Chirag Agrawal serves as Global Head of Data Science at Novelis. Novelis Inc. is a global leader in aluminum production and recycling, known for being the largest producer of flat-rolled aluminum products and the world's largest recycler of aluminum. Headquartered in Atlanta, Georgia, Novelis operates 33 manufacturing locations in nine countries and employs more than 14,000 people worldwide.
At Novelis, Chirag has spearheaded transformative technology initiatives since 2019. He previously worked as Data Science Manager at The Kini Group (2016-2018) and held strategic roles at CNH Industrial and Ernst & Young.?
He holds a Master's degree in Machine Learning from Harrisburg University of Science and Technology (2017-2019) and a B.Tech in Mechanical Engineering from IIT Dhanbad. His expertise spans AI/ML, cloud solutions, Industry 4.0, and sustainable manufacturing
I had the opportunity to interview Chirag Agrawal about AIoT and his views on how AI is impacting Industry 4.0 and industrial IoT. Here is the first half of the interview.
Carl Ford (News - Alert) (News - Alert): Chirag, I am looking forward to your keynote at AIoT World Expo, which takes place February 10-12, 2026 in Fort Lauderdale, Florida, and I am hoping you can give a little preview here about your thoughts about AI in manufacturing. Let me ask a specific question to get us started. Most manufacturing uses single purpose PLC to manage production. To me, this means that AI has to happen separately from the production. How do you see it tying back into impact production?
Chirag Agrawal: Before answering your question, I would like to thank you for the opportunity to be a keynote speaker at AIoT World Expo 2026 and for having me here today.
The integration of Artificial Intelligence (AI) into Programmable Logic Controllers (PLCs) is an interesting topic. Deterministic hard real-time control with validation is the primary functionality of PLCs, while the design of AI is primarily probabilistic with a focus on learning through iterations. Integrating both within the PLC can decrease the response time between PID controllers and AI output. At the same time, it may disrupt normal operations, which can become a problem.
I believe in keeping AI as a supervisory layer outside the PLC, where it senses, predicts, and recommends actions. When proven safe, AI can automate setpoints through the existing control strategy. This approach ensures changes are safe, repeatable, and deliver clear ROI.
By running AI on edge or near-line systems with heavier compute, we preserve the PLC’s role in deterministic control and safety. This separation of concerns also supports governance through Management of Change (MoC), enables faster iteration, and ensures fail-safe operation.
CF: Does this mean we should expect to see PLC become more intelligent?
CA (News - Alert) (News - Alert): PLCs will evolve over time, as new technologies emerge and exploration reaches its peak. They'll remain excellent at what they already do, with deterministic, real-time control, but they'll become more connected and more aware of their context. We already see PLC vendors integrating edge compute and AI hooks, so the PLC can take in advisory setpoints or execute lightweight anomaly detection. However, the hard work, deep learning, and optimization will still reside in edge servers or the cloud. PLCs will be “smarter,” but accurate intelligence will come from a layered architecture where PLCs remain the safety-critical nucleus and AI serves as the adaptive brain surrounding them.
CF: A lot of our audience is in the cellular industry and believe that cellular is the future for manufacturing. I believe the majority of industrial operations are still on the WAN and supplementing the plant using solutions such as BLE or Wi-Fi. Is this your experience and what do you expect to change in the near future?
CA: In factories today, the control and data infrastructure is still mostly hardwired (industrial Ethernet) with Wi-Fi and short-range radios (e.g., BLE) as support, divided by Purdue/CPwE with an IDMZ and edge compute at L2.5. We're architecting new solutions to be 4G/5G-ready on the edge, and we anticipate cellular will expand selectively where mobility, coverage, and quality of service dictate it, but not to displace wired or Wi-Fi in its entirety.
For large outdoor spaces and mobile equipment (forklifts, AGVs, mobile robots) where Wi-Fi roaming and coverage are challenging, mid-band cellular provides range, penetration, and controlled quality of service. For instance, in harsh RF/metal-abundant environments, where more power and licensed spectrum minimize interference vs. common Wi-Fi, particularly across several buildings, or temporary/quick stand-ups and redundancy (e.g., commissioning, turnarounds) where a fiber drop isn't feasible.
Wired connectivity is required for deterministic control (wired/TSN and APC (News - Alert) (News - Alert) in L1/L2); cellular is more suited for supervisory and telemetry applications today, rather than inner control loops. High-throughput indoor density with numerous clients and limited mobility; Wi-Fi 6/6E/7 economics are favorable within process halls compared to setting up a complete private cellular core.
CF: I recently came to the conclusion that when AI agentic solutions get adopted, AI will become the actuators tightly coupled to sensors. Is that a fair statement?
CA: Partially true. Agentic AI will shrink the sensing-acting gap but, in most industrial applications, AI won't become actuators themselves. It will instead control or manage actuators by a supervisory layer, typically through PLCs, edge controllers, or digital twins. Conceive AI as the decision-making brain sending commands, while physical actuators are still electromechanical devices.
CF: What does this mean for the future of Digital Twins? Will the model for digital twins be expanded by AI, or should we expect a new paradigm to occur?
CA: Earlier, digital twins mirrored physical assets or processes, showing what was happening in real time. They run simulations to forecast what might happen next (e.g., “If we increase speed, what’s the impact on quality?”), primarily a decision-support tool for humans.
Agentic AI enables goal-oriented behavior, allowing twins to coordinate workflows, negotiate trade-offs (e.g., energy vs. throughput), and initiate actions autonomously under governance.
Old approach: A twin predicts a pump will fail in 10 days, sends an alert, and a human schedules maintenance.
New approach: The twin predicts failure, checks spare parts availability, reschedules production, and automatically creates the work order, with human approval or even fully autonomously in low-risk cases.
This moves us from “observe and advise” to “sense, decide, and act”, a fundamental change in its role. Digital Twins are evolving from passive mirrors to active decision-makers. AI transforms them from ‘what-if’ tools into systems that can plan and act in real-time.
Read part 2 of Carl’s interview with Chirag Agrawal here.
Don’t miss Chirag Agrawal’s keynote at AIoT World Expo, taking place February 10-12, 2026 at the Broward County/Greater Fort Lauderdale Convention Center in Fort Lauderdale, Florida. His keynote will take place Wednesday, February 11, 2026, 11:30am. AIoT World Expo is the premier event for exploring the convergence of Artificial Intelligence (AI) and the Internet of Things (IoT) and serves as the gathering point for industry professionals to discover advancements, market opportunities, and understand the transformative power of AIoT across industries. Register today for AIoT World Expo, part of the #TECHSUPERSHOW.