Engineering AI Will Not Follow the Software Playbook

Agentic AI Hits a Harder Wall When the Environment Is Physical

The agentic AI conversation has been shaped almost entirely by software use cases: coding assistants, customer service agents, enterprise knowledge workflows. In those environments, failure modes are easier to observe, accountability structures are largely digital, and many errors can be contained within the workflow. In engineering-led sectors, the calculus is categorically different.

“In our domain, agentic AI cannot remain confined to workflows,” Chadha said. “It must operate in environments where decisions have physical consequences.”

For LTTS, real agentic capability means systems that can sense, decide, and act autonomously within defined boundaries in a manufacturing plant, a vehicle, or an energy system. Clients are comfortable with AI assisting decisions and running simulations. What they are not yet ready to hand over is full autonomy in safety-critical environments, where a wrong call can affect production continuity, product integrity, or human safety.

The path Chadha describes is deliberate and staged: AI recommends, then assists, then executes within guardrails, then operates autonomously in controlled domains. The final stage requires not just better models but engineering frameworks, validation systems, and accountability mechanisms. “Agentic AI in engineering will scale differently, and perhaps more cautiously, than in software,” he said. That caution is not merely defensive. It reflects the design constraints of engineering environments.

The bottleneck is not on the supply side. LTTS has built and deployed sophisticated AI systems, Chadha said. The harder problem is client absorption. Across automotive, industrial manufacturing, energy, and healthcare, AI maturity varies sharply. Some companies are building AI-native platforms from the ground up. Others are managing decades of legacy infrastructure and risk frameworks that were not designed for autonomous systems. That gap between what is technically possible and what organizations are operationally ready to deploy is, Chadha says, the defining constraint of this moment.

“The bottleneck is less about what we can build and more about how quickly industries can absorb and scale it,” he said.

Accountability Cannot Be Retrofitted After Deployment

When AI is embedded in physical systems, the question of who is responsible when something goes wrong becomes structurally more complex than in software. Chadha describes accountability as layered: at the system level, responsibility remains with the enterprise that owns the asset or operation; at the solution level, it is distributed across the engineering partner, the technology stack, and the operating framework.

That distribution has a direct implication for system design. Reliability, validation, and lifecycle governance cannot be added after deployment. They must be architected in from the beginning. That means the engineering partner’s role is not simply to deploy capability, but to embed accountability into the system itself.

“In practice, autonomy is always bounded,” Chadha said. “Systems operate with human-in-the-loop or human-on-the-loop models depending on criticality.”

Healthcare makes the principle concrete. LTTS’s lung navigation platform sits at the intersection of engineering precision, clinical workflow, and AI-enabled simulation. In that environment, Chadha says, AI can simulate scenarios, assist navigation, and provide decision support, but critical clinical decisions must remain with the clinician. Not because the model cannot perform, but because accountability cannot be algorithmically transferred. The same logic applies, with varying degrees of strictness, across every engineering domain where physical consequences are real.

The competitive implication is worth stating directly. Hyperscalers bring platform reach. Pure-play software firms bring model capability. What neither can easily replicate, Chadha argues, is deep multi-domain engineering expertise combined with the ability to embed AI into physical systems where outcomes are tied to real-world performance, safety, and lifecycle reliability. That convergence, not model sophistication alone, is where industrial AI will be differentiated.