Physical AI is the enabling technology for the industry’s next step, from automation to autonomy

Ritwij Kulkarni, Vice President & General Manager, India & APAC – Process Automation, Honeywell Technologies.

Physical AI is being positioned as the next evolution of smart manufacturing. What, in your view, distinguishes Physical AI from the AI, automation and connected technologies that manufacturers have already deployed, and why is it becoming important now?

Conventional automation executes pre-defined logic exceptionally well, but when exceptions arise, they are still largely managed by people Over the years, industrial operations have generated large volumes of data from sensors, physical assets and dashboards. However, this data has often lacked context, connection to physical assets and the ability to drive action in real time. Generative AI, on the other hand, is probabilistic in nature. That works well for knowledge tasks, but in operational technology, an operator needs near-absolute certainty before acting on a recommendation.

Physical AI bridges this gap because it has a deterministic core. It learns within the known laws of physics, chemistry and thermodynamics, as well as mass and energy balances, so design limits and engineering standards act as built-in guardrails. It works in a continuous loop of understanding, thinking, acting and learning in the context of physical operations, which means it keeps getting better over time.

It is becoming important now because several structural pressures are converging at once. Experienced operators are retiring, a large share of assets is ageing, traditional optimisation approaches are reaching their limits, and market volatility demands faster responses than conventional planning cycles allow. Physical AI is the enabling technology for the industry’s next step, from automation to autonomy.

As AI moves from analysing data to influencing decisions and physical processes, how could Physical AI change the way factories operate on a day-to-day basis? Which manufacturing decisions and workflows are most likely to become increasingly autonomous?

Day-to-day operations will shift from reactive to predictive and, eventually, prescriptive. Today, many process issues become visible only once an alarm is triggered. With Physical AI, abnormal situations can be detected earlier and addressed before they escalate. In pilots of our Experion Cognition, for instance, the platform made predictions an average of five to ten minutes before alarm incidents.

The decisions and workflows most likely to become autonomous first are those that are well-defined and repetitive. These include detecting and responding to abnormal situations, and monitoring sensor health so that an approved response is applied automatically when a sensor freezes or drifts. They also include routine procedures during start-ups, shutdowns and abnormal events, which can otherwise overwhelm operators with manual interventions. Maintenance is another area, where AI-driven assistants help teams diagnose issues and shorten the time to resolution. Over time, AI agents will coordinate with one another across the control and optimisation layers, while operators focus on supervision, optimisation and higher-value decisions.

Labour availability and skills gaps remain important challenges for Indian manufacturers. How can AI-driven autonomous workflows help address these constraints while maintaining process discipline, quality, and operational consistency?

The challenge for manufacturers is not only the availability of people, but the loss of experience as seasoned operators and engineers approach retirement. AI-driven autonomous workflows can capture that critical knowledge from procedures, documentation, operational history and subject-matter experts. They can then convert it into actionable guidance available to every operator within their workflow. By delegating routine cognitive tasks to autonomous agents, less experienced operators can run plants with knowledge comparable to that of a seasoned veteran.

This standardised, context-aware guidance also reduces variability across shifts. It helps teams make faster, more accurate decisions and achieve more predictable performance, which directly supports process discipline, quality and operational consistency. At the same time, automating routine tasks, procedures and responses reduces operator workload and allows teams to increase their span of control. Because Physical AI operates within known physical and engineering boundaries, this autonomy reinforces operating discipline rather than diluting it.

For Indian manufacturers looking to scale automation across multiple plants, what role can Physical AI play in making operations more scalable and adaptable—particularly when factories have different processes, equipment, legacy systems and workforce capabilities?

Scaling automation across multiple plants starts with recognising that most factories will not replace their existing systems overnight. The more practical approach is to add AI, cloud and edge capabilities to the control systems already running each plant, so that they can learn continuously, predict what is coming and amplify human expertise. A common intelligence layer such as Honeywell Forge connects and contextualises data from industrial assets. This gives manufacturers a consistent foundation across sites with different processes and equipment. Solutions such as Experion Cognition are also designed to integrate into existing control room environments.

Physical AI also makes operational knowledge easier to digitise. Once procedures and operational workflows are captured and automated, they can be extended well beyond a single unit. We are already seeing this progression globally. Following its proof of concept with Honeywell, Borouge International is assessing opportunities to scale autonomous operations across its Ruwais and global facilities.

Designing with scalability in mind from the outset is equally important. A maturity-based framework helps manufacturers understand where each plant stands on its autonomy journey, so that important steps are not skipped and sites with different capabilities can progress in a structured way.

