The Context Moat: Why 95% of AI Projects Fail and What the Winning 5% Know

The strange contradiction in enterprise AI

I sell AI-enabled software to institutional investors and fund managers across the UK and Europe, and I use AI daily in my own work. I sit on both sides of the adoption curve - practitioner and educator. And over the past year, two things have happened that are shaping my approach to both selling and using AI technology on a daily basis.

The first was a piece of research. MIT's study, The GenAI Divide: State of AI in Business 2025, analysed over 300 enterprise AI deployments. The finding: 95% of pilots delivered zero measurable P&L impact. This against an estimated $30-40 billion of enterprise AI spending. Only 5% of integrated systems created significant value.

The second was hearing it confirmed from the market itself. Clients and prospects began coming back to us with a consistent message: the generic AI systems they had deployed were not working. Accuracy was too low. Hallucination rates were too high. They needed more human control in the process. Several firms I have spoken to went further - they wound back their AI implementations entirely and returned to manual processes. An expensive round trip to end up exactly where they started.

So what separates the 5% from the 95%? I recently watched a keynote by an old friend, who leads AI strategy at a global retail technology firm, and he put language to something I had been observing in every sales conversation for the past year. His argument was simple:

AI models are commodities. Anyone can rent them. The competitive advantage is the context - the accumulated, structured, governed intelligence that makes those models useful. He calls it the context moat.

In practical terms, a context moat is the body of trusted knowledge that sits behind an AI system: approved content, historical decisions, institutional expertise, governance processes and structured data. The stronger that foundation, the more accurate, reliable and defensible the AI's outputs become. 

The market data backs this up. Venture
capital is not pouring billions into better models - it is flowing into the infrastructure and context layers around them. Data platforms, vector databases, protocols like MCP that connect AI agents to enterprise knowledge. The smart money has already decided where the value sits. It is not the model. It is what the model knows about your business.

What I see in the market

Every conversation I have with investment firms eventually arrives at the same place: AI architecture. But the firms making real progress have understood something the others have not - that underlying data which is audited, approved and accurate is essential for these systems to function at all.

The most common mistake I see is firms jumping into AI without proper due diligence on how these systems feed into their overall architecture. They buy the application layer and skip the intelligence layer. It is my miscalibrated power meter at enterprise scale: impressive technology producing confidently wrong outputs, because nobody verified the data underneath.

There is also a notable shift happening on the investor side, and fund managers should pay close attention to it. LPs are now asking managers directly: do you use AI to respond to our due diligence questionnaires, and what controls do you have in place to ensure those responses are accurate? AI governance has moved from an internal IT question to an investor-facing diligence question. Regulators are moving in the same direction - hallucination risk now features prominently in supervisory guidance on both sides of the Atlantic.

In other words, the question is no longer whether you use AI. It is whether you can evidence the controls, the approvals and the audit trail behind it.