Agentic AI Enters The Physical World
Krish Ghosh is CEO/CTO of DreamzTech US.
gettySo far, enterprise AI and agentic AI have focused mainly on things like chatbots, content creation, coding and knowledge assistance. Physical operations are the next logical space, and manufacturing companies and facility management firms are already deploying agentic AI in operational workflows. But the real world is less forgiving than those earlier applications, and companies need to approach implementation carefully.
A manufacturing facility typically runs many separate systems, including ERP, CMMS, field service management and warehouse management, with data streaming in from RFID and IoT sensors. All of that data is fragmented, and when something breaks, a technician often has to manually open the maintenance manual, review previous job records, check the ERP and call their supervisor before they can decide what to do. That process can take hours.
An AI agent can coordinate across all of those systems and respond in seconds. When a sensor detects an anomaly, it can pull the machine’s maintenance history, previous failures, relevant manuals, current parts inventory and available technicians with matching skills, then recommend the next action.
Work that previously required four or five people manually collecting data and checking dashboards is now a single coordinated response.
One of the biggest mistakes I see COOs and plant managers make is trying to implement AI everywhere all at once. Instead, it’s important to choose a specific problem first, whether that means downtime, maintenance response, inventory discrepancies, service prioritization or asset visibility. You should define the business outcome and the KPIs you’ll track before you evaluate any technology.
From there, map exactly how your process works right now: what the agent will observe, what context it needs, which systems it uses, what action it takes and where human approval is required. The agent will be trained on your actual process, and every business operates differently.
Then assess your data quality honestly. An agent needs trustworthy data, including asset history, current condition, location, maintenance records or inventory. I’ve seen a lot of early adopters put AI on top of poor data and be disappointed.
It’s critical to build controls before you deploy, including permissions, approval thresholds, audit trails, escalation paths, exception handling, manual overrides and cybersecurity. Permissions need to be explicit: What can the agent read, recommend, initiate or approve, and what can it never do? Every action should be auditable: what the agent observed, what information it used, why it recommended an action, who approved it and what happened afterward. You also need fallback procedures for when data is missing, systems are unavailable or the agent is uncertain.
Plant managers, technicians and supervisors should be involved from the beginning because they understand practical floor conditions that technical teams often miss. They’ve built substantial institutional knowledge about how things actually work, and that context is critical for training the model correctly.
You cannot take whatever AI produces in a production environment and simply act on it, because at the current level of development, AI agents can still sometimes hallucinate and suggest something absolutely wrong.
That’s why I advise starting with a narrow pilot where AI can deliver clear, measurable output, one that requires a thumbs-up or thumbs-down from a person on every recommendation. When the model is wrong, give it the overall context for why. That feedback trains it to understand your environment better over time. Increase authority only after you’re confident it has the required knowledge and judgment.
One of the most valuable things an agent can do in physical operations is work with an institutional knowledge base. I’ve seen businesses feed years of maintenance history and resolution records into a system, so that when an inexperienced technician faces a machine failure and a supervisor isn’t immediately available, the agent can walk the technician through exactly what was done before to fix that machine.
The future is not autonomous everything; it’s controlled intelligence connected to real operational context. Even though new models are coming online constantly, context and integration matter more in physical operations than which model you’re running. Within the next few years, I expect a significant shift to on-premises local models, as businesses look to avoid token costs and keep sensitive operational data out of the cloud.
I’m optimistic about the future of AI in physical operations. It can be extremely valuable if done correctly, but it will only work if we take responsibility for setting it up carefully from day one.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
