established methods improving production. The researchers examined a system that connected machine information with structured reporting of problems, operator-led diagnosis and corrective-action tracking across approximately 75 machines. They reported overall equipment effectiveness rising from about 73.1% to 80.3%, while scrap fell from 2.57% to 1.41%. The intervention did not include a deployed predictive AI model. Because the before-and-after periods differed and no comparison line was held aside, the study does not isolate the effect of the new workflow from other changes at the plant.
Both accounts concern factories using data and operational expertise to address specific difficulties. They also give newer AI applications a meaningful point of comparison: improvements in production may follow from better measurement, engineering or work practices without requiring a new class of model.
Specialist machine learning within established controls
Specialist learning has nevertheless been useful in automotive manufacturing for some time. Bosch reported in March 2021 that its anomaly-analysis system was helping powertrain plants identify disruptions from machine and production data. The company estimated savings of €1 million to €2 million a year at pilot plants and described a 15% reduction in cycle time at Hildesheim. It planned to extend the system to around 50 powertrain plants and more than 800 lines. In a subsequent company report, Bosch said the anomaly-analysis software was in use at around 50 plants, with more than 2,000 production lines connected. This establishes a wider operational footprint than the early pilots, although savings were not reported on a comparable basis for every site.
The analysis drew on information collected through Bosch's manufacturing systems and supported investigations by plant staff. This is a useful distinction from the expectation that an AI model must take over an entire process to contribute to its performance. In industrial operations, detecting a problem sooner or narrowing the possible causes can matter because established equipment and people can act on the result.
A September 2026 account by Apollo Tyres and AWS describes a more recent application to tyre extrusion. Changing between recipes can require adjustments before material profiles stabilise, producing startup rework. Apollo uses a machine-learning model to recommend the initial line speed from historical data. A closed-loop control system then adjusts line speed using live measurements, with AWS also describing a Bedrock-assisted agent operating within defined controls and operator oversight. The company reports 26% less startup rework, 40% faster stabilisation and approximately 20 additional production hours each month.
Bosch and Apollo illustrate a practical pattern in which trained models contribute to operating production systems rather than replacing them. Both companies report meaningful improvements in defined activities. Apollo's figures describe the combined prediction, agent-assisted control and production workflow, so the additional effect of any individual component remains difficult to isolate.
Quality prediction and the work of inspection
Inspection presents a further challenge because detecting something unusual is only part of deciding what the factory should do next. Fraunhofer IGD and ZF describe statistical image-based anomaly detection used on multiple lines inspecting airbag gas-generator components. Fraunhofer reports a false-positive rate below 0.1%, which is relevant to unnecessary rejection and reinspection. The public account does not provide a corresponding measure of missed defects, an important consideration for components with demanding quality requirements.
A 2026 retrospective study of 16,500 automotive workpieces illustrates what can happen when prediction is brought forward into a production process. The researchers developed models with an area under the ROC curve of up to 0.81, indicating an ability to distinguish workpieces more likely to be defective. Applying the threshold used at final inspection earlier in production nevertheless produced no warnings and missed all 15 defects in the later evaluation set.
Recalibrating the threshold could identify defects earlier, but one transferred threshold would have flagged 89% of the later workpieces, creating a substantial inspection burden. A different policy designed around recent predictions and available inspection capacity identified 14 of the 15 defects with a reported 36% false-alarm rate. Its advantage over simpler rules was not consistent across additional tests, and the policies were evaluated retrospectively rather than installed on a running line.
The two studies concern different applications and levels of maturity, but together they clarify an important condition for useful industrial prediction. Warning quality depends not only on a model's ability to identify a pattern, but also on how warnings are interpreted, how many can be investigated and what happens when a defective component is missed. These are production questions that cannot be settled by an accuracy score alone.
Production decisions under material constraints
Some manufacturing decisions are removed from the individual machine but can still affect what the factory builds. A shortage of one component can change the allocation of vehicle options, production volumes and commitments to different markets. Manufacturers have used optimisation, simulation and planning systems to examine such constraints long before the arrival of language-model agents.
In 2023, BMW, BCG and AWS described an operating system that combined supply forecasts, bills of materials, vehicle configurations and financial priorities to generate alternative allocations of scarce parts. Mathematical optimisation produced scenarios in seconds to two minutes, according to the project account, while cross-functional BMW teams retained responsibility for the decisions. The account describes an operating decision-support system built on mathematical optimisation and human review, although it does not quantify independently verified net savings.
Factory planning also takes place before equipment or vehicle configurations change on the line. In a June 2025 update, BMW described its Virtual Factory as using digital twins of more than 30 production sites to test layouts, vehicle clearances and production movements before physical trials. BMW reported that automated collision checks reduced one planning task from almost four weeks of real-world testing to around three days, while a possible reduction of up to 30% in wider planning costs was still a projection. Generative and agentic features were being added to this established simulation environment, rather than being identified as the source of those existing results.
Newer AI may improve particular parts of such workflows. Bosch Mobility reports approximately €800,000 in savings from AI-supported management of maintenance and operating spares, with a wider deployment programme involving more than 60 plants. An estimated €16 million in possible future savings is a projection, not an amount already achieved. Bosch's case concerns indirect production-support materials rather than the vehicle components allocated by BMW.
