Why enterprise AI projects fail

A company buys an AI tool, sets up a pilot, gives an encouraging demo and, months later, the project vanishes. Nothing visibly broke. It simply never became routine or delivered the expected result.

This script is common enough to have become a research topic. This article brings together what dated studies and surveys say about why AI projects in companies fail, and what that suggests for anyone about to start one.

How much actually fails

There is no official failure rate for AI projects. The figures in circulation measure different things.

• “More than 80%.” The RAND Corporation report of August 2024 cites estimates that more than 80% of AI projects fail, twice the rate of information technology projects that do not involve AI (RAND, 2024). RAND itself presents the figure as an estimate from other sources, not its own measurement.
• 42% abandonment. In S&P Global Market Intelligence’s Voice of the Enterprise survey of 1,006 IT and business professionals in North America and Europe, the share of companies that abandoned most of their AI initiatives before production rose from 17% to 42% in one year (S&P Global, May 2025). On average, respondents said 46% of projects are scrapped between proof of concept and broad adoption.
• 74% with no tangible value. Boston Consulting Group’s survey of 1,000 executives in 59 countries found that 74% of companies had yet to show tangible value from AI (BCG, October 2024).
• Forecasts. Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 (Gartner, July 2024). In June 2025, it predicted that more than 40% of agentic AI projects would be canceled by the end of 2027 (Gartner, June 2025).

A proof of concept (PoC) is a small test to see whether an idea works before investing in earnest. Abandoning a PoC is not always failure: sometimes finding out early that the idea does not work is the right outcome. The problem is when abandonment comes after heavy spending, or when the project should never have started.

Note also the type of each figure. S&P and BCG are opinion surveys of executives and professionals. Gartner’s are analyst forecasts. None of these figures was measured in Brazil.

Cause 1: the wrong problem, or a poorly explained one

The RAND report is one of the few that look at the subject from inside the projects. The authors interviewed 65 experienced professionals, 50 from industry and 15 from academia, between August and December 2023. One important caveat: the study covers machine learning projects and leaves out those that only used off-the-shelf language models through prompting.

The most cited cause was business leadership misunderstanding, or miscommunicating, what problem AI should solve. According to RAND, 84% of industry interviewees pointed to at least one failure of this kind as the main reason projects failed.

The report gives a simple example. Leadership asks for a model that sets the price of a product. The technical team optimizes for the price that sells the most units. What leadership wanted, however, was the price that yields the highest profit margin. The model works and still does not serve its purpose.

RAND describes other variations of the same mistake:

• using machine learning where a few simple “if this, then that” rules would do;
• expecting certainty from a system that is, by nature, probabilistic;
• underestimating the time needed to obtain, clean and explore the data (14 of the 50 industry interviewees reported this);
• changing priorities every few weeks, killing projects before they show results.

Cause 2: data that does not work

The second most cited cause in the RAND report was data. Thirty of the 50 industry interviewees talked about persistent quality problems.

Having a lot of data is not enough. RAND notes that legacy databases were usually built for record-keeping or compliance, not analysis. An online store may record which link a customer clicked, but not what was on the screen at the time or which search led them there. For training a model, that difference is decisive.

Other problems mentioned:

• imbalanced data, with few examples of the rare case that matters most, such as an uncommon disease;
• a shortage of data engineers, the professionals who build and maintain collection and cleaning routines;
• a lack of business knowledge on the technical team, which does not know which fields are reliable.

Gartner also lists poor data quality among the reasons generative AI projects are abandoned after the PoC. In systems that answer based on company documents, such as RAG systems, an outdated document becomes a wrong answer.

Cause 3: technology in search of a problem

RAND identified the reverse of cause 1: the technical team picks the newest tool because it is interesting, not because it is the best for the case. Sixteen of the 50 industry interviewees mentioned this pattern.

Gartner points to something similar in the agent market. In June 2025, the firm said most agentic AI projects were early-stage experiments driven by hype and that many use cases presented as agentic do not need agents. It also described agent washing: older products, such as chatbots and process automation, rebranded as agents without real capability to match. We explain what an agent is in what are AI agents and where they tend to fail in the limits of autonomous agents.

Cause 4: integration and infrastructure

A model that works on a data scientist’s laptop is not yet a production system. RAND points to underinvestment in infrastructure: pipelines that clean and deliver data continuously and monitoring that warns when a source changes format or stops arriving.

