AI Reality Check: ROI Struggles Force Delays

Three years after the explosion of generative artificial intelligence, businesses that rushed to adopt the technology are facing a sobering reality.

While many expected a quick revolution, executives and surveys now reveal that realizing a meaningful return on investment is proving far more difficult than anticipated.

When ChatGPT launched, companies worldwide scrambled to integrate generative AI into their operations. However, recent data suggests that the financial payoff has been minimal for most.

According to a second-quarter survey of over 1,500 executives by a leading research firm, only 15% reported improved profit margins due to AI over the last year. A separate study found that just 5% of 1,250 executives saw widespread value from their AI initiatives between May and July.

Rethinking the Timeline

While corporate leaders still believe generative AI will eventually transform their industries, they are recalibrating their expectations regarding speed. Predictions now suggest that by 2026, companies may delay approximately 25% of their planned AI spending as they grapple with the complexity of implementation.

Analysts point out that while tech giants have marketed AI as a rapidly transformative force, organizational and human adaptation is naturally slower.

This reassessment comes amidst unprecedented investment in infrastructure—from chips to data centers raising concerns that a failure to demonstrate clear revenue boosts or efficiency gains could trigger a market crash similar to the dot-com bust of the early 2000s.

The “Sycophancy” Problem

One major hurdle for businesses has been the tendency of AI models to be overly agreeable, a trait researchers call “sycophancy.”

CellarTracker, a wine-collection app, encountered this when building an AI sommelier. The chatbot was initially too polite, refusing to tell users they likely wouldn’t enjoy a specific wine even when the data suggested otherwise. The company spent six weeks fine-tuning the model to provide honest, critical appraisals.

“We had to bend over backwards to get the models to be critical,” said CEO Eric LeVine.

Accuracy and Consistency Issues

Beyond tone, reliability remains a significant barrier.

Cando Rail and Terminals, a North American railroad service provider, attempted to use an AI chatbot to help employees study safety regulations. The company spent $300,000 on the project only to find that the AI could not consistently or correctly summarize the industry’s critical operating rules.

The model sometimes misinterpreted safety standards or hallucinated entirely new rules, a common issue known as the “jagged frontier” where AI excels at complex tasks like coding but fails at seemingly simple information retrieval. Cando has since paused the project.

“We all thought it’d be the easy button,” said General Manager Jeremy Nielsen. “And that’s just not what happened.”

The Human Element Returns

The initial belief that AI would replace human workers in customer service is also being walked back.

Swedish payments firm Klarna rolled out an AI agent in 2024 intended to replace 700 full-time staff. By 2025, the CEO admitted they had to dial back the initiative because customers still preferred human interaction for complex issues.

Similarly, telecommunications giant Verizon is leaning back into human-staffed service for 2026. The company uses AI to handle routine data collection and call screening but relies on its 2,000 frontline agents for empathy and problem-solving.

“Empathy is probably the key thing that’s holding us from having AI agents talk to customers holistically right now,” said Ivan Berg, a Verizon executive.

The Path Forward: Specialized Partners

Recognizing these challenges, major AI providers like OpenAI and Anthropic are shifting their strategies. They are moving away from simply selling technology to becoming active partners, embedding experts within client companies to co-develop solutions.

New startups are also emerging to fill the gap, creating specialized tools for specific sectors like finance and law rather than general-purpose chatbots.

“Companies need more handholding in actually making AI tools useful for them,” noted May Habib, CEO of the AI startup Writer.

As the hype settles, the consensus is clear: AI is not magic. It requires significant customization, clean data, and a realistic understanding of its current limitations to deliver true business value.

READ ALSO: Arteta Challenges Jesus to Lead Attack