AI for the Physical World: Modeling, Measuring, and Governing What We Make
An interview with Matt Kaiser, Director of Product at UL Solutions
For more than a century, UL Solutions has been synonymous with product safety, testing, and compliance. From the wiring in your walls to the plastics in your kitchen, UL’s mark is shorthand for “this has been tested.”
But the nature of “testing” is changing. Evolving technology has consistently enhanced testing capabilities but arguably never as radically as is happening now with AI. AI can simulate the world, digitally modeling physical systems to predict their real-world behavior. From chemical safety to counterfeit detection, product compliance is becoming a data challenge as much as a materials challenge.
I had a conversation with Matt Kaiser, Director of Product at UL Solutions, to discuss how AI, simulation, and new data modalities are reshaping the field of product stewardship, the discipline that ensures products are safe, sustainable, and compliant with the world’s increasingly complex regulations and market access requirements.
When Compliance Becomes Design
Compliance has evolved beyond simple checklist completion. It now means intelligently integrating market requirements into the earliest stages of design.
“The product stewardship industry is under pressure from multiple directions,” Matt said. “AI and simulation technologies are helping companies align regulatory complexity and sustainability outcomes with their business operations.”
At the same time, the regulatory landscape is exploding. Rules are shifting faster than ever, and the line between regulation and sustainability is blurring. Companies are outsourcing compliance and seeking embedded orchestration layers that can connect directly into their design, sourcing, and manufacturing systems.
Modeling the Physical World
The promise of AI for the physical world starts with simulation. What began as static digital twins (virtual replicas of products or systems) is now evolving into collaborative simulation, where regulators, manufacturers, and auditors can model product performance together in shared digital environments.
“First you need structured data,” Matt said. “Then web-based collaboration spaces that are personalized by role. Simulation could be as simple as swapping one raw material for another and getting a green checkmark if a requirement is met.”
With digital twins (and their emerging “digital cousins”), upgraded physics engines, and AI reasoning already in place, simulation is poised to become the new test lab.
By centralizing product and chemical data, companies can now use AI to augment testing, identify failure risks earlier, and even replace certain physical tests entirely. “Thermal aging, for example, can take months,” Matt noted. “If you can train models to recognize early indicators of failure, you can dramatically shorten that cycle.”
UL and Siemens made history last year when UL issued a certification based solely on a digital twin of an industrial product, a milestone that hints at the future of testing.
Still, scaling that approach across materials and product types remains challenging. “The physical world isn’t a closed system like software,” Matt explained. “It deals with uncertainty – tolerances, environmental changes, chaotic behaviors. One of software’s biggest roles is to bring predictability to that uncertainty.”
PFAS and Chemical Regulation
Nowhere is this uncertainty more visible than in chemical regulation, particularly around PFAS – a large class of human-made chemicals used across industrial and consumer products. Often called “forever chemicals,” they don’t break down in the environment and can accumulate in people and animals over time.
“PFAS is a substance-level problem, which makes it systemic,” Matt said. “It’s in thousands of products because of its performance characteristics. The same traits that made it useful also make it difficult to regulate.”
It’s one thing for a regulatory body to agree to regulate a substance; it’s another thing entirely to do so in a coordinated way. Local, federal, and international rules often conflict – and NGOs and lobbyists add further complexity to the rulemaking process, even on topics where coordination should be straightforward. This ever-changing, patchwork system often gives rise to emergent, unpredictable behavior that leaves companies scrambling to address compliance retroactively, putting them at huge risk of causing issues for human health and the environment.
This is especially problematic for small and medium-sized product developers who face limited capital, lack of in-house expertise, and risks with adopting and scaling new advanced tech in preexisting, nonnative processes. As a result, Subject Matter Experts (SMEs) may fear disrupting existing processes and lack specialists to pilot new systems.
AI can help companies track evolving rules, extract relevant information, and map insights against product formulations automatically. As Matt explained, “This is one of the areas where I see AI helping the most; understanding and maintaining a unified rule set for regulated substances, even when that rule set will need to be dynamic as regulations change over time.”
Visual Data and Smart Labeling
Much of compliance still runs on unstructured data from static documents and images: labels, safety markings, inspection photos, certification documents. Extracting meaning from the mess is a perfect use case for AI.
