Faced with the rapid commoditization of AI, the only sustainable competitive advantage is the context graph: the historical memory that explains the “why” of decisions. Without it, the AI fails.
The business industry is witnessing a rapid succession of blockbuster AI trends. First, the great language models – the promise was clear: “to have the best model is to own the future”. Then came fine-tuning, then recovery-augmented generation (RAG), and more recently agentic frameworks. Ironically, all of these cycles follow the same pattern: an ingenious technical breakthrough is followed by frenzied corporate adoption, before rapid commoditization that makes it accessible to everyone.
We are currently experiencing this “agentic moment”. Boards of directors are flocking to agentic AI investments and vendors are rebranding their offerings and launching white-label solutions overnight. IBM’s 2026 technology forecasts clearly confirm this: competition is no longer based on AI models – which are “approaching commodity status” – but on the systems that surround them. This pattern is consistent enough for a serious technology leader to stop and ask: If every layer becomes commoditized at this rate, what, if anything, actually remains a sustainable and defensible advantage?
The answer is the context graph. And, no, most companies don’t have one, because it is neither a database nor a knowledge base that can be commoditized.
What businesses know – and what they don’t know
Most companies know what happened with a customer. Very few know why. An exceptional 20% discount granted five years ago leaves a trace in the CRM: the final figure. But the strategic relationship, the arbitration of the regional director, the negotiation which justified it appear nowhere. A support ticket that technically does not meet the emergency criteria is nevertheless escalated; the system records the resolution, never the judgment that triggered it.
The thinking that really runs an organization lives elsewhere: in phone or written conversations with the customer, in internal discussion threads on platforms like Slack, in email chains, and in the institutional memory of a regional sales manager. This invisible layer is what we call the “context graph” — the mapping of decisions and their reasons. Transactions, yes, but also the arbitrages that underlie them. And it’s the one thing in enterprise AI that can’t be quickly rebuilt, commoditized, or delivered via a software update.
Why does this change everything?
Foundation Capital, in an analysis published in December 2025, believes that the next trillion-dollar ecosystem will not be built by piling AI on existing data, but by capturing that “why” behind every decision – turning the exceptions and context trapped in casual conversations into a searchable asset for the business. This is something fundamentally different to build.
What makes the context graph structurally different from all the other layers that have become commonplace is time. An LLM can be replicated by a well-funded team and an agentic framework can be open-sourced. A context graph, however, is built on years of decision traces – exceptions, cancellations and precedents – that cannot be bought or reconstructed quickly.
This is precisely what creates a structural ceiling for start-ups that offer “context on demand”. Manually building context, use case after use case, is a great feat of execution, but it remains a service activity. Without being natively present in the execution path to capture the decision at the moment it is made, this context remains fragile and non-transferable. Evolving such a model does not create a cumulative feedback loop; it just increases the maintenance cost.
Compare that to a platform that has been managing customer journeys at scale for over a decade, processing millions of customer and internal employee conversations every day, across social/digital/voice channels, in multiple languages, and across industry-specific ontologies. The context graph that exists here is both broad and specific. A successful AI must know that the word “Apple” in a conversation with a tech player does not have the same meaning as in an exchange with a food distributor. She must understand that a “container” for a logistician has nothing to do with a “container” for a software engineer. She must be able to explain why a commercial discount was granted to one customer and refused to another with an almost identical profile. This understanding cannot be improvised: it is the result of years of learning accumulated in architecture itself.
The figures that should alert management
The first signals are already visible. Deloitte reports that 47% of enterprise AI users have made at least one major strategic decision in 2024 based on mind-blowing content. S&P Global reports that 42% of companies have abandoned the majority of their AI projects in 2025, up from 17% the year before. This failure is not a failure of the models – they have improved considerably. This is a failure of context. We ask AI to make decisions with serious consequences without providing them with the history, business specificity or institutional memory that would make them reliable.
The right question to ask
Companies that succeed in their AI transformation ask a different question during their assessments. Not “what model are you using?”, nor “how does your agentic layer work?” — these answers will be obsolete in 18 months. The real question is: “Show me your context graph. Show me what you know about this customer across each touchpoint. Show me the decision traces that explain why the last exception was granted.” If the response is a blank stare, the architectural problem will not be resolved by any model updates.
The race for AI is real, and its consequences are just as real. But the companies that define the next decade’s enterprise AI won’t be the ones with the best models. They will be those who have spent the last ten years building the context these models need to be right. This gap is not closing. It’s hollowing out.
