AI cannot currently scale to meet the demands of high-volume, low-error-tolerance business processes. Despite the Agentic AI and automation, foundational limitations remain: AI’s probabilistic nature inherently produces variance, which compounds across complex workflows. From payroll to compliance, enterprise use cases demand precision, not educated guesses. Until models overcome these structural issues, AI will remain most valuable in low-volume use cases.

A review of OpenAI’s recent paper on AI hallucinations, which argues that benchmark and evaluation incentives are the root cause effectively blaming the “test” for the model’s guessing behavior. While the explanation is elegant, it's a copout. I suggest the field needs deeper exploration into “IDK models”. If hallucinations can't be eliminated can AI companies design a sustainable, trustworthy tradeoff between confidence and hallucination?