Research into language-model agents is also beginning to test how natural-language requests can be connected to existing solvers. A 2026 study using automotive-supplier production data evaluated an agent system that orchestrated mathematical optimisation and simulation across seven planning experiments. The work tested the approach in simulated scenarios and reported improvements against manual planning activity, without establishing a continuing plant-wide financial or operational result. The question raised by such research is where language-based coordination adds value to established planning tools, rather than whether those tools have suddenly become obsolete.
Physical AI in factory trials
Robotic learning is developing along a different path. Manufacturers can already automate many repetitive handling operations, but tasks involving flexible materials, changing part positions or frequent reconfiguration may require extensive programming and tooling. Newer learned policies seek to make some of this work more adaptable, although their progress must be understood at the level of each task and trial.
Figure reported that its Figure 02 humanoids loaded three sheet-metal components into fixtures at BMW's Spartanburg plant in 2025, after which conventional industrial robots performed the welding. The company reported more than 90,000 parts handled and over 1,250 operating hours during the programme, which supported production involving more than 30,000 BMW X3 vehicles. This was participation in an operating line over an extended trial, although the disclosure does not make clear how hours were counted across the robots or provide a comparable cost assessment against a conventional loading cell. The Figure 02 generation has since been retired, and a June 2026 Figure 03 demonstration explored cart handling and logistics sequencing as different tasks.
Google DeepMind described another experiment in November 2026, using Apptronik robots at Mercedes-Benz production sites to pick brackets from bulk bins and place them into vehicle-specific kits. A model coordinating the task worked with a learned robot-action model, and DeepMind reported 96% task success. This was an in-factory proof of concept rather than a measured production-wide outcome. The reporting did not include a comparable shift throughput, a tally of human interventions or the economics of using the robots in regular operation.
Xiaomi's experiments offer a useful contrast with the automation already present in its vehicle plant, where the company describes more than 700 conventional industrial robots. Xiaomi reported different results for humanoid nut-loading trials during 2026, including a later company-reported 98% success rate. In November, it showed two humanoids sorting centre-console side covers during an eight-hour demonstration. The duration is noteworthy, but the video did not establish the number of parts handled, comparable human speed or the level of intervention required. Figures from the earlier nut-loading trials cannot establish performance in this different sorting task.
BMW, Mercedes-Benz and Xiaomi are therefore exploring different applications of learned manipulation, at different stages of development. The experiments show that researchers and manufacturers can now attempt tasks that merit attention beyond the laboratory. Whether the new methods make those tasks more economical than existing automation depends on repeatability, recovery from errors, reconfiguration effort, safety arrangements and the full costs of integration and maintenance. The public results considered here do not provide a common basis for making that comparison.
When engineering changes the automation problem
More adaptable robots are not the only way to address difficult handling tasks. Changes to the part, its packaging, the fixture or the production sequence can make automation easier without requiring a machine to cope with every variation.
Flexible wiring provides a useful example. Fraunhofer IWU's AuCA project feeds cable from spools and fixes it to a surface, reducing the need to grasp and place loose wires one by one. The work has been presented for samples and small-series applications, rather than as a proven high-volume vehicle installation. In Audi's Next2OEM demonstrator, partners considered wiring-harness manufacture, automation-compatible connectors and installation together. Both projects remain at limited development or demonstrator stages, but they show that part presentation and process design can change the automation task itself, alongside advances in robot dexterity.
This is also where the manufacturing story meets the earlier R&D article. The possibilities available to a vehicle designer can change as production methods develop, while design choices about components and their presentation can determine which forms of automation are viable. The two activities are linked, even when the results of individual research projects remain short of routine factory deployment.
Different priorities, and the evidence behind them
The strongest operating examples show that specialist AI already has a useful place in automotive manufacturing. Bosch's wider deployment and Apollo's process control involve trained models working alongside existing equipment and production expertise. Magna and the injection-moulding study show the continuing value of measurement, statistics and disciplined intervention. Several companies report meaningful operating improvements, but the extra benefit attributable to a particular learned model is often less clearly established than the improvement of the wider production system.
The inspection study makes clear why a capable predictor is only one part of a production decision. A warning becomes useful when it can be investigated within available capacity and acted upon at the appropriate stage. BMW's established shortage optimiser and Virtual Factory illustrate other ways that data, mathematical methods and simulation already support industrial decisions. For newer AI, the consequential advance is what it improves within those working processes, beyond producing an additional prediction or scenario.
Learned robotics extends the possibilities for physical work, but the comparison is similarly demanding. The factory trials by BMW, Mercedes-Benz and Xiaomi concern real or potentially useful tasks, though the available figures do not yet show whether learned manipulation offers better long-term operating economics than dedicated robots or redesigned tooling. This leaves room for genuine technical progress without treating every promising demonstration as an established industrial advantage. In manufacturing, the design of the part and the design of the process may be as consequential as the model controlling a machine.
Magna's study brings the distinction into focus. One system produced a reported operational improvement; two other ideas were held back by gaps in data and validation; and further investment in faster intervention was judged unnecessary for the benefit expected. Those were different responses to different evidence, not a single verdict on technology. Manufacturers start with different products, installed systems and priorities, and findings from other plants or research institutes may change the alternatives available to them. The value of following such work lies in understanding where a capability has become practically useful, where an experiment still has questions to answer and where an established method remains appropriate. That is a more defensible account of progress than assuming that every improvement in AI requires a corresponding change in production.