When AI Moves Physical Machinery, the Stakes Change
A router in Arkansas helped ground US flights; Cruise paid $2 million in penalties. As AI moves into machines, insurers and boards face unsettled questions.
By Amara Osei
4 min read
What's News
- A Nov. 13 FAA telecom disruption, traced to a new router in Arkansas and a Texas radar site, grounded departures nationwide.
- Cruise paid a $1.5 million NHTSA penalty and a $500,000 criminal fine after a robotaxi dragged a pedestrian about 20 feet in October 2023.
- The FAA had expanded a new AI system for flight planning nationwide three days before the ground stop; officials ruled it out as a cause.
- A single physical AI failure can fall between cyber and general liability coverage, with no settled answer on which responds.
On Nov. 13, routine telecom work helped bring American air travel to a standstill — and it showed what happens when digital systems reach into the physical world. Preliminary findings point to two FAA sites: a remote radio site in Arkansas, where a contractor had deployed a new router, and a radar site in Texas. A flood of connection attempts destabilized legacy equipment, and the FAA grounded departures nationwide. Every aircraft landed safely.
The root cause remains under investigation. Microsoft and Workday suffered major outages the same week. Officials at the FAA and Microsoft said artificial intelligence played no role, and nothing reported links Workday's outage to AI. Notably, the FAA had expanded a new AI system for flight planning and traffic management nationwide just three days before the ground stop; officials ruled it out as a cause.
Together, these incidents show how routine maintenance, upgrades and testing can ripple through interconnected systems in ways operators never anticipated. Now consider what happens when the technology being introduced can make its own decisions, initiate its own actions and move physical machinery. That is the emerging world of physical AI.
Where does AI cause real damage?
For most people, AI still lives on a screen. It drafts documents, answers questions and recommends decisions. When it fails, the consequences are usually financial, contractual or reputational.
That is changing quickly. Autonomous vehicles navigate public roads. Robots handle materials, operate manufacturing equipment and share warehouses with people. Intelligent systems regulate buildings' electrical and mechanical infrastructure. Agricultural machinery works without continuous human direction. Each combines AI with sensors, mechanical equipment and physical environments often designed long before autonomous decision-making was practical. A physical AI failure can damage property, halt production or injure people in seconds.
The October 2023 Cruise robotaxi incident in San Francisco shows what these failures look like. A pedestrian struck by a hit-and-run driver was thrown into the robotaxi's path. The vehicle braked and stopped, then attempted to pull over, dragging her about 20 feet because its systems did not detect that she was underneath. The failure did not sit neatly inside the AI — it arose where the software, its sensors, another driver's actions and an unanticipated situation met.
What followed was just as instructive. Cruise's initial reports to federal regulators omitted the dragging, which the National Highway Traffic Safety Administration identified after reviewing video it had requested. The company agreed to pay a $1.5 million NHTSA penalty and later a $500,000 criminal fine under a deferred prosecution agreement in which it admitted submitting a false report to influence a federal investigation. In physical AI, what the machine recorded, and how the company handled that record, can matter as much as what the machine did.
Why does one incident need several investigations?
A significant physical AI incident cannot be understood through a single discipline. Forensic engineers determine whether sensors, mechanical components, software, electrical systems or safety controls contributed. Human factors specialists examine how people interacted with the autonomous equipment, whether warnings were adequate and whether there was a reasonable opportunity to intervene.
Digital investigators reconstruct what the machine perceived, what it decided and what commands it issued. That depends on preserving sensor readings, model versions, operating logs and telemetry before they are overwritten, sometimes within hours. Financial experts then measure property damage, business interruption, lost profits, remediation costs and lasting impairment of valuable business assets.
For insurers, the coverage questions may be the least settled of all. When a compromised control system causes bodily injury or property damage, the loss can fall between cyber and general liability programs. An autonomous machine also raises questions conventional adjusting protocols do not answer: Can it be repaired? Must the model be retrained, at what cost and by whom? If the model itself was corrupted, is that a physical loss or a cyber loss?
What should boards ask before writing checks?
A manufacturer might justify a multimillion-dollar investment in autonomous equipment on labor savings and higher throughput. But a single preventable integration failure that causes injury or an extended shutdown could erase years of those savings. Even with insurance, the company may still absorb deductibles, uncovered losses, litigation, regulatory scrutiny and damaged customer relationships.
Before approving capital for physical AI, management and boards should require answers to three questions:
- If something goes wrong, can we reconstruct what happened, establish responsibility and measure the loss?
- Are our existing infrastructure, people and safety controls prepared for what these systems can do?
- Does our business case account for the cost of a failure, and not only the savings when everything works?
Brokers and underwriters should ask a parallel set before binding coverage: How was the system validated, and within what operating limits? Who can override it? Are incident logs retained? How are model updates and equipment changes controlled?
The business case for physical AI cannot end with whether the technology works, the piece argues. It must also show the organization can operate it safely, contain failures, preserve evidence and respond when something goes wrong. The organizations that answer those questions honestly will be the ones that capture the value of physical AI rather than litigate it.
Source: insurancejournal.com
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,175 articles