Brandon Barbello is co-founder and COO of Archetype AI, with over 10 years of product leadership across Fortune 500 and startups.
Most of us have had a “wow” moment with an AI agent. As agentic AI has advanced, we’ve seen it help complete more of our digital tasks: booking flights, clearing inboxes and drafting memos. The first time an agent did that for me, it felt like “Clarke’s third law“ made real: technology indistinguishable from magic.
Most AI has lived in the browser. If a digital agent gets something wrong, you catch it, fix it, move on.
But in factories, grids, fleets and roads, a mistake can mean a serious accident or a week of output that never comes back.
The physical world is also where agentic AI’s greatest promise lies, and we might not be too far off from bringing that reality to fruition. The physical world has been producing the data for years. The missing piece has been intelligence that can interpret those signals and help people act on them.
Why Almost No One Can Act On Sensor Data
As COO of Archetype AI, I’ve spent the last few years working with operators and engineers on one of the industry’s most persistent challenges: how to turn the enormous volume of data generated by physical assets into intelligence that helps people make better decisions.
The consequences of failing to make that connection are significant. Siemens reports that unplanned downtime accounted for 11% of annual revenue among the world’s 500 largest companies in 2024, roughly $1.4 trillion, up from 8% in 2020.
Recovery is also getting slower: The average restart time after a stoppage was 81 minutes, up from 49 in 2019. An idle line at a major automotive plant can cost up to $2.3 million per hour of downtime.
After a decade of industrial IoT investment, both numbers moved in the wrong direction. The problem isn’t a lack of data. It’s that much of that data remains difficult to interpret and act on before something goes wrong.
Meanwhile, experienced operators often have a “sixth sense” that lets them predict a machine is going bad before any numbers appear on-screen. The signals are there, but we still don’t have the intelligence to interpret them at scale and generate useful decisions.
Why The First Wave Of AI Struggled In The Physical World
Operational failures are rare, so labeled failure data is scarce. The events that matter most are often the ones that cause the most damage.
Every machine is effectively bespoke, with sensor types that don’t standardize. Unlike a line of code, physical failure isn’t reversible. Equipment also drifts over time, so last year’s calibrations are no longer accurate.
The industry built what it could at the time: fixed thresholds, beefed-up dashboards and one-off detection models rebuilt by hand with each asset change. But operators were left a step behind.
Like the transformation of digital AI, physical AI works in the real world today because a few things finally caught up: compute became more accessible, models became better at understanding physical data and the architecture transformed to one better equipped for complex processes.
What’s different now is the emergence of new AI technologies that can generalize and adapt across machines, sites and use cases by learning how things move, wear, vibrate and behave, without the need for individual customization. Much like an LLM learns how words flow, sensors can recognize when a machine is drifting toward a failure no one labeled, predicted or wrote a rule for.
By fusing multiple sensor types with video, text and audio, a model can surface anomalies, identify hidden root causes and explain them in natural language.
What This Looks Like On Real Equipment
Speed is half the equation. A hand-built model only finds the failures you design it to look for. In my experience, most of the effort in building a model goes into specifying those failures in advance.
When designed correctly, though, a model already fluent in how physical systems behave can catch a rare, high-cost fault without pre-made examples, verify a task on every pass and transform a video of a correctly completed task into written procedure. The right approach no longer lives just with the most experienced employee.
The value is legibility. For decades, the physical world has run on gut feel, expert knowledge and gauges that tell you something broke after the fact. But a physical world model lets them see what’s happening, ask why and decide before production stops.
It’s Bigger than “Productivity”
A well-run predictive maintenance program can eliminate 70% to 75% of equipment breakdowns before they happen. These numbers turn a “wow” moment into a budget line.
But productivity is too narrow a frame for what’s at stake. Digital work is forgiving. Physical AI output isn’t. A day of production that never ran is gone.
The same applies to expertise. Every plant has an operator who “just knows” based on knowledge earned over 20 years. Today, that knowledge is gone when the operator retires. Making equipment legible with physical AI preserves and distributes that judgment.
What Comes Next
Bringing AI agents into the physical world requires more than powerful models.
Industrial environments are constantly changing, and the intelligence supporting them must adapt as equipment ages, conditions shift and new problems emerge. Reliability, transparency and human oversight will be essential, particularly when the decisions informed by AI affect worker safety or critical operations.
The opportunity isn’t to replace human judgment, but to give people a more complete understanding of the systems they manage. When AI can interpret real-world signals, identify emerging problems and explain what is happening, operators can move from reacting to failures to anticipating them.
The first wave of AI agents changed how we work with digital information. The next will change how we understand and operate the physical world. The real breakthrough won’t be an agent that can do more on a screen. It will be one that helps us see what we couldn’t see before—and act on it before it matters most.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
