<h2><strong>Executive Summary</strong></h2><ul><li><p><strong>Physical AI</strong> (systems that sense, decide and act in the real world) <strong>is the next frontier</strong>, and most organisations have not yet worked out what effective deployment requires.</p></li><li><p>Running <strong>AI on devices</strong> rather than in the cloud is emerging as <strong>a cheaper, faster and more secure default option</strong> for frontline work.</p></li><li><p>India's mix of <strong>engineering talent, cost discipline and a people-first operating model is a genuine advantage</strong> as AI moves off the screen and onto the factory floor.</p></li><li><p>Adoption increases or decreases, <strong>depending on how workers experience the technology.</strong> </p></li><li><p>A narrow, well-chosen use case, proven early, <strong>builds the credibility needed to attempt something larger later</strong>.</p></li></ul>.<p>Manufacturing has moved well past the pilot phase of AI adoption. Today, instead of just analysing the data that comes out of such systems, companies are testing whether they can act reliably in the physical world. More and more companies are using AI across supply chains, quality control and predictive maintenance, but the persistent gap between deployment and real value creation is hard to ignore. India's manufacturing sector, which is expanding rapidly but remains heavily people dependent, sits at an inflection point. At a recent joint session of the India CEO and CFO Forums in Bangalore, Tom Bianculli, Chief Technology Officer of Zebra Technologies, examined where AI is creating measurable value on the factory floor, what a genuine enterprise scale deployment of physical AI demands and how organisations should sequence their investment.</p><h2><strong>From Reactive AI to Agentic Execution</strong></h2><p>Frontline AI has moved through three stages:</p><ol><li><p>A reactive, chatbot style tool, where a worker types a question and gets an answer. This model is hard to justify to CFOs because there is no clear way to measure the savings it generates.</p></li><li><p>Context aware agents. This is a step up, because the model understands the task in front of a worker and breaks it into steps rather than leaving them to work out what to ask.</p></li><li><p>Agentic execution, where a companion layer sits between a device's operating system and its applications, and simply gets the workflow done on its own.</p></li></ol><p>To illustrate, a material receiving task, which took 26 seconds under a manual entry process, required 7 seconds with auto-filled label scans, but less then 1 second once the process became fully autonomous.</p><h2><strong>Why the Economics Favour the Edge Over the Cloud</strong></h2><p>In Zebra's frontline deployments, most AI tasks run on the device or within the facility, reducing cloud costs, response times and data exposure. Cost control matters as much as infrastructure. Zebra uses financial operations (FinOps) to set monthly budgets for teams and track AI spending in near-real-time. Firms could also consider using a hybrid model, running frequent workloads on their own servers and sending less frequent, compute-heavy tasks to the cloud. An early estimate suggests that this could triple usage while increasing costs by about 30%. For factories that are fully disconnected from the internet, larger workloads can run on a local edge appliance instead of the public cloud. </p><h2><strong>What Feeds Physical AI</strong></h2><p>The main challenge is giving the AI model a continuous view of what is happening on the factory floor. Handheld devices can combine cameras, RFID, Bluetooth tags and 3D sensors, allowing them to capture context continuously rather than through individual barcode scans. When this data is combined with a company's ERP, inventory and transport systems, a domain-specific model can understand what should be happening and trigger an action when something goes wrong. It could, for example, warn that a delayed trailer may cause the company to miss a delivery commitment. Capturing how experienced workers perform tasks can also help spread knowledge across sites and improve consistency. This matters because much of the knowledge within manufacturing organisations is tacit, meaning it is held by workers rather than recorded in formal procedures.</p><h2><strong>India's Case for a Different Path</strong></h2><p>India's manufacturing AI story does not have to repeat the automation-first path taken in the West. The country went straight to mobile-first computing rather than first building out fixed-line infrastructure. Resultantly, quick commerce now runs on delivery windows measured in minutes. Because so much of Indian manufacturing still runs on the back of people rather than machines, human-AI collaboration rather than replacement may work best in this setting. This is close to what industry analysts call ‘decision entrepreneurism’, pushing the authority to notice a problem and fix it as far down toward the frontline as it will go, rather than up through several layers of approval. India's combination of workforce scale, technical talent and a people-dependent manufacturing model could become a competitive advantage.</p><h2><strong>Sequencing the Change</strong></h2><p>Getting this right depends heavily on change management. A practical starting point is to map workflows end-to-end, select the first few use-cases where real-time physical data can create the most value, and compare the current process with the proposed one. Once these use-cases show results, the organisation can build internal support and move towards a broader plan for the plant and its workforce. Trying to introduce the full transformation in one go creates too much change for the organisation to absorb. Adoption also depends on how the technology is introduced. Workers are more likely to accept systems that remove steps, reduce training requirements and give them greater control than systems that appear to monitor them. Management support and clear expectations remain central to whether the deployment succeeds.</p><p>Ultimately, the aim is to shorten the cycle from sensing what is happening to directing and completing the right action, then using the resulting data to improve the next cycle. The faster a factory can repeat this process, the more responsive and consistent its operations become.</p>
