The transition of artificial intelligence from digital code to physical presence marks a pivotal shift in how we interact with technology. As autonomous systems leave the confines of software and enter the tactile world of warehouses, delivery routes, and city streets, the stakes for safety and governance rise exponentially. Laurent Giraid, a leading technologist specializing in machine learning and the ethics of embodied AI, joins us to discuss this evolution. His work focuses on the bridge between agentic autonomy and real-world infrastructure, examining how global frameworks are adapting to a reality where a software glitch could have physical consequences.
The conversation explores the amplification of risks when AI controls critical infrastructure, such as smart grids or transport networks, and the emerging governance models designed to mitigate these dangers. We delve into the concept of agentic AI—systems capable of multi-step planning and tool interaction—and how institutions like Singapore’s IMDA are setting new standards for accountability. The dialogue also covers the industrial scaling of robotics in response to labor shortages in Japan and China, the transformation of retail through “super agents,” and the shifting workforce landscape in the global banking sector.
When AI systems transition from digital environments to physical infrastructure like smart grids or delivery networks, how does the nature of risk change for the organizations managing them?
The shift from the digital domain to the physical one represents a fundamental change in the consequences of failure. In a software environment, an error might lead to a data leak or a biased output, but once AI is embedded in a vehicle or a power grid, that same error can result in property damage or even a threat to human life. We are seeing a shift in focus toward operational safety, similar to what you would find in aviation or heavy industry, rather than traditional software regulation. For example, Dr. Ya-Qin Zhang has pointed out that any risk present in the digital world is essentially amplified once it touches physical systems like drones or logistics networks. Organizations now have to worry about a “long tail” of unpredictable real-world issues that simply don’t exist in a controlled software testbed.
With the rise of agentic AI that can plan and execute complex tasks across multiple systems, what new layers of control are necessary to prevent these agents from acting outside their intended scope?
The complexity of agentic AI requires a move away from one-time certifications toward a more dynamic, iterative governance process. Singapore’s latest framework for agentic AI highlights the need for technical controls like “least-privilege” permissions, where an agent only has access to the specific tools it needs for a task. We have to be very deliberate about human-in-the-loop requirements, especially for high-stakes or irreversible actions. It is no longer practical for a human to review every single workflow at scale, so we have to identify significant checkpoints for approval. This approach helps mitigate “automation bias,” where human supervisors might become too reliant on the agent’s capability and miss critical errors.
Given the inherent unpredictability of the real world, how are companies effectively validating the safety of autonomous robots before they are deployed in public spaces?
The gold standard for validation right now is a heavy reliance on simulation and telemetry before any large-scale rollout occurs. Grab’s work in the Punggol district is a great example of this, where they use closed and open courses to stress-test their delivery robots. They don’t just jump to deploying hundreds of units; they “crack it” in a simulated environment first to ensure the logic is sound. Once the robots are on the street, continuous monitoring and the ability to take a system offline if it malfunctions become the primary safeguards. It’s a process of iterative testing where you are constantly looking for unexpected behaviors that only emerge in a messy, real-world environment.
In the financial sector, we see institutions like JPMorgan and OCBC integrating AI into complex workflows; how is this technology reshaping the roles of human professionals in those environments?
We are witnessing a structural shift where the workforce is being reshaped to prioritize AI expertise over traditional roles. For instance, the leadership at JPMorgan has explicitly stated that they expect to hire more AI specialists and fewer traditional bankers in the coming years. In specific applications like the source-of-wealth analysis used by OCBC, the AI doesn’t make the final decision but instead parses documents to draft memos for human review. This keeps the human professional as the final validator, especially in high-risk areas like credit or onboarding decisions. It transforms the job from manual data synthesis to a more strategic oversight role, though it also requires employees to develop new skills to effectively assess AI-generated outputs.
The industrial sector in Japan and China seems to be moving at a different pace regarding robotics; what is driving the large-scale adoption of AI-powered robots in these regions?
The motivation is largely driven by necessity, particularly chronic labor shortages and the desire for industrial commercialization. In Japan, a recent survey of nearly 500 companies showed that about one-third are already using or considering AI-powered robots, with the transportation equipment sector being the most proactive. There is a massive effort to collect 100,000 hours of robotics data to support foundational models that can handle various industrial tasks. Meanwhile, in China, startups like Galbot are focusing on semi-structured environments like warehouses and autonomous retail stores because these settings are more controllable than the open street. These government-backed testbeds and industrial partnerships are creating a path for robotics to become a standard part of the commercial landscape.
When an autonomous system involves developers, manufacturers, and infrastructure operators, how do we determine who is responsible when a failure occurs?
Accountability is one of the most difficult hurdles because the value chain is so fragmented. You have the AI model developer, the robotics manufacturer, the semiconductor supplier, and the operator who actually deploys the system. The consensus emerging from frameworks like the one in Singapore is that the organization deploying the agent remains ultimately accountable for its actions. However, there is a push for clear responsibility across the entire chain, ensuring that tool providers and platform developers are also held to specific standards. This becomes even more complex when systems continue to learn and adapt after they are deployed, making it hard to point to a single moment of failure in the original code.
Walmart is moving toward a future defined by “super agents” for shoppers and employees; what does this mean for the everyday consumer experience?
Retail is becoming much more proactive and personalized through the use of agentic tools. Walmart’s “Sparky” agent is already evolving from a simple shopping assistant to a tool that can plan entire events or suggest recipes based on what it sees in your fridge via computer vision. They are building a whole ecosystem of agents, including those for store associates and suppliers, to streamline every interaction point. The goal is to make the AI the primary entry point for all groups, effectively removing the friction of traditional search and logistics. While there is a lot of talk about how these tools will change job roles, the immediate impact for the consumer is a much more integrated and automated shopping journey.
What is your forecast for the future of embodied AI in public infrastructure?
I believe we are moving toward a “purpose-built” era where AI is not a generic solution but is specifically adapted to the industrial ecosystems it inhabits. We will see a massive increase in the use of specialized sensors and more energy-efficient computing architectures to support these physical agents. The next five years will likely be defined by the development of global safety standards, similar to the Hiroshima AI Process, which will allow different countries to share data and governance models. Ultimately, the success of embodied AI will depend on our ability to create a “digital twin” of our physical world that is accurate enough to predict failures before they happen in real time.
