A number that doesn't appear in any AI presentation, but should: 70 to 80 percent of all AI projects fail before they reach production.

This is not a rumor. McKinsey, Gartner and IDC arrive at similar figures in independent studies. Much of the billions that companies invest in AI ends up in projects that never make the step into real operations.

The question is not whether this happens, but why. And what the 20–30% who succeed do differently.

Mistake 1: The technology takes center stage, not the problem

The most classic pattern: a company sees an AI demo, is impressed and starts a project. The technology is the starting point, not a concrete business problem.

The result: a system that works technically, but that nobody needs.

What works instead: Start with a clearly defined problem that costs money or time today. Not: "We need an AI agent." But: "Preparing our quotes takes 3 days and holds up sales. How can we get that down to 4 hours?"

Mistake 2: Too big, too slow, too expensive

Many companies launch AI as a transformation program with an 18-month roadmap, four departments and a seven-figure budget. After 12 months, there is an architecture presentation, but no AI in operation.

In the meantime, the market has moved on. The original assumptions are outdated. The project gets reprioritized, or shut down.

What works instead: Get a single, small, measurable use case into production in 4–6 weeks. Only scale once it is running and showing ROI.

Mistake 3: Poor data quality is underestimated

AI systems need clean, structured, accessible data. In reality, company data is scattered across legacy systems, Excel files, email inboxes and isolated solutions.

Cleaning up this data takes more time than the actual AI project, and it is regularly left out of the planning.

What works instead: Before the AI project: a data audit. Where is the relevant data? In what quality? How accessible is it? Who is responsible? You can clarify this in 2 weeks, and save months of overruns.

Mistake 4: The team is not brought along

AI projects often don't fail because of the technology. They fail because of people who don't use the system. Because they don't understand it, don't trust it, or because it makes their work harder instead of easier.

Adoption is not a downstream change management topic. It is a design principle.

What works instead: Involve end users early. Not as testers at the end, but as co-designers from the start. AI that takes work off employees' plates gets used. AI that is forced on them does not.

Mistake 5: No clear success criteria

"We want to use AI" is not a success criterion. Without measurable goals (cost reduction in euros, time saved in hours, conversion improvement in percent), no project can be judged a success or a failure.

This leads to something many companies are familiar with: the project "went well, somehow," but nobody knows whether it was worth it.

What works instead: Before you start, define three metrics that are measured after 30, 60 and 90 days. If the numbers are off, adjust early, not after 18 months.

What the 20–30% do differently

Companies whose AI projects successfully make it into production have three things in common:

  • They start small. No transformation project. One use case, one team, one measurable goal.
  • They think in outcomes, not in technology. The question is not "What can this AI do?" but "What gets better in our business once this is running?"
  • They have someone who knows what they're doing. Process understanding, change management and experience with AI projects that failed: that is the real competitive advantage.
"AI works. It delivers real ROI. But only if the fundamentals are right, and the mistakes that cost 70–80% of projects are known and avoidable."