Why AI Projects Fail—and What the Successful 5% Do Differently

At a Glance

• According to the MIT study “State of AI in Business 2025,” 95% of companies are not deriving any measurable benefit from generative AI.

• The main reason is not the technology, but the approach: tools that do not learn and do not fit the process.

• Companies that focus on a specific pain point and deeply integrate AI into their workflows are the ones that succeed.

The Study: What MIT Investigated

In the summer of 2025, the NANDA Initiative at the Massachusetts Institute of Technology (MIT) published the report “The GenAI Divide: State of AI in Business 2025.” The authors analyzed more than 300 publicly documented AI initiatives, conducted 52 structured interviews, and surveyed 153 executives. The study period: January through June 2025. [1]

The results are sobering: Despite investments of $30 to $40 billion, 95 percent of organizations are not achieving any measurable return on investment. Only about 5 percent of integrated AI pilot projects generate significant added value—the vast majority have no measurable impact on the income statement.

Do Almost All of Them Really Fail? A Contextual Analysis

The figure is striking, and it deserves some context. [3] The observation period—about six months—was short; it does not capture long-term effects. The sample is not representative, and “success” was narrowly defined as a measurable impact on the income statement. Productivity gains by individual employees are barely reflected in the data—especially since many of them use personal AI tools that do not appear on any official balance sheet.

Nevertheless, the core finding remains robust: there is a deep chasm between experimenting with AI and deriving real business value. The study calls it the “GenAI Divide.”

Five Reasons Why AI Projects Fail

1. Lots of Testing, Little Adoption

General-purpose tools like ChatGPT or Copilot are widely used: Over 80 percent of organizations have tested them, and just under 40 percent are using them. However, they primarily boost individual productivity—not the company’s bottom line.

The gap is widest when it comes to customized, task-specific solutions: 60 percent of companies have evaluated such solutions, 20 percent reached the pilot phase—and only 5 percent went into production.

Important: This does not mean that specialization is the wrong approach. It means that most specialized solutions are too rigid, do not adapt, and operate separately from the actual process. The following reasons illustrate exactly that.

2. The Learning Gap: Systems That Don’t Learn

The study identifies a “learning gap” as the core problem: Most AI systems do not retain feedback, do not adapt to context, and do not improve over time. Users complain that AI does not learn from corrections, requires too much context for every query, and fails in special cases.

The result: For simple tasks like drafting emails, 70 percent of respondents are happy to use AI. For complex, long-term work, however, they prefer humans by a ratio of 9 to 1.

Generic tools like ChatGPT are excellent for individuals because they’re flexible. In a corporate setting, however, they falter because they don’t learn from or adapt to workflows.

– Paraphrased from Aditya Challapally, lead author of the MIT study, in an interview with Fortune [2]

3. Thinking Too Broadly

Many projects start with a grand vision: an AI platform for the entire company, centrally managed by an AI lab. The study shows that these projects are particularly prone to getting stuck. Companies that delegate responsibility to line organizations—and focus on a specific problem—are more successful. Challapally describes successful startups this way: They focus on a pain point, execute it well, and form smart partnerships. [2]

4. The Budget Is Being Allocated to the Wrong Areas

According to the study, more than half of the budget for generative AI goes to sales and marketing—areas where results are visible. However, the researchers found the highest return in the back office: in the automation of routine processes that had previously been outsourced or handled by external service providers.

5. In-House Development Without Experience

Solutions implemented in collaboration with specialized partners went live in about 67 percent of cases. Solutions developed entirely in-house succeeded in only about one-third of cases.

Side note: Shadow AI

Only 40 percent of companies have purchased official licenses for AI language models. Yet in over 90 percent of the companies surveyed, employees regularly use private AI tools for their work. This reveals two things: The demand is enormous—and company data often ends up where no one is monitoring it.

What the Successful 5% Do Differently

According to the study, companies on the right side of the divide follow a recognizable pattern:

• They work with specialized partners instead of building everything themselves.

• They delegate responsibility to the line departments—to the people who know the process.

• They demand deep integration into their operations and measure solutions against operational results, not model benchmarks.

• They rely on systems that retain context and learn from feedback.

The ENKO Approach: Targeted Rather Than A Scattergun Approach

If you read the study closely, a clear picture emerges: Success isn’t about buying the biggest model, but about precisely solving a specific problem—and ensuring that the solution gets better every day. That’s exactly how we work.

Why AI Projects Fail | How We Work at ENKO

Generic tool with no process relevance | Small model, trained for exactly one task in your process

The system doesn’t learn on its own | Corrections from your experts feed into the next training round—the model improves as it’s used

Too broad an approach | One pain point with quantifiable costs—then the next

Success isn’t measurable | Success criteria before launch: Detection rate, false alarms, turnaround time, hours saved

Pilot runs alongside the process | Integration into Business Central, production data acquisition (BDE), and quality management (QM) systems—right where the work happens

In-house development without experience | Partner with over 25 years of process experience—plus enablement to ensure the knowledge stays within your team

Concerns about your own data | Operated locally or in data centers in Germany or the EU

Small but Precise

Instead of applying a large language model to everything, we use small, specialized models exactly where they provide measurable benefits: in visual inspections, the classification of complaints, the extraction of information from documents, or the explanation of error codes. A model with a clear task is faster, more cost-effective to operate, and, above all, verifiable.

Learning on the Fly

We bridge the learning gap with a simple principle: The model makes a suggestion, and your experts confirm or correct it. Every correction becomes training data for the next round. Unclear cases are handled by humans, not automatically approved. This way, quality improves with every iteration—and so does employee trust.

Example: Pre-sorting Complaints

Incoming complaints are currently reviewed, assigned to a specific error type, and forwarded to the appropriate department—manually. A small model handles the pre-sorting directly within the QM system. The quality team reviews the suggestions, makes corrections where necessary, and each correction improves the model. Metrics are tracked from the very beginning: How often is the model correct, and how much time does the team save?

Integrated into the Process, Not Separate from It

AI that requires employees to open yet another window is rarely used. That’s why we integrate the results right where work is already being done: in the ERP system, in production data collection, and in quality management. This is the part where most projects fail—and the part we know best from over 25 years of software and ERP projects in the SME sector.

Checklist: Six Questions Before Launching Your AI Project

1. What specific problem are we solving—and what is it costing us today?

2. How will we know after three months whether the project is worth it?

3. What data is available, and who in-house can evaluate it from a technical perspective?

4. How does the system learn from corrections?

5. Into which existing system will the results be integrated?

6. Where will our data be processed—and who has access to it?

Conclusion

95 percent is not a law of nature. Above all, the MIT study shows how AI projects don’t work: too broad, too generic, without a learning loop, and disconnected from the process. On the other hand, those who start with a specific problem, train a small model specifically for that problem, measure success from the outset, and integrate the solution right where the work is being done have the best chance of ending up on the right side of the divide.

Sources

1. Challapally, A.; Pease, C.; Raskar, R.; Chari, P.: “The GenAI Divide: State of AI in Business 2025,” MIT NANDA, July 2025, report (PDF) – https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf

2. Fortune: “MIT Report: 95% of Generative AI Pilots at Companies Are Failing,” August 18, 2025, via Yahoo Finance – https://finance.yahoo.com/news/mit-report-95-generative-ai-105412686.html

3. AI in Human Resources: “MIT Study ‘State of AI in Business 2025’: Are Almost All of Them Really Failing?”, ki-im-personalwesen.de – https://www.ki-im-personalwesen.de/mit-studie-state-of-ai-in-business-2025-scheitern-wirklich-fast-alle/