It is now just over three years since ChatGPT was launched publicly by OpenAI, and ignited a wave of interest in AI and huge investment in companies providing AI products and services. OpenAI itself has been publicly pondering a 2026 IPO at a potential valuation of $750 billion or more. NVIDIA, which makes the specialist chips used in most AI processing, is the most valuable company on the planet. NVIDIA had a market capitalisation of $4.3 trillion in December 2025 compared to $337 billion in October 2022, just before ChatGPT appeared. This is a 12.7-fold increase in share price in just three years. ChatGPT and its rivals, Claude, Gemini and others, are being used for everything from writing marketing material to coding, from translation to therapy. Now that the world has been using generative AI for three years, what impact has it really had?
Curiously, despite the avalanche of interest in AI from the public, companies and government policymakers, and concerns about AI replacing whole industries or even causing the end of the world, the real-world impact has so far been more of a whimper than a bang. Multiple recent reports from AI implementations reveal a number of practical problems in moving from demonstrations and experiments to practical projects that can be put into production.
A Boston Consulting Group report that emerged at the end of September 2025 found that just 5% of the 1,250 companies surveyed were achieving significant value from their AI initiatives. This is consistent with an August 2025 MIT report, based on interviews with hundreds of company executives, which found that 95% of AI projects fail to deliver any measurable monetary impact.
For example, Cando Rail and Terminals tested an AI chatbot on a task that should be right up the street of a large language model (LLM). It was asked to summarise a 100-page document that lays out industry safety standards for the railway. Although the LLM produced summaries, close inspection of them showed them to be wrong. Sometimes the model would forget an important rule, at other times it would misinterpret rules, and sometimes it would just make up new rules entirely, which is the well-known LLM phenomenon of “hallucinations”.
Sometime earlier, back in 2024, there had been the widely reported case of the fintech firm Klarna, which replaced 700 customer service staff with an AI chatbot, but then had to rehire those staff as customer complaints about the new chatbot mounted. In another example, Zendesk uses LLMs to handle more than half of customer service requests, but has encountered many issues. Shashi Upadhyay, president of product, engineering and AI at customer-service platform Zendesk, said that the idea that AI could handle everything was “oversold”. AI has made progress in software coding and debugging, but even here, the actual effect on productivity has been a very mixed story. The largest survey in the industry, the DORA report, found a negative impact on programmer productivity and an increase in code instability. One “industry” that has definitely found AI useful has been hackers, since generative AI brings a whole range of new security vulnerabilities, quite apart from the issue of deepfakes.
To be sure, there are some generative AI success stories, but fewer than you might have expected after three years of implementation and huge global investment. When I have dug deeply into some published AI success stories, many of the more impressive ones turn out to have been based on older machine learning AI, rather than generative AI or agentic AI, which itself is immature and has been struggling with handling real-world problems. Public perception of AI reveals that people are more concerned than excited about AI.
As we look out into 2026 and beyond, it seems clear that there are many challenges with converting what is, without doubt, a technology that has generated excitement, into something that delivers widespread value. Many of the AI failures can be attributed to using generative AI for tasks to which it is ill-suited, due to its probabilistic nature. Generative AI will struggle in situations where consistency and accuracy are important, since by their very nature, LLMs are inconsistent and prone to hallucinations. In some situations, this does not matter much or at all, such as in image generation, where LLMs are undoubtedly making strides and are starting to impact the industries of graphic design, advertising, and even modelling. LLMs will continue to develop and improve in many ways, and people will become better at figuring out where they are well-suited and where they are not. However, 2026 is shaping up to be a year in which AI needs to deliver greater economic value than it has so far for corporations, public services and consumers, rather than just to investors in AI companies.
