Why 8 out of 10 AI projects fail and what the others do differently

Florian Maier
Jul 19
5 min read

The figures on the state of AI in business use are sobering. Over 80 percent of AI projects fail, twice as many as traditional IT projects [1]. 42 percent of companies abandoned the majority of their AI initiatives in 2025, compared to 17 percent the year before [2]. And an MIT study concludes that 95 percent of all GenAI pilots have no measurable impact on business results [3].

What's remarkable about these studies is that the technology is almost never the cause. The models work. What fails is the surrounding organization. We analyzed the most robust studies of the last two years and compared them with what we see in our own projects. Five patterns explain the majority of failures, and none of them are related to the choice of model.

1. The project starts with a technology instead of a problem.

The RAND Corporation surveyed 65 experienced data scientists and engineers about failed AI projects. The result is clear: 84 percent of the industry respondents cited management decisions as the main cause, primarily a misunderstood or poorly communicated problem [1]. Projects are launched because the board has AI on its agenda, not because someone has quantified a specific pain point.

A principle older than any language model can help counter this: Eliyahu Goldratt's Theory of Constraints. Every organization has exactly one limiting factor at any given time, and any improvement outside of this bottleneck is an illusion of progress. Applied to AI: Before you ask what the technology can do, quantify which process is demonstrably costing you revenue, margin, or time today. Only once this question is answered is it worthwhile to look at tools. Otherwise, you end up with what we regularly encounter in initial consultations: a dozen half-finished automations, none of which deliver any business value.

2. The pilot is brilliant, but production never happens.

According to S&P Global, the average company discards 46 percent of its AI proofs of concept before they ever go live [2]. The demos are impressive, but they don't survive the transition to everyday use. The reason is rarely the model itself. The operational infrastructure is lacking: defined triggers, connected data sources, a verification step, a responsible party, and monitoring.

MIT researchers describe this same finding as a "learning gap": Generic chat tools are brilliant for individuals, but they don't learn anything about the company's processes and don't adapt to them [3]. An employee copying results from a chat window isn't automation, but rather a person with a faster, pre-defined text snippet. If that person is unavailable, the process grinds to a halt. Measurable effects only emerge when AI is embedded in a workflow that functions even without a specific individual. Interestingly, according to MIT, purchased, specialized solutions reach production readiness about twice as often as in-house developments [3].

3. The data basis decides before the model calculates.

In the RAND survey, 30 out of 50 interviewees discussed data quality as a chronic problem; one participant put it this way: 80 percent of AI work is the dirty work of data engineering [1]. The global CDO Insights Survey confirms this: data quality and readiness are the most frequently cited obstacle to AI projects, at 43 percent [4].

In practice, this doesn't mean that medium-sized businesses lack data. Meeting notes, quotes, support histories, project documentation: the material is there, it's just scattered across mailboxes, drives, and people's minds. No system can compensate for a neglected CRM. Those who automate before clarifying where information originates, where it's stored, and who is responsible for its accuracy will produce the same errors as before, only faster and in greater numbers.

4. The budget ignores the people.

The Boston Consulting Group recommends a resource allocation of 10/20/70 for AI transformations: 10 percent for algorithms and tools, 20 percent for technology and data, 70 percent for people and processes [5]. Most companies budget in exactly the opposite way. They buy licenses and treat training, process redesign, and responsibilities as a footnote.

The consequences are measurable. McKinsey shows that managers systematically misjudge AI usage and their employees' concerns: Employees are three times more likely than their superiors to believe that AI will replace 30 percent of their work within a year, and 21 percent report receiving little or no support in using AI [6]. At the same time, 70 percent of marketing professionals state that their employer offers no training on generative AI [7]. A tool that no one can use confidently and that no one trusts delivers no ROI, no matter how good the model is.

5. Without rules, AI becomes a compliance risk.

Where companies don't offer AI processes, employees build their own. Studies on shadow AI show that 77 percent of employees using generative tools also copy company data into public services, from customer data to source code [8]. Samsung restricted the use of generative AI internally after just such an incident. For companies subject to the GDPR and the EU AI Act, this is not a minor offense, but a liability issue.

The solution isn't a culture of prohibition, but a concise, written set of rules: Which tools are permitted, which data is allowed, who reviews results before they reach customers, and which decisions are always made by a human. Because ultimately, responsibility cannot be delegated to a system. Automation shifts control to the appropriate place; it doesn't eliminate it.

The cognitive bias behind all five patterns

Investor Howard Marks distinguishes between first-level and second-level thinking: First-level thinking assesses the immediate consequences of a decision, second-level thinking the consequences of those consequences. Most failed AI projects are based on first-level decisions. The license saves time immediately, so it's purchased. What happens in twelve months when the necessary expertise is lacking because entry-level tasks have been automated, when core processes depend on a single provider, or when no one remembers why the system makes the decisions it does—none of this is addressed in any presentation.

Three questions before every AI decision bring a second layer of thinking into play: What will change immediately? What will change in a year? And what if the provider doubles prices, changes its model, or disappears from the market? Anyone who cannot answer the third question is not making a strategic decision, but rather following an impulse.

The order that works

The studies and our own project work reveal a clear sequence. First, quantify the problem: a bottleneck, measurable damage, and a responsible party. Then, define the process that the AI should take over, including the criteria for recognizing a good result. Next, clarify the data basis. Then, train the people and establish the rules of engagement. And only at the end, select the tool, as the smallest and most easily replaceable part of the system.

This sequence is unremarkable, and precisely why it's so rarely followed. But it explains the difference between the 42 percent who drop out and the companies where the investment pays off: The competitive edge doesn't come from a license that anyone can buy tomorrow. It comes from clearly defined processes and a database that becomes more valuable with each month of operation.