The Secret Sauce for AI Success – “Human In The Loop” Design

I have had so many great conversations this week ithat t was a struggle to decide on which to expound on.

• The causes of “Failure Demand”
• Digital transformation “quicksand”
• The “sugar-rush” of silo thinking
• “Data lies when processes are broken” but it’s still valuable . . .
• Agentic AI and the “end of the world as we know it” . . . .

. . . .  and so many more,

But an article in Forbes by Carter Busse, really sparked me with its well thought out challenge on process design in an “AI First” world . .

Carter’s article “The Secret To Successful Enterprise AI? ‘Human-In-The-Loop’ Design” observes some high-profile examples of AI-enabled processes that could have benefited from a “human in the loop” (HITL);

• Air Canada’s chatbot mistakenly providing undesirable discounts (maybe desirable for the customer 😉)
• McDonald’s drive-thru AI serving up bacon on ice cream and placing hundred-dollar McNugget orders

The author argues that the way to accelerate AI and LLM progress is to “design-in” human involvement at critical decision points in the process.

This is the secret sauce for AI Success – HITL design.

Recent research by MIT and Accenture shows that, when nudged to review LLM-generated outputs, humans are more likely to discover and fix errors.

There is the corresponding challenge that we can be encouraged to laziness in the assumption that the LLM will generate the right answers. But that’s on us. Human behaviour remains one of the biggest challenges in exercising intelligence, be it human or artificial.

The research addressed this risk.

“Those who were given no details about the accuracy or sourcing of the information—the typical experience for most commercial LLM users today—were more likely to overlook mistakes or inaccuracies compared to those provided with error labeling. The more information provided about the origins and accuracy of the results, the better the users are at detecting problems.”

Human-in-the-loop design is powerful and already common in many processes. Increasingly, companies are using AI to generate results quickly, but have several rounds of reviews to identify any errors in critical outputs.

The author of the research, Dr. Renée Richardson Gosline, speaks of the value of adding speed bumps between AI output and delivery, what she refers to as “beneficial friction.”

Carter describes an example in his own organization, an orchestration between the Salesforce, Slack, Outreach and an LLM (GPT4), where the final step is not automatically deployed without a decision based on valuable context gained through years of experience and human relationships.

Done well, these human decision feedback loops will make LLMs (and SLMs, but that’s a topic for another day . . .) smarter, leading to more intelligent processes over time.

To get to that level of AI maturity, companies must start thinking about humans as a key ingredient in orchestrating their AI-enabled processes. Then, they can weave together people, systems, processes and AI to enhance productivity and efficiency.

It all starts with human-in-the-loop (HITL) design today.

You can read Carter’s article, “The Secret To Successful Enterprise AI? ‘Human-In-The-Loop’ Design” here . . . 

You can read the the research “Nudge Users to Catch Generative AI Errors” in the Sloan Management Review here . . . 

The Secret Sauce for AI Success – “Human In The Loop” Design.

Thanks for reading . . .