ER&D GCCs, learn to use vision AI to turn complex manufacturing data into outcomes

This HFS Highlight is for CTOs, CIOs, and ER&D GCC leaders building vision AI capability for defect detection, real-time process monitoring, and shop-floor safety in manufacturing.

Improving speed, scale, and product quality, and employee safety continues to be a persistent challenge for manufacturers. In the semiconductor industry, vision AI is emerging as a powerful tool to automate the 2,000-plus steps involved in chip production by identifying areas the human eye can’t reach. CTOs and CIOs have an opportunity to use this technology to improve quality and throughput while ensuring shop-floor safety, potentially saving their enterprise millions of dollars.

ER&D GCCs, including those operating offshore, can bring vision AI’s capabilities to the plant. The tech integrates hardware, software, firmware, and human intelligence to not only identify defects much more accurately but also monitor shop floors, track manufacturing performance, and improve overall efficiency.

These opportunities, as well as the challenges of using vision AI, were discussed at the TechForward session with ER&D GCC leaders, held in collaboration with Micron, IIIT-H, and HYSEA, in July 2026.

ER&D GCC Leaders at the Tech Forward Roundtable hosted by Micron Technology India

In the picture: Leaders from Micron, Thermo Fischer, Enovix, Silicon Labs, Cognizant, Octave, Hexagon, BOSCH, Qualcomm, Stellantis, and HYSEA, and IIIT-H professors

Vision AI works as an additional pair of eyes for automating chip manufacturing

ER&D GCCs are already leveraging vision AI in limited use cases such as monitoring occupancy ratios for assets, inventory, and raw materials to maximize usage, and detecting defective products. An HFS survey identifies its growing potential in manufacturing (see Exhibit 1). When integrated with multimodal AI, GCCs can expand its scope by experimenting, failing fast, and learning to deploy more quickly to interpret a wider range of data to drive better outcomes and ROI.

Exhibit 1: Real-time monitoring and defect detection is the next AI frontier which has huge potential

Sample size: 31 Forbes Global 2000 manufacturing enterprise decision makers on the AI use cases; The total does not add to 100% due to multiple selection options
Source: HFS Research Pulse, December 2025

ER&D GCCs can use vision AI to streamline the 2,000–2,500 steps involved in chip manufacturing, but integrating it across environments is complex. It requires coordination across various AI models, cloud services, and other components influenced by the plant, process, and personnel to capture data, analyze it, interpret the results, and trigger the right action.

Service providers are ahead in exploring the potential of vision AI in engineering decisions

HFS Agentic and GenAI Compass points out that service providers are already addressing some of these complex challenges. Some use cases include defect detection, physical testing of automobiles, and reduced incidents in plants, all made possible by integrating vision AI into edge devices, with cameras trained on vision language models (VLMs).

Here are some examples of provider use cases:

• TCS worked with NVIDIA Omniverse to help a European automaker that makes commercial trucks. It created a camera calibration application to check systems virtually, eliminating the need for physical tests in the field. Instead, it used vision AI to inspect 10–14 foot-long truck chassis at the control center, reducing manual work and improving output quality and productivity.
• Cognizant launched a sovereign physical AI platform on the Cognizant Intelligence Spine, which unifies computer vision, sensors, video, SCADA, and edge into a single, governed architecture that can be implemented on any shop floor.
• LTTS, as part of its new Engineering Intelligence (EI) initiative, leverages VLM to let machines interpret real-world data, assess conditions, and act accordingly to keep humans or assets safe by retracting from the object in hazardous situations.

These use cases show that applying vision AI needs a strong data foundation, continuous model training, low latency, and skilled AI talent to effectively deploy these solutions and achieve the right outcomes.

Moving vision AI beyond use cases means overcoming data, model, interoperability, and talent challenges

The leaders at the roundtable talked about many challenges. Four areas emerged that needed attention:

• Data foundation and model training: Traditional automated optical inspection systems often lack full accuracy and produce many false positives. A single 300mm wafer generates about 70k images, and with 30k to 50k wafers produced daily, that’s around 100 million images monthly. This requires real-time, high-speed processing, making manual inspection of every image impractical.
• Continuous model retraining: When process changes occur, new materials are introduced, lens contamination happens, and lighting variations degrade the model, leading to defects downstream. Models should be retrained every three to six months due to process drift and new equipment. While self-supervised learning methods are being explored, constant changes in products and processes make it difficult to maintain adequate training datasets.
• Interoperability and latency constraints: A model trained on one tool, one fab, or one product variant fails to generalize to the next fab. This requires hybrid approaches that combine ML with physics-based models rather than pure deep learning. While cloud provides significant compute power, communication latency can become a bottleneck. Some low-latency cases require on-premises GPU deployment rather than relying solely on cloud infrastructure.
• Limited SMEs: The semiconductor industry is not facing an AI/ML talent scarcity. Rather, domain knowledge is concentrated among a few experts and the data team, which can’t be transferred downstream due to the governance and protocols of the companies.

Addressing these challenges requires more than improving individual models. ER&D GCCs must build a vision AI capability model that combines a reusable image data foundation, continuous model training, and hybrid edge-cloud architectures that balance interoperability, accuracy, and latency across processes and plants to deliver measurable improvements in quality, throughput, efficiency, and ROI.

The Bottom Line: Vision AI will deliver value at scale only when ER&D GCCs leaders build the data foundation, industrialize models, and develop specialized talent to move manual steps into measurable ROI.