Why most AI projects fail in finance

A recent study by MIT’s Media Lab revealed a sobering statistic: across the companies surveyed, 95% of corporate AI initiatives fail to deliver a measurable return. The report clarifies that this is not a technical failure, but rather a lack of P&L impact in the short term. While many organizations have explored or even deployed general-purpose tools like ChatGPT or Copilot, the study found these are primarily enhancing individual productivity, not translating into company-wide financial gains. While 80% of companies have explored or piloted these tools, high-impact enterprise-grade systems are seeing limited adoption, with only 20% of surveyed companies reaching the pilot stage and just 5% making it to full production.

These findings directly contradict the “AI gold rush” narrative that has been a fixture of the recent news cycle and bear a striking resemblance to the dot-com bubble of the mid-1990s. The head of OpenAI himself, Sam Altman, has acknowledged the situation, stating that he believes the AI market is in a bubble and that “smart people get overexcited about a kernel of truth”. While AI certainly has the potential to transform industries, unrealistic expectations about its short-term impact have led many businesses to over-invest in projects that ultimately provide little to no return. We have already seen cautionary tales, such as Klarna, which had to rehire staff after realizing its AI chatbots could not completely replace human agents.

While AI has been a game-changer for media and tech, traditional industries like finance have had a much harder time. The reason is that custom solutions often stall due to integration complexity and a poor fit with existing workflows. To be useful for financial operations, AI implementations must integrate with a company’s legacy systems, which are often built on massive, siloed, and poor-quality data. Since AI’s utility is only as good as the data it’s trained on, the quality of these tools is often lower in such cases.

In addition, major regulatory and ethical hurdles make AI adoption a difficult task in the financial sector. The use of AI in finance demands Explainable AI (XAI) and Transparency by Design to satisfy regulators and build customer trust. The risk of automation bias and the need for human oversight are essential requirements to avoid perpetuating historical biases in lending or fraud detection models.

A critical takeaway for any finance corporation aiming to adopt AI is to start with a specific business problem, not a technology solution. Many companies have launched initiatives with AI simply to be an early adopter, which has forced them to retroactively search for a problem to solve. This often leads to implementations that are never fully utilized. Corporations should always identify their pain points before deciding on a solution. For example, Deutsche Bank’s fraud detection overhaul began with a clear issue: high false positive rates were killing the customer experience. This business problem then led them to the solution of using AI to overhaul their fraud detection system, resulting in a successful implementation that directly addressed a key pain point. Moving the needle on P&L requires this kind of focused, high-impact approach in areas like fraud detection, back-office automation, and predictive analytics for credit and underwriting.