The theory is elegant. Accumulate enough organizational knowledge; workflows, decisions, exceptions and you build a defensible position. The more context you feed the system, the more irreplaceable it becomes.
It is a comforting thought. It is also wrong.
The Confidence-Correctness Gap
Manuel Koelman captured the zeitgeist well in his recent piece 2026: When AI Hits Reality: “Intelligence without context is a demo. Context without runaway costs is the product.”
The framing is sharp. But it obscures a fundamental architectural flaw.
Context compounds confidence, not correctness.
An LLM with ten years of enterprise data doesn’t become more accurate. It becomes more convincing. The system sounds authoritative because it has absorbed your company’s vocabulary and decision patterns. But the underlying architecture remains unchanged: pattern matching without causal understanding.
This is the moat that floods.
The 2025 Reality Check: High Spend, Low Yield
We don’t have to guess. The numbers from late 2025 are brutal.
According to Menlo Ventures, enterprise GenAI spend hit $37 billion in 2025, a 3.2x increase year-over-year. We are pouring concrete.
But what are we building?
The Abandonment Rate: S&P Global’s 2025 survey reveals that 42% of companies abandoned the majority of their AI initiatives, up from 17% in 2024. Nearly half of all POCs (46%) were scrapped before ever hitting production.
The ROI Void: MIT’s “The GenAI Divide” report is even more damning: 95% of GenAI pilots delivered zero measurable ROI.
Enterprises are accumulating context, deploying systems, and watching them fail.
Why “Smarter” Models Aren’t Fixing It
The industry hope was that “Thinking Models” (Agentic AI) would solve the reliability gap. They haven’t.
Hallucinations Persist: The Vectara Leaderboard (updated Dec 18, 2025) shows that frontier reasoning models; including the latest from OpenAI, Google, and Anthropic, still hallucinate >10% of the time.
Agents are Stalling: Deloitte (Dec 2025) reports that only 11% of enterprises have agentic AI in production. Gartner predicts 40%+ of these projects will be cancelled by 2027.
We are gaining reasoning depth at the cost of factual accuracy. A 10%+ error rate is acceptable for a chatbot; it is disqualifying for autonomous enterprise operations.
Why Context Can’t Fix Architecture
The core issue isn’t data quality. It is architectural.
Large Language Models are correlation engines. They learn statistical associations between tokens. Given enough enterprise context, they become extremely good at predicting what sounds right based on your history.
But they do not model causation. They can’t answer why things happen, only what usually follows what.
When a system fails because it lacks context, you can fix it by adding data. When a system fails because it fundamentally can’t distinguish correlation from causation, no amount of context helps. You have simply empowered the system to hallucinate in your corporate voice.
The Lock-In Problem
Here’s where the moat metaphor breaks down entirely.
If context were truly a moat, switching costs would protect you. But if your context-rich system is confidently wrong in 1 out of 10 complex decisions, what have you locked in?
You haven’t locked in value. You’ve locked in liability.
The reliability ceiling isn’t a feature… it’s a wall. It is why 42% of projects are being abandoned.
The Real Moat: Causal Reliability
The winning position in enterprise AI won’t be who has accumulated the most context. It will be who has built systems capable of genuine reliability.
Reliability requires understanding why things happen.
Correlation: “What usually happens next?”
Causation: “If I take this action, what will be the consequence, and why?”
This requires moving beyond stochastic token prediction to architectures that model causation. Systems that can distinguish between “this pattern matches historical data” and “this action will produce this outcome.”
Context matters. But context without causal understanding is just correlation with extra steps.
The moat isn’t memory. It’s truth.
Kris Ledel Founder & CEO, Sapiexo Inc.
Read original on Substack ↗