Copilots have been all the rage. There is something perhaps reassuring about the concept in the brave new world of AI. These AI systems work alongside humans to ensure accuracy, finding productive applications in software development, product design, sales lead generation, and content creation, among many others. However, our current framework for determining when humans should step aside may reflect outdated thinking. This risks missing groundbreaking AI use cases, especially in the most complex and sensitive foundational industries.
Common wisdom suggests that human-in-the-loop copilots are best for tasks requiring significant human judgment and where errors have substantial consequences. Meanwhile, full automation —or what I will refer to as autopilot — is deemed suitable for repetitive and well-defined tasks. However, this conventional view may prevent us from utilizing AI's full potential, particularly in foundational industries.
An alternative framework, credited to Sergiy Nesterenko, CEO of Quilter AI, was recently discussed on the Terpentine podcast. This framework resonated with me because it focuses on what we can do, rather than what we cannot.
The team at Quilter AI is developing an AI system for printed circuit board design, with customers including Tesla, Meta, and Anduril. Sergiy emphasizes that Quilter’s AI does not merely assist electrical engineers; it fully automates the design process. In this context, the AI acts as an autopilot, not a copilot. When dealing with the intricate web of resistors, capacitors, inductors, and diodes, if a computer handles 90% of the job, the remaining 10% becomes even more complex for a human to complete. Imagine threading the last few components in a design containing hundreds or thousands. This represents consequential thinking around the critical distinction between copilots and autopilots.
From this example, an obvious yet contrarian framework emerges: If AI supplementation makes task completion more difficult — often due to the task's complexity — full automation or autopilot should be pursued. Many of these tasks involve high judgment importance and significant error consequences and might not be repetitive or well-defined. Common wisdom would suggest these are opportunities for human-in-the-loop copilots, but this framework advocates for full automation as the only viable approach.
Consider the task of routing a package within the global supply chain. If AI gets a supply chain analyst 90% of the way there, the remaining decisions without context become challenging. Similarly, if a model analyzes more driving scenarios than a human could experience in thousands of lifetimes, it will not prevent many accidents by merely making suggestions to a human driver. In both cases, despite the complexity and need for judgment in an undefined problem space, the human should step aside in favor of autopilot, even if copilot feels reassuring.
When AI supplementation complicates tasks completion, full automation is more effective, even in complex, high-stakes scenario. Today’s common wisdom should not steer technologists off from pursuing the ambitious that is complex and consequential. instead technologies for monitoring and validating systems should empower AI systems to solve problems for us, as autopilots.