“This problem is harder than people think,” Matt cautioned. “Even standardized documents like an SDS (Safety Data Sheet) vary wildly. You need near-perfect accuracy because that data affects regulatory outcomes downstream.”
Smart labeling and visual data retrieval are early steps toward multimodal compliance systems, where text, images, and sensor data combine to monitor performance and maintain certifications in real time.
“Figuring this out will be a game changer,” Matt said. “Not just for initial testing and certification (speeding up tests or allowing companies to circumvent certain testing altogether through simulation), but also for maintaining certifications.” Testing and certification is not a one-and-done activity, but access to production data will help to cut down on the need for recurring field visits and ongoing sample testing.
Fighting Counterfeits and Building Trust
At Chicago’s mHUB innovation center, Matt has seen startups using computer vision to detect physical anomalies, tiny manufacturing differences that can help drive all kinds of product stewardship outcomes, including helping to reveal counterfeit products.
“Imagine verifying a product’s digital fingerprint with physical material samples from a production line,” he said. “That kind of traceability would transform supply chain trust and quality assurance.”
AI can also help customs agents and auditors identify forged documentation, a major issue in cross-border trade. Identifying issues with regulatory documentation, especially what’s required for moving product through customs, would benefit from the same applications that help manufacturing companies manage their regulatory documents. If a customs agent can easily call a document management system to retrieve or compare a source document with the document they have on hand, it’s a lot easier to determine whether that document was forged or is out of compliance.
But the key, Matt emphasized, is maintaining a human-in-the-loop: “Trust systems need feedback. There’s no such thing as a perfect AI model. People still need to guide and validate.”
The Governance Layer: Toward a “Compliance Copilot”
Regulation is the connective tissue of product stewardship, and one of its biggest bottlenecks.
Products can be regulated in numerous ways. They may be subjected to substance or material regulations, standards that govern best practices for components or systems; they may also be in scope for extended producer responsibility policies or rules governing sustainable product development. Many of these rules are jurisdiction or geography-specific. Matt tells me “AI can help extract and contextualize that information, but interpretation still requires human judgment.”
When I asked him what an AI solution might look like, he highlighted simplicity and usability (something I wrote about a couple weeks ago), saying “I don’t think this should be too complicated from a look and feel perspective. A simple chat-style interface would do, with predefined templates to help visualize the data.”
Underpinning that, the copilot would need to:
Access structured product and supply chain data to compare against regulations.
Validate the accuracy and completeness of product data, filling gaps where needed.
Cite and verify sources for regulatory information.
Maintain audit trails and version control to ensure traceability.
“Training AI on deep, complex subject matter is resource-intensive and prone to hallucinations or model drift,” Matt said. “The goal isn’t to automate judgment, it’s to offload the tedious workflow tasks so experts can focus on advancing the science.”
Agentic systems, for instance, could automatically flag out-of-compliance formulations, draft regulatory responses, or generate summaries for human review. The expert remains in control; the system simply reduces the cognitive and administrative load.
“The software won’t exist unless SMEs help develop it,” Matt said. “We need them to build the systems that will ultimately improve outcomes.”
The Road Ahead
The world’s physical systems – sitting at the intersection of materials, regulation, and human safety – are entering a new era. They will increasingly be designed, tested, and governed by AI.
According to Matt, the biggest near-term opportunities are twofold:
Driving efficiencies through workflow automation and deeper collaboration.
Mining data for context and insight to improve product lifecycle outcomes across the supply chain.
AI will soon be embedded in every stage of the product stewardship process, from data aggregation and verification to impact reporting, trend analysis, and design optimization. In turn, this will enable feedback loops that continuously improve efficiency, accuracy, and compliance.
When I asked Matt what “AI for the physical world” might look like in five years, his answer was both pragmatic and visionary:
“Accessible product lifecycle data means SMEs can be more creative in driving product stewardship outcomes, helping to simulate and meet end-market requirements early in the design phase to improve formulation, engineering, storage, waste, and reuse outcomes. This will ultimately help companies reduce risk while improving efficiency, environmental outcomes, and human health.”
In that future, AI will do more than accelerate testing or simplify compliance; it will contribute to making the physical world more intelligent, resilient, and sustainable.