Why AI pilots stall before production, and how to pick one that will not

The most useful figure comes from S&P Global Market Intelligence. In its 2025 survey of about 1,000 IT and business respondents, 42% said they had abandoned most of their AI initiatives before production, up from 17% a year earlier. Note what that measures: the share of companies that dropped the majority of their initiatives. It is often misquoted as 42% of AI projects failing.

Gartner predicted in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. That is a forecast, not a measurement.

The figure you will see most often is that 95% of generative AI pilots fail, from an MIT report in 2025. It rests on a small set of interviews and survey responses and a narrow definition of success, so we do not use it. There is also evidence in the other direction: a 2025 survey of about 800 senior leaders by Wharton and GBK Collective found that roughly three in four reported a positive return on generative AI.

Both pictures can be true. Use is broad and many leaders see value. Getting a specific system into production, with a measured effect, is where many efforts still stop.

In 2024 RAND interviewed 65 experienced data scientists and engineers about failed AI projects and identified five root causes. The most common was that the problem was misunderstood or poorly communicated. The others were inadequate data, a focus on the technology instead of the problem, weak infrastructure, and problems that were simply too hard for the technology.

None of the five is about which model was chosen. Put in practical terms, a pilot stalls when nobody can say how often it is right, what it costs, or who owns it.

The best independent evidence of benefit is a 2023 study of 5,179 customer support agents by Brynjolfsson, Li and Raymond. Agents given an AI assistant resolved 14% more issues per hour on average. The newest agents improved by 34%, and the most experienced barely changed.

The pattern in that result is worth copying. The AI assisted people on high-volume work. It did not replace them, and a person stayed responsible for the outcome.

Retrieval-augmented generation (RAG), where the model answers from your own documents, reduces made-up answers. It does not remove them. In 2024 researchers at Stanford tested commercial legal research tools that already use retrieval and found incorrect or misgrounded answers on more than 17% of queries for two products and more than 34% for a third.

This is why a serious build includes a test set of real questions with known answers, run on every change, and a fallback for answers the system is unsure of. Without that you have an opinion about accuracy, not a measurement.

A good first project passes all five of these.

• Volume: The task is done hundreds of times a month, so a saving is visible.
• A number you already track: Time per item, cost per document or error rate, measured before you start.
• Data you can reach: The documents exist in one place and someone can grant access to them.
• A tolerable failure: A wrong answer is caught by a person before it costs money or trust.
• An owner: One person decides whether it worked.