Why do most enterprise AI projects fail?

Four separate research strands, from RAND, BCG, MIT NANDA and McKinsey, arrive at the same conclusion by different routes: failure is not caused by insufficiently good technology. It is caused by poor problem alignment between leadership and technical teams, by data that is siloed, incomplete or inconsistently formatted, by organizational resistance from people who were never told why, and by relying on one horizontal generic assistant applied identically everywhere. All four are independent of how good the model is.

What is the 70/20/10 rule in AI adoption?

Boston Consulting Group measured where failures actually originate and found that roughly 70% come from people and processes, 20% from infrastructure and data, and only 10% from algorithms. The same proportion shows up from other angles: MIT NANDA found that 95% of generative AI pilots show no measurable financial impact, and McKinsey found that 88% of organizations already use AI in at least one function while only 39% see an effect in EBIT. Three sources, one structure: adoption fails at scale-up, not at prototype.

How do you fix AI adoption if the technology is not the problem?

None of the common failure modes is fixed by a better model. Problem alignment is a conversation to be facilitated between leadership and technical teams. Data quality is a process discipline, not an infrastructure upgrade. Organizational resistance is untreated fear, not a bug. And over-reliance on one generic assistant is a prioritisation decision, not a model parameter. A good implementation partner solves the system side. The part where 70% of projects die stays on the table of the people who have to decide, align, and feel safe enough to say out loud that they do not yet understand what the new tool is for.

1 August 2026 · the Facilitare.ro team

Why 8 in 10 AI projects fail, and why it is not a technical problem

Of the 684 billion dollars invested globally in AI in 2025, close to 547 billion went into projects that did not deliver the expected result, more than 80% of deployed capital. The figure does not come from an opinion piece, it comes from RAND Corporation estimates. The useful question is not why AI fails so much, but where exactly the process breaks, and the answer is almost never the algorithm.

Colleagues in an office discussing around laptops and printed pages, in a working meeting. — Seventy percent of the failures sit with people and processes, not with the algorithm.

Four different research strands, from RAND, BCG, MIT NANDA and McKinsey, arrive by separate routes at the same conclusion: failure is not a problem of insufficiently good technology. It is a problem of alignment, of clean data, of people who were never told why, and of the horizontal architecture of the rollout. Four distinct ways to fail, all independent of how good the model is.

Four different ways to fail

1. Poor problem alignment

The organization buys a large language model and only then goes looking for a problem to point it at. Business leaders and the technical team are not speaking the same language about what the project is supposed to solve.

Only 15% of employees say their organization has communicated a clear AI strategy. Projects that begin with technology selection rather than with the business problem have the highest abandonment rates. Source: Talyx, "Why 90% of Enterprise AI Implementations Fail".

2. The data quality tax

Models depend on the data they ingest. Siloed, incomplete or inconsistently formatted data is not fixed by changing the model. It is fixed by the order in which the organization keeps its information, which is itself a human process.

Data contamination of just 20% degrades model performance by almost 10 percentage points. Gartner reports that 85% of projects fail because of poor data quality. Source: BERI, "Why 80% of AI Projects Fail: The Complete 2026 Analysis".

3. Organizational antibodies

Passive or active resistance to change. People find workarounds to avoid the new system, out of mistrust, fear of redundancy, or the absence of clear rules about what is allowed.

67% of executives surveyed by McKinsey name organizational resistance as the main barrier: it blocks precisely the feedback loops AI needs in order to be trained on the company's real cases. Source: BERI, "Why 80% of AI Projects Fail".

4. The horizontal illusion

The organization relies exclusively on a generalist, horizontal assistant applied the same way everywhere. Productivity gains distribute asymmetrically, a few people save time and the rest feel nothing, and the financial impact stays invisible in the reports.

Models adapted to a specific use case are far more likely to generate direct return, through revenue or cost, than a generic assistant spread across every workflow. Source: FullStack, "Generative AI ROI: Why 80% Fail".

The 70/20/10 rule

Boston Consulting Group measured where the failures actually originate, beyond their taxonomy. The result is the pivot of this whole article:

70% of failures come from people and processes, 20% from infrastructure, 10% from algorithms.

Source: AI Readi, "Why 80% of Enterprise AI Projects Fail", citing the BCG research. The same proportion appears from different angles in MIT NANDA, where 95% of generative AI pilots show no measurable financial impact, and in McKinsey, where 88% of organizations already use AI in at least one function but only 39% see an effect in EBIT. Three sources, one structure: adoption fails at scale-up, not at prototype.

The missing link is the human process, not the software

None of the four failure categories above is solved by a better model. Problem alignment is a conversation to be facilitated between leadership and the technical teams. Data quality is a process discipline, not an infrastructure upgrade. Organizational antibodies are untreated fear, not a bug. The horizontal illusion is a prioritisation decision, not a model parameter.

A good implementation partner solves the system side. The part where 70% of projects die stays, every single time, on the table of the people who have to decide, to align, and to feel safe enough to say out loud that they do not yet understand what the new tool is for. That is facilitation work, even when the topic in the room is AI. If you recognise the pattern in your organization, we have also written about what a facilitated AI adoption programme looks like.

The 70% nobody sells you

Alignment at leadership level, hands-on labs on your team's real workflows, and an internal network of champions that keeps adoption alive after the facilitators leave.

See the programme