The AI Centre of Excellence Trap: Why 85% Fail and How to Build One That Ships

ClarityARC's latest enterprise study dropped a number that should worry every CTO: somewhere between 70 and 85 percent of AI Centres of Excellence fail. Not stumble. Not underperform. Fail.

The pattern is predictable. A company launches its AI CoE with fanfare. Six months later, business units are routing around it. Eighteen months later, it's either disbanded or exists as a PowerPoint factory that nobody consults.

The research, which analyzed May 2026 enterprise restructures, found that most CoEs become the exact bottleneck they were designed to prevent. Teams wait weeks for approvals. Innovation slows to a crawl. The very structure meant to accelerate AI adoption ends up strangling it.

The PowerPoint Factory Problem

According to CIO magazine's analysis, most companies are stuck in what they call "AI pilot purgatory." The CoE publishes guidelines. It creates frameworks. It holds workshops. What it doesn't do is ship.

FinTekCafe's research identified the core issue: most CoEs lack delivery mandates and executive air cover. They can advise. They can recommend. They can't actually build anything.

The result? Business units learn to work around them. Shadow AI projects multiply. Governance becomes a joke. The organization ends up with dozens of disconnected pilots and no path to production.

Why Hub-and-Spoke Beats Everything Else

The research points to a clear winner among organizational models: hub-and-spoke. For organizations with more than 500 employees, this structure consistently outperforms both fully centralized and fully embedded alternatives.

Here's why it works. The hub maintains standards and provides shared services. The spokes, embedded in business units, actually build things. Each spoke has its own budget and delivery targets. The hub can't become a bottleneck because the spokes have autonomy.

ClarityARC's design framework emphasizes a critical detail: budget ownership determines outcomes. When the CoE controls the budget, it ships products. When it doesn't, it publishes guidelines.

The Data Quality Wall

Even well-structured CoEs hit a wall: data quality. Gartner's research estimates poor data quality costs organizations 15 percent of revenue annually. For AI projects, the impact is worse. Sixty percent of AI projects unsupported by AI-ready data will be abandoned through 2026.

Organizations with poor data quality programs see 60 percent higher project failure rates. The math is brutal. If your baseline AI failure rate is 70 percent, bad data pushes it past 90 percent.

This creates a vicious cycle. The CoE can't deliver wins because the data isn't ready. Without wins, it can't justify the investment needed to fix the data. The whole structure collapses.

From 200 Pilots to 5 That Matter

MIT's NANDA Initiative studied over 300 enterprise AI deployments. Their finding: 95 percent of enterprise generative AI pilots deliver zero measurable P&L impact.

The successful 15 percent of CoEs share a ruthless focus on portfolio triage. They don't try to support 200 pilots. They identify the five that could actually move revenue or cut costs, then put all their resources behind those.

One Fortune 500 retailer's CoE started with 180 proposed AI projects. After triage, they focused on three: inventory optimization, customer service automation, and personalized pricing. All three shipped within six months. All three delivered measurable ROI within a year.

The Delivery Mandate

The difference between CoEs that ship and CoEs that drift comes down to structure. Successful ones have four non-negotiable elements:

First, they own budgets. Not influence budgets. Own them. Second, they have delivery targets tied to business metrics, not activity metrics. Third, they report to someone with P&L responsibility, not to IT or innovation. Fourth, they can hire and fire their own teams.

Without these four elements, a CoE is just another corporate initiative. With them, it's a product organization that happens to focus on AI.

The 85 percent failure rate isn't inevitable. It's a choice. Most organizations choose the comfortable path: centralized control, consensus-driven decisions, activity over outcomes. The 15 percent that succeed choose differently. They build CoEs that look more like startups than committees.

The question for builders isn't whether you need a CoE. It's whether you're willing to build one that actually ships.