true77
As we entered 2025, OpenAI CEO Sam Altman was hyping up two game-changing developments that he believed would revolutionize our lives. One was the highly anticipated GPT-5, a major upgrade to the Large Language Model (LLM) that propelled ChatGPT to tech stardom. The other was AI Agents that not only answer queries like ChatGPT but also execute tasks for you. Altman predicted that we might witness AI agents entering the workforce in 2025 and significantly impacting company output.
Fast forward eight months, and Altman’s prediction already seems to be falling short. While many companies are eager to embrace AI Agents, the feedback on their performance has been overwhelmingly negative. Reviews describe them as glitchy, inconsistent, and failing to live up to the hype.
A study conducted by Carnegie Mellon University in May 2025 found that even the best-performing AI agent failed at real-world office tasks 70% of the time. OpenAI’s GPT 4.0-powered entry fared even worse, failing over 90% of the time. The upcoming GPT-5 may improve these numbers slightly, but the fundamental challenges remain.
Researchers point out that errors tend to accumulate over time in AI Agents, leading to performance degradation and potentially catastrophic mistakes. As a result, some AI Agents have made critical errors in judgment, such as unintentionally deleting a customer’s database after working on a task for just 9 days.
Despite the anticipation surrounding GPT-5, there are concerns that regulatory constraints and company policies may limit the capabilities of AI Agents. For instance, Amazon has restricted AI Agents from browsing and purchasing across its platform, citing control over the customer experience and advertising interests. Additionally, vulnerabilities in AI Agents, such as susceptibility to phishing attacks and data manipulation, pose significant risks that could compromise user data security.
In light of these challenges, a Gartner paper predicts that 40% of AI Agent initiatives by companies may be canceled within two years due to inflated expectations and deployment complexities. While advancements like GPT-5 may enhance the reliability of AI Agents, it remains to be seen if they can overcome the inherent limitations and vulnerabilities that currently plague the Agentic AI landscape.
