AMNA 2026
From connected data to virtual factories: How AI delivers value in automotive manufacturing
At Automotive Manufacturing North America 2026, experts from Autoliv, BorgWarner and Kistler discussed where AI and digital twins already improve manufacturing performance, why scaling remains difficult and how to balance automation with human expertise.
From connected data to virtual factories: How AI delivers value in automotive manufacturing
SEO teaser: At Automotive Manufacturing North America 2026, experts from Autoliv, BorgWarner, Kistler and the automotive manufacturing sector discussed where AI and digital twins already improve production, why scaling remains difficult and how to balance automation with human expertise.
Automotive manufacturers are collecting more production data than ever before, but turning that information into measurable operational improvements remains a challenge. AI agents and digital twins offer new opportunities to identify bottlenecks, improve quality and automate processes. Yet the technology alone is rarely the deciding factor. Data consistency, reliable models and a clear business case are equally important.
At Automotive Manufacturing North America (AMNA) 2026, experts from Autoliv, BorgWarner and Kistler, alongside manufacturing digitalisation specialist Rajeev Kalamdani, discussed how manufacturers can move beyond individual AI pilots and establish digital solutions that deliver value across production operations. Their experiences highlighted a common challenge: connecting systems is relatively straightforward compared with ensuring that the information they provide can actually support better decisions.
Where AI is already delivering results
At Autoliv, agentic AI is helping address a challenge in inventory management. Nitin Sethi, Vice-President and Head of IT at Autoliv Americas, explained how planners traditionally spent considerable time investigating potential material shortages. To avoid disruptions, they often increased minimum inventory levels, resulting in additional costs.
Autoliv has started using AI agents to perform some of the checks previously carried out manually, including analysing purchase orders and advance shipping notices. The objective is to identify potential shortages earlier and reduce the need for excessive safety stocks.
At BorgWarner, Maria Beemer, Global Chief Engineer, Transformation and Digitalization, described how AI is being used to improve compliance auditing in engineering processes. Rather than discovering problems during final audits, AI-supported checks help identify potential issues earlier in development. This approach has increased auditing capacity and improved downstream process quality, according to Beemer.
Rajeev Kalamdani provided a more production-focused example, explaining how AI can help manufacturers identify bottlenecks by analysing equipment-state data. The system continuously monitors whether individual machines are running, stopped, blocked or starved of material. AI agents then examine historical data across weeks or months to identify persistent production constraints.
By analysing where equipment states transition between blocked and starved, the system can locate potential bottlenecks and determine whether excessive cycle times, downtime or reliability problems are contributing factors.
This allows production teams to identify the stations causing the greatest disruption without manually examining large volumes of data. For Kalamdani, combining AI with digital twins is particularly valuable because it provides actionable insights directly to the people responsible for running production.
Data quality matters more than data quantity
For Sethi, one of the biggest misconceptions is that connecting operational technology (OT) and IT systems automatically creates business value. Instead, manufacturers should first define the operational problem and the performance indicators they want to improve.
This becomes particularly important when data from manufacturing execution systems (MES), product lifecycle management (PLM) and other applications needs to be combined. Differences in terminology, data ownership and master data structures can undermine otherwise promising applications.
Beemer described similar challenges at BorgWarner. Even established metrics such as the cost of poor quality are not always calculated consistently across factories. Some plants include expedited shipping costs, while others do not. Without common definitions, comparing performance across sites becomes difficult.
Hiroyuki Natsume, Sales Director Industrial Production at Kistler Group, argued that data collection should always reflect the intended application. Certain process parameters require high-frequency monitoring, while condition monitoring may need significantly less information.
Kalamdani added another dimension to the question of data reliability. Before manufacturers use digital twins to support operational decisions, they must ensure that the underlying measurements accurately represent physical processes. He pointed to measurement system analysis as an important discipline for identifying uncertainty and avoiding misleading conclusions from apparently precise data.
