Where AI actually delivers in manufacturing: lessons from Automate Live

At Automate Live in Chicago this month, three industrial practitioners answered a question manufacturers have been circling for years: where, exactly, does AI produce results on the factory floor? The conversation, moderated by Jimmy Carroll of the Manufacturing Matters Podcast, featured Darcy Bachert, CEO of Prolucid Technologies; Charlie Long, VP and GM of Machine Vision and Fixed Industrial Scanning at Zebra Technologies; and Michele Silva, engineering manager at Reynolds & Moore. According to reporting by A3's online team, the panel's clearest finding was that the technology itself is less often the bottleneck than the data, infrastructure, and engineering decisions that surround it.

Rules-based vision still wins, until variability spikes

Conventional, deterministic machine vision remains the right tool for a wide range of industrial tasks. When a measurement is consistent and the acceptable range is well-defined, rules-based algorithms are fast, auditable, and low-maintenance. The calculus shifts when variability enters the picture.

Panelists described several application categories where AI now closes gaps that traditional programming cannot: OCR on damaged or inconsistent labels, complex surface inspections with natural material variation, and multi-modal inspections that fuse visual data with temperature, vibration, or production metadata. Autonomous mobile robot perception was also cited as an area where AI-driven pattern recognition outperforms hand-coded logic.

The panel was careful to note that AI should extend conventional vision systems, not displace them. Operators who layer AI onto applications that a deterministic algorithm already handles reliably risk adding complexity without adding value.

Data preparation is where most projects fail early

A consistent theme across the discussion was that teams routinely underestimate how much work sits between a compelling AI demo and a production-ready deployment. Structured, high-quality datasets built from real production conditions are the prerequisite, not the afterthought.

The panelists suggested operations and engineering teams work through a set of foundational questions before any model training begins: Is the available data accurate and truly representative? Does this specific application require AI, or would a simpler solution suffice? Are there enough real-world examples to train a robust model? And should inference run at the edge or in the cloud? The answers to those questions, they argued, determine project outcomes more reliably than the choice of AI framework or hardware platform.

Edge computing is now a prerequisite for real-time industrial AI

Consumer AI can afford to wait seconds for a cloud response. Industrial AI often cannot. Inspecting hundreds of parts per minute or enabling an AMR to navigate a crowded warehouse aisle requires decisions measured in milliseconds. Communication latency introduced by cloud routing is enough to affect both throughput and safety.

The panel pointed to edge devices equipped with GPUs or purpose-built AI accelerators as the deployment pattern gaining traction on production floors. Cloud infrastructure still plays a role, particularly for training large models on aggregated datasets, but inference is increasingly moving to the machine or the line. For procurement and IT teams evaluating AI platforms, that distinction carries direct implications for hardware budgets and network architecture.

Engineering judgment isn't optional, it's the control system

Panelists pushed back firmly on the idea that AI reduces the need for engineering expertise. In manufacturing, AI outputs still require experienced engineers to validate results, apply domain knowledge, manage cybersecurity exposure, and maintain regulatory compliance. Documentation workflows, quality management processes, and software development assistance were cited as areas where AI lifts productivity, but the engineer remains the accountable decision-maker.

Safety added another layer of nuance. AI-driven perception systems are enabling robots to work alongside people beyond traditional fenced cells, recognizing individuals and adapting to dynamic environments. But the standards governing AI-enabled safety systems are still evolving, and the panel stressed that careful validation and human oversight remain non-negotiable before any such system reaches a live production environment.

What this means for your team

• Audit your data before your model: confirm that training datasets reflect actual production variability, not idealized lab conditions, before committing engineering resources to a deployment.
• Map inference location to latency requirements: if your application requires sub-second decisions, budget for edge hardware with onboard GPU or AI accelerator capability rather than assuming cloud connectivity will suffice.
• Apply the rules-based test first: if a deterministic algorithm already solves the problem reliably, quantify what AI adds before introducing it, complexity has a maintenance cost.
• Track evolving safety standards: teams deploying AI-guided collaborative robots or AMRs near personnel should monitor standards development and build explicit human-oversight steps into commissioning processes.