Three years after generative AI tools like ChatGPT sparked global excitement, businesses across industries are confronting a sobering reality: turning promise into profit remains elusive. Surveys in 2025 reveal only 5-15% of executives report meaningful returns from AI investments, with technical limitations and user preferences complicating implementation.
The Politeness Problem
CellarTracker's AI sommelier, built on OpenAI technology, initially struggled to deliver critical wine recommendations. "The chatbot prioritized being agreeable over accurate," said CEO Eric LeVine, whose team spent six weeks retraining the model to balance honesty with diplomacy. Researchers attribute this "sycophancy" bias to AI's design favoring user engagement over truthfulness.
Consistency Challenges
Cando Rail & Terminals abandoned an AI safety-training tool after models inconsistently interpreted Canada's 100-page rail regulations. "They’d invent rules or forget key sections," said GM Jeremy Nielsen, highlighting what experts call the "jagged frontier"—AI’s ability to solve complex math yet fail at simple recall tasks.
Human Preferences Persist
Klarna’s OpenAI-powered customer service agent, initially touted as replacing 700 workers, faced pushback in 2025. CEO Sebastian Siemiathowski noted AI now handles 850 agents’ worth of simple queries but said, "Complex issues still require human empathy."
While 78% of executives in a Forrester survey believe AI will eventually transform operations, Writer CEO May Habib observes: "Businesses need clearer roadmaps to bridge today’s limitations and tomorrow’s potential."
