Most AI projects don't fail because of a bad model, but because of a missing path to production. Companies underestimate what it takes in terms of data quality, governance and internal ownership to turn a working prototype into a reliable system.

Based on interviews with 65 experienced data scientists and engineers, RAND concluded in 2024 that more than 80% of AI projects fail. That's not an incident. That's a pattern.

"The 95% failure rate for enterprise AI solutions represents the clearest manifestation of the GenAI Divide", according to the MIT report. What stands out: companies that buy AI tools and partner with specialized vendors have a 67% success rate according to the same research, versus 33% for building everything in-house. Wanting to build it all yourself sounds ambitious. In practice, it's mostly expensive and slow.

A CEO who wants to scale AI successfully doesn't need extra enthusiasm, but a concrete business case per use case, an owner for that project and a team that can go beyond the demo phase. That sounds boring. It works.

McKinsey figures show that only 39% of organizations see measurable bottom-line impact from AI investments. That's not a technology problem. That's an execution problem.

In the Netherlands that percentage is even lower. Research puts successful AI projects at 21%, the lowest of six European countries studied, with 38% internal resistance as a striking factor. Resistance is not a technical problem. It's an organizational problem, and you can't prompt your way out of it.

This is exactly where the demand for tech talent is shifting. No longer AI project management as a loose buzzword, but people who understand how to guide a pilot through governance, security and integration. We see that demand with clients who recently needed a security specialist with a sharp eye for compliance for exactly these kinds of projects.

The layoffs in the tech sector are creating a larger candidate pool, but not automatically the right people for AI implementation. Many of those candidates have experience with building. Companies are now mainly looking for people who can scale and operate.

That's an important difference. A developer who is good at building a prototype is not automatically good at making it production-ready. That requires different skills: data engineering, MLOps, process automation, and the ability to translate a pilot into something an organization can trust.

We notice this in practice. Vacancies for a developer who also understands infrastructure and deployment attract more attention than a year ago. Not because the role has changed, but because companies are finally realizing that building and scaling are two different skills.

Not every former tech employee fits here. That's exactly why screening for AI projects now needs to be more precise than ever.

Companies are no longer looking for people who can experiment with AI. They're looking for people who can land a pilot in an existing system, with the right checks on security, cost and reliability.

IDC research on AI agents shows that 88% of proofs-of-concept never reach production. That number says something about technology, but even more about the people who have to guide it.

The roles that are scarce right now are not the classic AI project lead who builds a sprint planning. They're engineers who understand both the technology and the organization. Someone who builds a pilot, but also knows which questions a CISO will ask before it's allowed to go live.

At a client managing a sensitive environment, the recent need wasn't for an AI specialist, but for an SRE who can guarantee stability under increasing automation. That's exactly the shift: from experimenting to running at production level.

In practice, this means that a job description asking only for "experience with AI tools" attracts the wrong people. Companies that get this right explicitly ask for experience with scaling, not for enthusiasm about new models.

We saw this at a scale-up that was stuck on an AI pilot for two quarters without a clear owner. Not a technical problem. Nobody had the role to push the project through to production.

When the vacancy was rewritten, with a focus on integration experience and collaboration with compliance teams, a shortlist of candidates who understood the difference between a demo and a system a company can build on came together within weeks.

That's exactly where a structured recruitment process makes the difference. Not sending around CVs based on a trending buzzword, but an intake that figures out what a role really requires, before a single profile is proposed.

AI pilots don't fail because the technology falls short. They fail because companies underestimate what it takes to turn an experiment into a reliable system.

That's no reason to stop with AI. It's a reason to recruit differently.

No longer based on enthusiasm. Based on proven ability to make something actually run.

At Doghouse we see that shift every day in the vacancies that come in. Want to know how to sharpen a role that fits what your organization needs now, instead of what was trending two years ago? Let's just have a chat about it.