A series of studies have now shown that corporate AI projects are failing at a disturbingly high rate. MIT in August 2025 found that 95% of AI projects fail to deliver any economic return. The same was found at the same 95% failure rate by Boston Consulting Group in September 2025. Similar, albeit slightly lower, failure rates have been reported in studies from Rand Corporation (80%) in August 2024, and 85% by Gartner. Reasons for this poor performance include poor data quality, workforce resistance, misaligned expectations and difficulty in integration with corporate systems. Large language models (LLMs) are probabilistic in nature, and yet we are used to deterministic computer systems that behave consistently and reliably. LLMs are good at things like brainstorming, pattern matching, researching websites and generating text, but struggle with simple arithmetic, truthfulness, consistency, causal understanding and novel problem-solving. What are you trying to use an LLM for? Is it likely to perform well, or are you trying to put a square peg (an LLM) into a round hole (a problem it is ill-suited to)? Some projects where LLMs fail were doomed from the start as they were simply the wrong tool for the job.
This all raises a key question. If 95% are failing, what are the 5% doing right? Without doubt, there have been successful AI projects. Some of these are based on machine learning rather than large language models. For example, BMW uses AI for predictive maintenance, claiming enhanced efficiency. In 2023, PayPal used deep learning models to achieve an 11% improvement in losses through fraud. Walmart used a store-floor robot to monitor shelf inventory and trigger restocking decisions. This resulted in a 35% reduction in inventory and a 15% improvement in inventory accuracy. Other AI tools are also in use at the company. Siemens used AI (machine learning) to achieve a 35% improvement in factory downtime by predicting machine failure better. Massachusetts General Hospital used AI automated note-taking and updates to patient records. Results of such note-taking in general have mixed reports, with one case finding 62% improvement in note-taking that resulted in increased patient/doctor face time, yet in other cases, not so. Nonetheless, doctors seem happy with it.
What are the common factors of successful AI projects? One seems to be taking a limited scope and tackling niche problems that cause measurable pain. An example might be compliance reporting or customer churn prediction, or demand forecasting. Use cases are selected where there is clear economic value and where they are feasible. Another common factor seems to be a keen awareness of the readiness of data to support the project. The relevant data is accurately mapped, its quality assessed and monitored on an ongoing basis. Successful projects have good business sponsorship rather than being IT-driven, and where security reviews are baked into the project plan. LLMs in particular have a range of peculiar security risks like prompt injection and data poisoning vulnerability as well as model drift, and steps need to be taken to mitigate these risks. A 2026 PwC report, “AI Business Predictions”, found that successful companies set up centralised “AI studios” of experts to support line-of-business initiatives.
Many of these themes are actually not peculiar to AI but could apply to any new technology. Ensure that you have a rigorous business case for investment and prioritise manageable projects that should have a rapid return on investment. Set up a centre of excellence with expertise that can support specific business projects. Carefully review the technologies used and the risks involved (including security). Ensure that projects have strong business sponsorship, and that the technology infrastructure and data to support that project are in place. Carefully consider ongoing monitoring of projects (and AI models) and maintenance, not just implementation. These should be basic disciplines, yet in all the excitement about generative AI, some of these old lessons seem to have been forgotten. Those who remember them will be more likely to be in that rarefied 5% air of success.