This problem has been known for a long time. A paper by Sculley and co-authors presented at the 2015 NeurIPS conference described the technical debt of machine learning systems: large, ongoing maintenance costs that do not show up in the prototype (Sculley et al., 2015). The risk factors listed include data dependencies, hidden feedback loops and changes in the external environment.

Gartner adds that integrating agents into legacy systems can be technically complex and require costly changes. In many cases, according to the firm, the best path is to redesign the workflow rather than fit AI into the old process. We cover how to map processes before automating in how to automate processes with AI.

Cause 5: high costs and value that is hard to prove

Cost and value show up in every market source.

• Gartner cites escalating costs and unclear business value as reasons for abandoning projects, both generative AI (2024) and agents (2025). In its 2024 press release, it estimates that certain generative AI deployment approaches cost from US$ 5 million to US$ 20 million and states that these costs are less predictable than those of other technologies.
• In S&P Global’s 2025 survey, costs were cited as a challenge by 37% of respondents. Cost was also the most common criterion for choosing projects and the worst-performing metric. Among companies that invested in generative AI, 46% said no business objective had seen a “strongly positive impact.”

Part of the problem is measurement. Individual productivity gains do not always turn into financial results. We discuss how to calculate costs and benefits in AI ROI in business.

Cause 6: governance and risk control

“Inadequate risk controls” appear in both Gartner forecasts as a reason for abandonment. The S&P Global survey found a complementary pattern: companies with lower failure rates were comparatively more likely to consider compliance, risk and data availability when choosing projects.

AI governance is the set of rules, roles and processes that defines who decides, who is accountable and how risks are tracked. The AI risk management framework from NIST, the US standards institute, treats governance as a cross-cutting function present in every stage of mapping, measuring and managing risks (NIST AI RMF 1.0, January 2023). The document is voluntary, but it serves as a roadmap.

Without this care, an error that reaches the customer can cost more than the project. The case of Air Canada’s chatbot, which gave wrong information and led to the company being held liable, is discussed in AI in customer service. Before releasing a system, it is worth testing it with real cases, as we show in how to evaluate AI responses.

Cause 7: people and changes to work

BCG’s 2024 survey concluded that about 70% of implementation challenges come from people and processes, 20% from technology and only 10% from algorithms. The companies BCG classifies as leaders follow this proportion when allocating resources. Among the scaling factors most tied to people and processes, the report lists change management, workflow redesign, talent and governance.

Other sources point in the same direction:

• RAND reports that business experts sometimes passively resist AI projects out of fear for their own jobs. Without them, the technical team does not know what the data means.
• S&P Global observed that companies with more failed projects report more resistance from customers and employees. A shortage of qualified professionals was cited as a concern by 27% of respondents.

Cause 8: what AI cannot do

Finally, RAND records projects that fail because the problem is too hard for the available technology. In the report’s words, AI “is not a magic wand.” To avoid this mistake, the recommendation is to involve technical experts in assessing feasibility before approving the project. Errors inherent to the models, such as hallucinations, also factor in.

Questions to ask before you start

Based on RAND’s recommendations and the factors pointed out by the other sources, the newsroom suggests a minimum checklist:

1. Which decision or task will change, and which number will show that it changed?
2. Would a simple rule or a better process solve it without AI?
3. Does the necessary data exist, is it reliable, and can someone on the business side explain it?
4. How will the system connect to current systems, and who will maintain it after the pilot?
5. How much does it cost to operate, not just to build?
6. Who is accountable for an error, and how will the error be detected?
7. Did the people who will use the tool take part in its design, and was the workflow adjusted?

Key takeaways

• Failure figures depend on what is measured. In 2025, S&P Global found 42% of companies abandoning most of their AI initiatives. The “more than 80%” estimate cited by RAND comes from other sources.
• According to RAND (2024), the most cited causes are a poorly defined or poorly communicated problem and inadequate data.
• Integration, infrastructure and maintenance cost more than the prototype suggests.
• Unpredictable costs, hard-to-prove value and weak risk control are among the reasons for abandonment cited by Gartner.
• People and processes matter more than the algorithm: in BCG’s survey, about 70% of the challenges come from there.