The echoes of a decades-old observation are resonating in boardrooms across France. In 1987, Nobel laureate Robert Solow famously noted, “You can see the computer age everywhere but in the productivity statistics.” Today, a similar sentiment is emerging as companies grapple with the implementation of generative artificial intelligence. Despite the intense hype surrounding tools like ChatGPT, tangible economic gains remain elusive. A recent, large-scale survey by the National Bureau of Economic Research, encompassing 6,000 business leaders in the United States, Europe, and Australia, found that nearly nine in ten companies have seen no discernible impact on either employment or productivity from generative AI over the past three years.
This lack of immediate impact isn’t simply a matter of slow adoption. It speaks to a fundamental disconnect between the promise of the technology and the realities of integrating it into established organizations. French “grands comptes”—leading companies like Air France, Total, and Axa—are investing heavily in generative AI, with some projects exceeding six-figure budgets. Yet, as reported by L’Express, the results on the ground are often modest. The core challenge? Data. Generative AI relies on large language models (LLMs) that require vast amounts of information to function effectively. But many large organizations struggle with fragmented and inaccessible data, a legacy of decades of disparate systems.
The Data Silos of Legacy Systems
Unlike nimble startups that build databases on modern architectures, established companies often contend with what professionals call “data silos”—isolated pockets of information that don’t communicate with each other. Imagine a 30-year-old accounting software system struggling to interface with a cutting-edge marketing tool. This “legacy” infrastructure presents a significant hurdle. The difficulty isn’t simply technical; it’s organizational. Breaking down these silos requires a fundamental shift in how data is managed and shared across departments, a process that can be slow and politically fraught.
At LCL bank, Didier Lellouche, head of artificial intelligence, describes a “graveyard of POCs”—proofs of concept that demonstrate the technology’s potential but fail to translate into operational systems. “As soon as we wanted to integrate a generative AI project into the information system, it stopped working. Too complicated, too expensive,” he explained. The cost of “inference”—the process of running AI models, billed per unit of text known as a token—is also a concern, with companies seeking greater transparency in pricing.
Limited Production, High Expectations
The challenges extend beyond data and cost. Bpifrance, a French investment bank, provides a stark illustration of the difficulties. Of 240 tests using generative AI, only 17 have been put into production, according to Lionel Chaine, the bank’s director of information systems. “Everyone is learning,” Chaine stated, “and unfortunately, this time, You can’t buy experience from consulting firms to guide us.” Ghislain de Pierrefeu, a consultant at Wavestone, which has assisted numerous large French companies with their AI transformations, echoes this sentiment. “There’s a real disconnect between the prevailing discourse—the idea that generative AI is changing the world—and the reality of the gains. Everyone is struggling.” One executive went even further, stating bluntly, “Generative AI is oversold.”
The Cost of Inaction and the Demand for Strategic Investment
The current situation isn’t necessarily a condemnation of the technology itself, but rather a reflection of the complexities of implementing it within large, established organizations. The initial wave of enthusiasm has given way to a more sober assessment of the challenges involved. Companies are realizing that simply throwing money at AI isn’t enough. Successful implementation requires a strategic approach that addresses data infrastructure, organizational culture, and employee training.
The lack of immediate returns is also prompting a reevaluation of investment strategies. Rather than pursuing broad, ambitious projects, companies are beginning to focus on specific, well-defined use cases where generative AI can deliver tangible value. This often involves automating repetitive tasks, improving customer service, or enhancing decision-making processes.
Beyond the Hype: A Realistic Outlook
The initial promise of generative AI—a rapid and transformative impact on productivity—appears to be overstated, at least in the short term. The National Bureau of Economic Research findings align with earlier research from MIT, which found that 95% of enterprise AI projects fail to deliver a return on investment. This isn’t to say that generative AI is a failure, but it does suggest that the path to realizing its potential will be longer and more challenging than many initially anticipated.
The focus now is shifting towards building a solid foundation for future growth. This includes investing in data infrastructure, developing internal expertise, and fostering a culture of experimentation. Companies are also exploring new approaches to AI governance and risk management, recognizing the potential for bias and misuse.
The situation in France mirrors a global trend. While the potential of generative AI remains significant, realizing that potential requires a pragmatic and strategic approach. The initial hype cycle is subsiding, replaced by a more realistic assessment of the challenges and opportunities ahead. The next key development will be the release of more detailed case studies demonstrating successful implementations of generative AI within large organizations, providing a roadmap for others to follow.
What are your thoughts on the challenges of implementing generative AI in large organizations? Share your experiences and insights in the comments below.