<h2><strong>Executive Summary</strong></h2><ul><li><p><strong>Physical AI</strong> (systems that sense, decide and act in the real world) <strong>is the next frontier</strong>, and most organisations have not yet worked out what effective deployment requires.</p></li><li><p>Running <strong>AI on devices</strong> rather than in the cloud is emerging as <strong>a cheaper, faster and more secure default option</strong> for frontline work.</p></li><li><p>India's mix of <strong>engineering talent, cost discipline and a people-first operating model is a genuine advantage</strong> as AI moves off the screen and onto the factory floor.</p></li><li><p>Adoption increases or decreases, <strong>depending on how workers experience the technology.</strong> </p></li><li><p>A narrow, well-chosen use case, proven early, <strong>builds the credibility needed to attempt something larger later</strong>.</p></li></ul>.<p>Manufacturing has moved well past the pilot phase of AI adoption. Today, instead of just analysing the data that comes out of such systems, companies are testing whether they can act reliably in the physical world. More and more companies are using AI across supply chains, quality control and predictive maintenance, but the persistent gap between deployment and real value creation is hard to ignore. India's manufacturing sector, which is expanding rapidly but remains heavily people dependent, sits at an inflection point. At a recent joint session of the India CEO and CFO Forums in Bangalore, Tom Bianculli, Chief Technology Officer of Zebra Technologies, examined where AI is creating measurable value on the factory floor, what a genuine enterprise scale deployment of physical AI demands and how organisations should sequence their investment.</p><h2><strong>From Reactive AI to Agentic Execution</strong></h2><p>Frontline AI has moved through three stages:</p><ol><li><p>A reactive, chatbot style tool, where a worker types a question and gets an answer. This model is hard to justify to CFOs because there is no clear way to measure the savings it generates.</p></li><li><p>Context aware agents. This is a step up, because the model understands the task in front of a worker and breaks it into steps rather than leaving them to work out what to ask.</p></li><li><p>Agentic execution, where a companion layer sits between a device's operating system and its applications, and simply gets the workflow done on its own.</p></li></ol><p>To illustrate, a material receiving task, which took 26 seconds under a manual entry process, required 7 seconds with auto-filled label scans, but less then 1 second once the process became fully autonomous.</p><h2><strong>Why the Economics Favour the Edge Over the Cloud</strong></h2><p>In Zebra's frontline deployments, most AI tasks run on the device or within the facility, reducing cloud costs, response times and data exposure. Cost control matters as much as infrastructure. Zebra uses financial operations (FinOps) to set monthly budgets for teams and track AI spending in near-real-time. Firms could also consider using a hybrid model, running frequent workloads on their own servers and sending less frequent, compute-heavy tasks to the cloud. An early estimate suggests that this could triple usage while increasing costs by about 30%. For factories that are fully disconnected from the internet, larger workloads can run on a local edge appliance instead of the public cloud. </p><h2><strong>What Feeds Physical AI</strong></h2><p>The main challenge is giving the AI model a continuous view of what is happening on the factory floor. Handheld devices can combine cameras, RFID, Bluetooth tags and 3D sensors, allowing them to capture context continuously rather than through individual barcode scans. When this data is combined with a company's ERP, inventory and transport systems, a domain-specific model can understand what should be happening and trigger an action when something goes wrong. It could, for example, warn that a delayed trailer may cause the company to miss a delivery commitment. Capturing how experienced workers perform tasks can also help spread knowledge across sites and improve consistency. This matters because much of the knowledge within manufacturing organisations is tacit, meaning it is held by workers rather than recorded in formal procedures.</p><h2><strong>India's Case for a Different Path</strong></h2><p>India's manufacturing AI story does not have to repeat the automation-first path taken in the West. The country went straight to mobile-first computing rather than first building out fixed-line infrastructure. Resultantly, quick commerce now runs on delivery windows measured in minutes. Because so much of Indian manufacturing still runs on the back of people rather than machines, human-AI collaboration rather than replacement may work best in this setting. This is close to what industry analysts call ‘decision entrepreneurism’, pushing the authority to notice a problem and fix it as far down toward the frontline as it will go, rather than up through several layers of approval. India's combination of workforce scale, technical talent and a people-dependent manufacturing model could become a competitive advantage.</p><h2><strong>Sequencing the Change</strong></h2><p>Getting this right depends heavily on change management. A practical starting point is to map workflows end-to-end, select the first few use-cases where real-time physical data can create the most value, and compare the current process with the proposed one. Once these use-cases show results, the organisation can build internal support and move towards a broader plan for the plant and its workforce. Trying to introduce the full transformation in one go creates too much change for the organisation to absorb. Adoption also depends on how the technology is introduced. Workers are more likely to accept systems that remove steps, reduce training requirements and give them greater control than systems that appear to monitor them. Management support and clear expectations remain central to whether the deployment succeeds.</p><p>Ultimately, the aim is to shorten the cycle from sensing what is happening to directing and completing the right action, then using the resulting data to improve the next cycle. The faster a factory can repeat this process, the more responsive and consistent its operations become.</p>
