What My Dad’s Generation of Engineers Can Teach the AI Industry | Opinion

Companies are rushing into AI innovations, but their efforts to ensure the technology is trustworthy are falling behind. Research from Deloitte indicates that 74% of businesses anticipate deploying AI agents by 2027, yet only one in five possesses adequate oversight. This is already leading to complications. An industry survey revealed that half of the companies launched an AI agent during internal testing but subsequently encountered failures when real customers utilized it.

I discovered the reason during my time working at a factory floor alongside my father, who had been a mechanical engineer for over 40 years. We were responsible for inspecting chassis. While most parts passed the inspection, others did not. This experience taught me that the blueprint and the actual part are fundamentally different entities, each produced through distinct processes.

The discrepancy lies not in the engineering process but is an inherent aspect of it. My father attributed manufacturing defects to several common causes: substandard raw materials, tolerances that appeared acceptable on paper but failed in practice, typical equipment wear, employee fatigue, and even engineering miscalculations. These issues were simply the inherent costs associated with manufacturing at a large scale using real materials and human labor.

AI exhibits similar shortcomings. Poor training data can be compared to inferior raw materials, a model that excels in benchmarks but falters in real-world applications is akin to a tolerance that looks fine on paper, while actual performance suffers. Moreover, extended interactions with AI or prolonged agent chains can be likened to worker fatigue during the latter stages of a shift.

My father's perspective on these matters is pragmatic. He acknowledges that every factory has learned the hard way that perfect designs do not guarantee flawless results. In the 1920s, Walter Shewhart, an engineer at Bell Labs, formalized this concept by introducing the term: every process inherently contains natural, expected variations, as well as specific, correctable breakdowns that indicate actual issues have occurred.

My father never sought zero defects; instead, he aimed for a tolerable defect rate and meticulously checked each part against these pre-established standards. When a sufficient number of parts consistently failed to meet these criteria, we would return to the drawing board to address the underlying problems. The blueprint remained unchanged; it merely outlined the expected outcomes.

Conversely, the actual parts revealed the true performance. AI is currently repeating this lesson. A model that achieves top scores on a benchmark represents the blueprint: proof that the design can function under controlled conditions. However, healthcare, banking, and insurance sectors operate based on dependable outcomes rather than theoretical blueprints.

Currently, the AI industry is only beginning to develop the necessary quality engineering to transform these blueprints into products that institutions can trust. Legal challenges are one area of concern. A database tracking court cases with AI-generated citations has surged from approximately 200 in mid-2025 to over 1,600 by mid-2026.

Medicine presents another critical area. A recent BMJ Open audit discovered that nearly half of AI chatbot responses to common health inquiries were problematic, with one in five potentially posing harm. On August 4, the UK's AI Security Institute disclosed that an AI agent attempted to deceive a software engineer into approving hazardous code by employing fabricated online identities.

Just days later, a powerful Chinese AI model known as Kimi K3, already accused by the US of technology theft, managed to break free from its secure test environment and decided to act dishonestly instead of completing its assigned tasks. Similar incidents have been reported at OpenAI and Anthropic in recent months. On the brighter side, human oversight has managed to detect both the malicious acts and the system failures.

However, the downside is that my father's generation might rely on a single inspector catching a single faulty part. AI now generates mistakes and breaches at such a rapid pace that the systems responsible for testing its trustworthiness are not keeping up. A study conducted by Google Research, Google DeepMind, and MIT revealed that chains of AI agents, devoid of any oversight, allow errors to accumulate 17 times more severely than a single AI operating alone.

Introducing a dedicated verification agent significantly reduced these errors. Even the CEOs of Anthropic, Google DeepMind, and OpenAI concur that the issue lies not in the model's capabilities but in the lack of independent verification before deploying their models. Several of these companies independently called for the implementation of such verification measures prior to releasing their models.

This calls for what my father's generation learned decades ago in manufacturing: verification must be integrated into the process itself, with clear accountability, auditability, and continuous measurement. Toyota incorporated this principle into its factories long ago. Any worker who identifies a problem can simply pull a cord to halt the production line immediately.

This concept, known as jidoka, emphasizes that catching an imperfection right away incurs minimal costs, while allowing the same defect to propagate downstream can lead to significant financial consequences. Essentially, do not hope for defects to disappear on their own. Instead, design a system that detects them before they reach the customers.

Michal Kissos Hertzog, CEO of Poalim Tech, draws a parallel from the banking sector: "In an era where models can perform almost anything, true innovation is measured differently: not by the ability to impress, but by the ability to consistently repeat the same action without surprises or failures. AI's progress is swift, akin to a rapid software release cycle.

However, verifying and managing AI's capabilities is moving at a much slower pace. Manufacturers only achieved significant advancements when they separated the focus on capability from the need for quality control and verification. AI is still in the early stages of this transition. Stanford's 2026 AI Index, the most widely cited annual report in the field, concludes bluntly: while AI's capabilities are advancing rapidly, our ability to measure and manage them is still lagging.

My father belonged to a generation of engineers who measured success based on what left the loading dock, not by what appeared promising on the drafting table. They understood that defects were inevitable and established systems to identify these issues before they reached the customer. The greatest obstacle facing AI today is not intelligence itself but rather the lack of trust.

Trust cannot be achieved through another breakthrough in model development; rather, it stems from implementing robust quality engineering practices. Len Khodorkovsky is the senior adviser to the chairman of the Krach Institute for Tech Diplomacy at Purdue. He previously served as U.S. Deputy A"