A recent experiment with AI workflows has highlighted a puzzling trend: while AI excels at execution, it often falters when it comes to making small, yet critical decisions. Tasks such as writing content, summarizing information, and handling multi-step workflows are tackled with ease, but the failures that occur are not due to a lack of capability. Instead, they are attributed to judgment calls that humans make instinctively, such as choosing the right context, accounting for edge cases, and knowing when to stop and ask for clarification.

A notable example of this phenomenon was observed when attempting to automate a basic lead qualification and outreach flow using AI. The system performed flawlessly with clean data, but when faced with messy inputs, incomplete information, and ambiguous intent, it continued to execute, albeit incorrectly. This raises questions about the current state of AI development and where the main bottleneck lies.

Some approaches, such as 60x AI, focus on structuring context and decision layers around workflows, rather than solely improving prompts or tool chaining. This has sparked curiosity about how people perceive the current limitations of AI. Is the primary challenge improving model outputs, such as refining prompts and retrieval, or is it enhancing how decisions are made across a system, including context, logic, and orchestration?

Photo by cottonbro studio on Pexels
Photos provided by Pexels