The problem is never the tool. It's the process around it

An accounting firm deploys an assistant that cuts the time spent reviewing case files in half. It works. It is faster, more consistent and cheaper than the old process. Six weeks later, almost nobody uses it. The tool is still there, untouched. The organisation has simply gone back to doing things the old way.

This is not an isolated anecdote. It is the pattern behind most failed AI projects, and it is worth understanding why before spending a euro on the next tool.

The figure nobody wants to look at

According to a RAND Corporation analysis, more than 80% of AI projects fail, twice the rate for technology projects that do not involve AI. The problem is not access to technology. It is adoption.

A Deloitte report on the state of AI in business in 2026 finds something similar: although a significant share of companies believe they have an AI-ready strategy, far fewer feel prepared when it comes to infrastructure, data, risk management and talent. Most projects break down in the gap between intention and execution.

A recent survey of executives found that most face serious adoption problems despite their investment. Fewer than a third of companies see a significant return from generative AI, despite notable individual productivity gains. The pattern repeats: a tool gets bought, the results are measured badly, and the conclusion is that AI “wasn’t ready”.

Why this happens

Most consultancies agree on one point: AI is not a technology problem, but a leadership and organisational design problem. A Google vice president put it this way: AI’s capabilities are advancing faster than organisations can absorb them.

Here is what happens in practice. A company has a process that has worked in a particular way for years: who approves what, who reviews whose work, and which steps protect people if something goes wrong. When a tool arrives that speeds up the process, it does not affect only the process. It also affects the accountability structure built around it.

That is why resistance is rarely explicit. Nobody says, “I don’t want to use this because it takes away my power.” They say, “I don’t trust the result” or “I’d rather check it myself, just in case.” Those doubts are not irrational. They reflect the fact that nobody has redesigned the process around the tool: who decides, who answers when it fails, and what happens to accountability. The old structure has simply been left on top of a new tool.

A study by Harvard Business School and the University of Hong Kong makes a similar point: most companies fail to capture real value from AI because people, processes and internal politics get in the way, not because the technology fails. Fear of being replaced, rigid workflows and established power structures explain more failures than any technical limitation of the model.

What is usually needed instead

When an automation project works in the medium term, there is almost always one thing in common: someone took the time to redesign the process instead of simply dropping a tool into the existing one. That means asking what each approval step protects, whether that risk still exists, and who takes responsibility when a system rather than a person produces the result.

Gartner has been pointing out for months that unclear governance is one of the fastest ways to stall responsible AI adoption. When the rules about data use, ownership of information or who answers for an error are ambiguous, people default to saying no. Or worse, they use the tool out of sight and without any oversight.

For a small business, the practical conclusion is easy to state and hard to act on: before automating a process, decide who is responsible for the result, what happens if the system makes a mistake, and which parts of the old process no longer make sense. No tool can solve that. It takes a conversation that most projects skip.