When CellarTracker rolled out an AI-powered sommelier last spring, the goal was to give users blunt, personalized wine advice. Instead, the chatbot proved too agreeable, praising bottles that users were almost certain to dislike. It took weeks of fine-tuning to convince the system to offer honest criticism, a small but telling example of the broader challenges companies are facing as they race to deploy generative artificial intelligence.
Three years after ChatGPT ignited global excitement, businesses across industries have poured money and talent into AI projects, only to find that meaningful financial returns remain elusive. According to executives, advisors and multiple surveys cited by Reuters, most companies are still struggling to translate AI experimentation into improved margins or sustained revenue growth. A Forrester Research survey of more than 1,500 executives found that only 15 percent reported profit margin improvements from AI over the past year, while a Boston Consulting Group study showed just 5 percent of executives saw widespread value from the technology.
Despite continued belief that generative AI will eventually reshape business operations, executives are becoming more cautious about timelines. Forrester now expects companies to delay roughly a quarter of planned AI spending into 2026, reflecting a growing recognition that organizational change moves more slowly than the technology itself. Analysts say early expectations were fueled by tech companies promising rapid transformation that underestimated the complexity of human workflows and corporate systems.
A key obstacle has been AI’s tendency toward “sycophancy,” a bias that pushes models to please users rather than challenge them. CellarTracker’s experience highlighted how this trait can undermine decision-making tools meant to offer clear guidance. Other firms have encountered more serious problems, including inconsistency and factual errors. At Canadian rail services provider Cando Rail and Terminals, an internal chatbot failed to reliably summarize a core safety rulebook, sometimes misinterpreting or inventing regulations altogether. After spending about $300,000, the company paused the project, underscoring how far AI still is from being a dependable shortcut.
Customer service was once viewed as one of AI’s most promising targets, but here too expectations have softened. Klarna initially claimed its AI assistant could replace hundreds of human agents, only to later acknowledge that customers still wanted human interaction for complex issues. Verizon has also shifted back toward human-staffed support, using AI mainly to triage calls and handle routine questions. Executives told Reuters that empathy remains a major barrier to fully automated customer service.
Researchers describe these uneven capabilities as AI’s “jagged frontier,” where systems can excel at advanced tasks like coding or math but fail at simpler, contextual ones. Data complexity further complicates deployment, particularly in industries like finance, where inconsistent formats can cause AI tools to misread patterns. At Dutch investment group Prosus, an internal AI agent designed to answer portfolio questions still struggles with basic geographic and time-based concepts, limiting its usefulness.
As a result, companies are demanding more hands-on support from AI providers. OpenAI and Anthropic are increasingly embedding engineers and applied AI specialists directly with clients, aiming to tailor tools to specific business needs. OpenAI executives have warned against betting on massive, billion-dollar AI projects too early, instead encouraging companies to start with smaller, high-impact use cases. Startups focused on specialized industry models are also gaining traction, arguing that bespoke systems outperform general-purpose tools.
The shift reflects a sobering reassessment of generative AI’s role in the enterprise. Rather than a magic solution, executives now see AI as a powerful but demanding tool that requires deep integration, clean data and ongoing human oversight. As one industry leader put it, the technology is impressive, but making it truly useful still takes far more work than many companies first imagined.