Digital twins must evolve with production
Keeping digital twins aligned with real production conditions presents another challenge. While sensor data and equipment signals can continuously update certain aspects of a digital model, production lines undergo changes that are more difficult to capture.
Kalamdani acknowledged that maintaining this alignment remains a challenge in practice. Changes to PLC and CNC programmes, for example, require disciplined version control and change management to ensure that digital models reflect the configurations actually running in production.
Although live equipment data can remain continuously updated, other elements of the digital representation may gradually become outdated. Maintaining accurate master copies and tracking modifications are therefore essential, particularly across production systems containing multiple identical machines.
Natsume noted that feedback loops are often relatively well established during prototyping but become weaker once series production begins. Engineering modifications, supplier changes and process adjustments can create discrepancies between the virtual representation and the physical factory.
He suggested systematically reviewing deviations six and twelve months after production launch to improve existing models and transfer the findings into future projects.
For older production equipment, Kalamdani identified retrofitted sensors and communication gateways as practical ways to introduce connectivity without replacing entire machines. However, the suitability of such solutions depends on the equipment and the data required.
How much autonomy should manufacturers give AI?
The discussion also explored when AI should move from providing recommendations to making decisions independently.
At BorgWarner, AI applications follow a phased implementation approach. Initially, the technology reviews human-generated work and identifies potential errors. Once its reliability has been established, it moves into a coaching role, providing feedback during ongoing tasks.
Only after these stages does BorgWarner consider applications in which AI generates results independently, initially with human review. Fully autonomous operation remains limited to suitable use cases.
"Just because you're using AI, it does not absolve you from the responsibility of the decisions being made," Beemer emphasised.
Safety- and security-critical decisions therefore continue to require human oversight. Natsume similarly stressed that AI models must be validated against physical manufacturing processes before their conclusions can be trusted.
Kalamdani suggested that relatively straightforward maintenance applications could provide an appropriate starting point for automation. For example, a CNC machine requiring maintenance after 50,000 operating cycles could automatically trigger a work order when that threshold is reached.
Such clearly defined, rules-based applications offer immediate opportunities to reduce manual intervention without introducing the uncertainties associated with more complex predictive decisions.
Looking further ahead, Kalamdani also distinguished between manufacturing processes with different levels of automation potential. Body shops, where welding and joining operations are already extensively automated, appear better suited to greater autonomy than final assembly, which continues to involve substantial manual work.
Scaling AI requires manufacturing expertise
Turning successful pilots into solutions that work across multiple plants requires more than technical validation.
Beemer explained that BorgWarner increasingly evaluates AI projects from a systems perspective, considering the different functions, stakeholders and processes affected from the outset. This helps determine whether an application is suitable for broader deployment.
Sethi described how Autoliv is bringing IT specialists, manufacturing IT teams and automation engineers together to work on specific business problems. He also sees opportunities for technology suppliers to accelerate adoption by offering data models tailored to the automotive industry's requirements.
Kalamdani emphasised that decisions about developing AI capabilities internally or relying on external providers must reflect the manufacturer's specific requirements. While specialised technology suppliers can offer established solutions, ongoing support and experience with the relevant manufacturing processes are important considerations.
Another key issue is preserving engineering knowledge as AI takes over repetitive tasks. An audience question raised concerns about how younger engineers will develop expertise if established processes increasingly become automated.
Beemer sees an opportunity in AI-supported coaching, where knowledge contributed by experienced engineers can help younger employees improve their work. BorgWarner is also teaching employees basic data science skills to improve collaboration between process specialists and AI developers.
The wider conclusion was that greater autonomy does not eliminate the need for engineering expertise. AI can accelerate analysis, automate routine activities and make accumulated manufacturing knowledge more accessible. But manufacturers still need people who understand physical processes, question unexpected results and take responsibility for production decisions.
Ultimately, the transition from connected data to virtual factories depends less on the sophistication of individual AI applications than on manufacturers' ability to integrate reliable data, process knowledge and digital tools into everyday operations.
