IBM Patent Leverages AI to Fix AI

What happens when agents go awry? 

The results can range from frustrating to disastrous, and IBM is developing a potential remedy: a system for “coordinating a conversational agent” with a large language model for “conversation repair.” When an agent fails to understand a user query, IBM’s tech would essentially refer back to a more sophisticated AI model to set it on the right track. 

When a failure is detected, IBM’s system pulls a descriptive prompt that explains the chatbot’s purpose and capabilities. That information, plus the context of the failed conversation, are sent to a more capable large language model, which determines the user’s intent as well as where the original agent went wrong. 

If possible, the language model hands the task back to the agent, this time with a clarified user intent. If the task is outside of the agent’s capabilities, the language model handles it. 

“Traditional (conversational agents) are limited to a specific set of topics and are not able to respond to input outside those topics,” IBM said in the filing. “They can only produce responses that are pre-written or pre-approved by humans, i.e., they are non-generative.”

While IBM’s tech uses AI to fix AI, generative models don’t always get things right. The domain knowledge of large language models tends to be far more expansive than that of a simple question-answer chatbot. But when conversation expands outside of the realm of its training data, these models can hallucinate, potentially leading to broader issues than a closed-domain chatbot running out of things to say.

Plus, while AI agents are the current hot topic of the tech world, as far as customer service goes, many consumers prefer the real thing: A Gartner study found that 64% of customers reported they’d prefer that companies not use AI for customer service. About 53% reported that they’d consider switching to a competitor because of it.

The bottom line: Before using AI to make chatbots better, remember the tech is far from a panacea. Enterprises should consider where it’s best to keep a human in the loop.