The loudest AI conversations in manufacturing tend to orbit the same themes: chatbots, copilots, and the looming question of job displacement. Bijan Saeerozari thinks the industry is looking in the wrong direction. The director of AI at Rockwell Automation’s Center of Excellence argues that AI in manufacturing isn’t overhyped — it’s underhyped. The capabilities that matter most are only beginning to emerge, and they live not in the cloud but at the edge of the plant floor, where milliseconds and trust determine whether a technology gets adopted or ignored.

Saeerozari joined Tim Wilson on SME Media’s Advancing the Line — AI in Manufacturing podcast to lay out how edge computing, self-learning engines, and generative AI are converging to give operators something they’ve never had: a data-driven assistant that earns its keep by making their jobs less stressful, not more complicated.

Saeerozari’s career began at a Stanford-born startup that was among the first companies to attempt using neural networks for model-predictive control. That startup was eventually acquired by Rockwell, and Saeerozari has spent the years since assembling a team that works at the intersection of machine learning, control systems, optimization, and complex system modeling.

His framing for what AI offers a plant manager is deliberately simple: a new set of eyes. Any manufacturing facility generates thousands of data points simultaneously — temperature, line speed, moisture, pressure — and extracting actionable insight from that data has traditionally required enormous engineering overhead. AI can identify multi-variable relationships and surface patterns that would take human analysts far longer to find, if they found them at all. More importantly, it can translate those insights for any stakeholder in the organization, not just the engineer who understands the raw data.

When Saeerozari’s team engages with plant managers, the conversation almost always turns to edge computing — and the preference is driven by physics, not philosophy. Two factors dominate the discussion.

The first is data security. Manufacturers are protective of proprietary process data and reluctant to expose it beyond the plant’s walls. An edge-based solution avoids creating new attack surfaces entirely.

The second is latency. In a high-speed stamping operation, decisions must be made in milliseconds. The round-trip delay of sending data to the cloud and waiting for a response is not merely inconvenient — it’s operationally unacceptable. A decision engine that sits close to the machine eliminates that lag.

A solution at the edge does not have to expose any new surface for outside attack, and it manages the latency and response effectively. The justification for sending data out and bringing it back needs to be weighed against whether the entirety of the decision is available at the edge.

Saeerozari is emphatic that edge AI is not a transitional technology. Even in a future with better cloud infrastructure and more robust cybersecurity, processing decisions at the point of operation will remain the most optimal engineering solution for many applications. The cost of edge hardware is falling simultaneously, removing what was once a significant barrier to adoption.

Rockwell’s approach deploys self-learning engines at the edge that consume operational data and make recommendations within guardrails defined by the plant. The deployment follows a deliberate progression: advisory mode first — where the AI surfaces information for human review — and closed-loop decision-making only after trust has been established.

That trust, Saeerozari stresses, is the make-or-break factor. If operators don’t understand what a system is recommending, they won’t act on it. Full stop. Building explainable models is the baseline, but it isn’t sufficient on its own. An optimization engine might determine that a facility should shut down Chiller A and start Chiller B for energy efficiency — but if the operator can’t see why that recommendation was made, the system will never run in closed loop.

This is where generative AI enters the picture, and Saeerozari’s team is using it in a way that sidesteps the hallucination concerns that dog GenAI in other contexts. Rather than generating answers from scratch, they use generative AI as a translation layer — converting complex mathematical outputs from proven machine learning models into plain-language explanations that operators can evaluate. The accuracy comes from the underlying decision engine. The explainability comes from GenAI. Every engine Rockwell produces now carries this explainability layer as standard.

The workforce conversation around AI in manufacturing typically centers on displacement. Saeerozari reframes it around agency. For the foreseeable future, operators will still be needed. What changes is how much support they have.

His example is vivid: an operator responsible for a wastewater pumping station watches the reservoir level begin to rise unexpectedly. If it spills, it’s an environmental disaster. The stress is immediate and intense. A data-driven decision engine that can simultaneously analyze influx rates, pump availability, and reservoir capacity doesn’t replace the operator — it assists with thinking under pressure, potentially making a better decision and certainly reducing the cognitive burden of making it alone.

The right AI technology can give that operator more agency to perform their functions better. It reduces the stress of the operation. And perhaps it frees up time for more productive, more business-oriented output that otherwise they would not have.

A maintenance technician confronted with an unfamiliar fault code can consult an AI assistant that retrieves the relevant manual, identifies the steps, and walks through the response. The technician remains in charge. The AI handles the lookup. The result is faster resolution and a worker whose expertise grows with every interaction rather than being replaced by it.

Saeerozari’s framework for moving past the pilot stage is characteristically practical. First, identify a problem that genuinely matters to the operation — the proof being that the relevant people actually show up to the meeting. Second, design the solution to be self-maintaining, because any system that burdens an already-stressed operations team to keep running will be abandoned. Third, build for transferability. In manufacturing, no two lines or plants are identical, and a solution that can’t adapt to similar-but-different conditions will never scale.

The deeper point is that AI adoption in manufacturing won’t be driven by executive mandates or vendor demonstrations. It will be driven by operators who see a system reduce their stress, improve their decision-making, and respect their expertise — and who then refuse to go back to working without it.

Shares: