“Human in the loop” is probably one of the most oft-repeated phrases in today’s business technology lexicon. It seems like the most sensible of common sense: artificial intelligence or AI agents cannot run fully unattended, we need to have a human reviewing the output or results.
However, a prominent AI proponent from MIT says it may be time to think about taking humans out of some loops when it comes to AI applications.
Not every loop needs a human, writes Paul Cheek, senior lecturer at the MIT Sloan School of Management, co-founder of the AI-Driven Enterprise Institute, and founder of Entonomy.
In his new book, No One Works Here, he urges managers and business leaders to think twice before throwing a roadblock into the flow of what could be very efficient automation. ‘The companies that are most likely to fall behind their peers are the ones that insist on keeping a human in the loop for decisions that a machine could have made yesterday,” he states.
Thus, we reach a fine line when it comes to deciding where a human in the loop is needed, and where it may be an impediment. Humans in the loop serve as the necessary “default” for the AI world, especially in financial services, according to Ted Paris, head of analytics, intelligence and AI for TD Bank US, in recent CIO article.
AI already has major trust issues, said Paris. “The question is where can AI create meaningful value while keeping the right human judgment, oversight and accountability in place? In financial services, that distinction matters. In an industry built on trust, those capabilities are not optional – they are foundational to how we serve customers, manage risk and earn confidence every day.”
Still, there are areas where maybe a human in the loop will make AI worse, not better, Cheek argues. For example, “a modern motor vehicle equipped with an intelligent emergency braking system reacts before a human driver even perceives a hazard or has the time to react. In this moment, the human’s biological latency is a liability; the machine’s autonomous agency is the lifesaver.”
The delay inherent in human decision-making can have its costs from a corporate or economic perspective as well. “Pricing adjustments, supply chain routing, and customer support queries” all may function quickly and efficiently in a fully automated fashion. “In these domains, the hesitation of the human operator – the need to schedule a meeting or build consensus – is not a safety feature; it is a bottleneck,” he states.
Many AI applications and agents start out with stringent human oversight and audits before they are unleashed on their own, Cheek adds. “When you first deploy an agent, latency often spikes. You will spend more time auditing the agent’s output than you would have spent doing the work yourself. This is the latency J-curve: Things get slower and riskier before they become instant.”
Once all the potential bugs and issues have been worked out, an AI may achieve “trust-based autonomy” in which humans can step aside.
AI efforts tend to be hampered by what Cheek calls “structural debt,” in which organizations have historically been designed “to solve for the limitations of human communication. We created hierarchies because one person couldn’t manage a thousand people. We created departments to specialize knowledge because one brain couldn’t know everything. We created middle management to route information because we didn’t have networks.”
As a result, data for decision-making “has to travel up the chain of command,” he observes. “It gets aggregated, summarized, PowerPointed, and debated. By the time the decision is made and the communication comes back down, the market has moved.”
Granting greater autonomy to AI across key processes may help organizations move faster, and avoid the paralysis of multi-level decision-making, he urges. Still, he adds, make sure there are guardrails and governance policies in place at all times.