The moment AI begins influencing physical operations, trust becomes critical. What governance mechanisms, human oversight and validation frameworks are necessary before manufacturers can confidently allow AI to make or execute operational decisions?

When AI begins to influence physical operations, trust must be engineered into the system from the start. The first element is boundaries. Physical AI learns within the laws of physics, chemistry and thermodynamics, so design limits, reliability constraints and decades of engineering standards become guardrails embedded in the AI itself. The second is visibility. Operators must be able to see why the system is recommending a particular action. Explainability is key to overcoming the trust deficit that has historically slowed AI adoption in safety-critical operations.

The third element is human oversight. Human-autonomy teaming provides operators the opportunity to review, approve, monitor, or override actions while autonomous agents conduct routine monitoring and response activities. Autonomous actions should be restricted to validated responses, so that the system only recommends or performs what has been validated. 

Manufacturers should also adopt autonomy in phases rather than all at once. A structured maturity model provides a clear framework and set of guidelines for integrating AI. It ensures that critical steps are not skipped as plants progress from AI-assisted to more autonomous operations.

Greater connectivity and autonomy can also expand the cybersecurity risk landscape. How should manufacturers rethink OT security and cyber resilience as AI systems increasingly interact with industrial control systems, machines and critical processes?

As AI takes on a greater role in physical operations, autonomy and cyber resilience must advance together. Industrial organisations face increasingly sophisticated threats targeting critical operations. Those with a unified cybersecurity strategy across both digital and physical environments are best placed to build operational resilience and avoid potentially dangerous and costly attacks.

In practice, this means moving from a reactive to a proactive security posture. AI and behaviour-based analytics can establish a baseline of normal system operations and detect anomalies in OT cyber behaviour before they turn into attacks. Manufacturers should also inspect removable media and storage devices before they enter critical environments, secure the data pathways between systems, and continuously monitor OT environments for early signs of an attack. Their security programmes should align with industrial standards such as ISA/IEC 62443 and the NIST Cybersecurity Framework.

Architecture matters too and deploying industrial AI securely on-premises limits exposure. Much like IT-OT integration, cybersecurity must be aligned and designed in from the outset, as it is fundamental to a strong digital foundation for any plant.

Rather than replacing people, could the more practical model for Indian factories be a ‘human-plus-AI’ workforce? How can manufacturers redesign roles and workflows so that AI augments human expertise while improving worker safety, productivity and decision-making?

Yes, a human-plus-AI model is both the practical and the right approach. Even as the shift toward autonomy accelerates, the most essential element of that change remains people. The objective of industrial AI is to help physical assets work harder and people work smarter, amplifying human expertise rather than replacing it.

This changes how roles and workflows are designed. Operators can move away from repeatedly responding to routine abnormal situations and instead focus on optimisation and higher-value activities, while increasing their span of control. The knowledge of the most experienced personnel can be captured and made available to every operator as best-practice guidance. Maintenance teams, in turn, can use AI-powered assistants to diagnose issues and resolve them faster.

Worker safety and productivity also improve. The experience gap created by an ageing workforce has been driving up human-error-related upsets, unplanned shutdowns and production losses. By detecting and mitigating abnormal situations before they happen, AI can help reduce process mistakes and downtime, supporting safer operations and better decision-making on the shop floor.

Looking ahead, what does Honeywell Technologies see as the roadmap for Physical AI in Indian manufacturing? What needs to happen across technology, infrastructure, skills and organisational culture for Indian factories to move from AI-assisted operations towards genuinely autonomous and scalable production?

Honeywell Technologies views autonomy as a journey along a continuum rather than a single leap. Plants typically progress from controlled and optimised operations to intelligent, remote and resilient operations, before reaching fully autonomous operations. For Indian manufacturers, moving along this continuum will require progress across technology, infrastructure, skills and culture together.

On technology, the foundation is deterministic, physics-grounded AI that operators can trust, deployed across cloud and edge environments. On infrastructure, manufacturers need connected and contextualised data from their installed assets, along with robust OT cybersecurity. Skills are equally critical, as demand for AI and cybersecurity talent in India continues to outpace supply. 

Organisations should begin with use cases that deliver measurable outcomes, build operator trust through explainability, and align leadership vision with plant-level KPIs. India is well positioned for this transition. It is one of Honeywell Technologies’ fastest-growing markets, and the innovation coming out of India increasingly combines software, hardware and basic research. That makes India a strong base for co-creating solutions not just for India, but for the world.