AI agents more effective than GenAI for enterprise productivity: Deloitte Study 2024

NEW DELHI: Artificial intelligence (AI) agents can be a more effective tool than large language models (LLMs) or GenAI applications, a study by British professional services firm Deloitte found. This opens up new possibilities for driving enterprise productivity and program delivery through business process automation.
The study said that cases that were previously deemed too complex for GenAI can now be enabled at scale securely and effectively with the aid of an AI agent.
By definition, the AI agent is an autonomous intelligent system that uses AI techniques to interact with its environment, collect data, and perform tasks without human intervention.

Explaining the difference between Gen AI and AI agents, the study adds that typical LLM-powered chatbots usually cannot understand multistep prompts.

“They (LLM or Gen AI) conform to the “input-output” paradigm of traditional applications.
and can get confused when presented with a request that must be deconstructed into multiple smaller tasks. They also struggle to reason over sequences, such as compositional tasks that require consideration of temporal and textual contexts. These limitations are even more pronounced when using small language models (SLMs), which, because they are trained on smaller volumes of data, typically sacrifice depth of knowledge and/or quality of outputs in favor of improved computational cost and speed,” it said.
The study says that GenAI use cases have mostly been limited to standalone applications such as generating personalized ads based on a customer’s search history and reviewing contracts, among others.
On the other hand, AI agents excel in addressing the limitations while also leveraging the capabilities of domain- and task-specific digital tools to complete more complicated tasks effectively.

“AI agents equipped with long-term memory can remember customer and constituent interactions–

including emails, chat sessions, and phone calls–across digital channels, continuously learning and adjusting personalized recommendations. This contrasts with typical LLMs and SLMs, which are often limited to session-specific information,” the study adds.
The study further adds that while individual AI agents can offer valuable enhancements, businesses also need multi-agent AI systems, given the limitations of single AI agents.
However, the study notes that AI agents also introduce new risks that necessitate robust security and governance structures.
“A significant risk is a potential bias in AI. algorithms and training data, which can lead to inequitable decisions. Additionally, AI agents can be vulnerable to data breaches and cyberattacks, compromising sensitive information and data integrity,” it adds.