Why Context Is Capital In Enterprise AI
Anand “Andy” Logani is executive vice president and chief AI officer at EXL, a global data and AI company.
gettyEvery week, we are inundated with headlines about the newest, most powerful large language model (LLM). Does that mean everything you built based on last week’s or last year’s LLM is suddenly obsolete?
It should not. Models are getting more capable and cheaper every quarter, which is exactly why access to the “best” model is no longer an advantage. The more important business question is, what does your AI spend leave behind when the next one arrives? For many enterprises, the answer is “not enough.” McKinsey estimates (registration required) that just 37% of businesses can attribute any enterprise-level financial impact to their use of AI, and Gartner projects that more than 40% of agentic AI initiatives will be canceled by the end of 2027.
The symptoms vary, but the pattern is the same: The AI model works well in the pilot, but struggles when applied to complex, real-world scenarios.
The enterprise AI systems that improve over time are not necessarily using better models than everyone else. They are capturing what their business learns every time a decision is made. Put simply, the smartest enterprise AI systems treat LLMs as a rental, temporary and depreciating, while the context that informs them becomes an appreciating enterprise asset.
I call it context capital, the AI era’s counterpart to the software estates of the client-server era, the digital channels of the internet era and the data estates of the cloud era. For those able to harness it, context capital is the asset that outlasts every model it serves.
Business leaders do not need advanced degrees in computer science to understand the value of context capital. It makes their business their business—the decision data, company-specific ontologies and unique approaches to problem solving that define the way their operations run. Many AI integration projects fail because a foundational LLM was never designed to address those company-specific protocols or incorporate the critical feedback loops that allow teams to learn from their cumulative experiences.
To put this in a real-world context, consider the challenges insurance and financial services business leaders have faced trying to integrate AI into underwriting and compliance workflows. While, on the surface, these repetitive, process-heavy functions should be the natural fits for AI-powered automation, many initiatives stumble when they encounter missing data points or edge-case decisions that require a level of institutional knowledge and finesse to solve. An LLM does not know which claims specialist can locate a missing accident report, or which compliance expert has the final word on a KYC anomaly. These critical pieces of domain-specific context are what make a company’s business processes function, and without that level of nuance built into the feedback loop, an AI workflow stalls exactly where the value is.
Context capital demands an architecture of its own. It cannot be an afterthought bolted onto a data platform because the work of forming the asset and the work of using it follow different pathways.
The context build path is where the capital is formed. It draws selectively on enterprise metadata, business definitions and relevant operational signals to build and maintain a context graph: ontologies that define what things mean, resolved entities that establish identity across systems, relationships that connect them and decision traces that record what was decided. The graph organizes selected context while preserving links to underlying data in source systems.
The context serving path is where that capital is put to work. At decision time, it assembles a working set for each agent, user and task: the relevant slice of meaning, facts, rules, rights and memory, layered from enterprise truth down to the case at hand.
It draws on the context graph and accesses additional source-system data as needed, based on the task’s requirements for latency and freshness.
These two paths converge when company-specific data and accumulated enterprise insight are available across the entire organization to support complex decision making and operational processes. For example, in a wealth-management context, one firm I work with developed a context graph across wealth-management data that gives agents a shared understanding of clients, products and opportunities, with decision-histories built into the loop.
The goal is to accumulate institutional knowledge and deep domain expertise in a form that any AI model can harness, including models not yet developed.
Every enterprise running AI is already accumulating context capital. The question is whether it’s being built deliberately or accumulating by accident. To get to that point, business leaders need to ask themselves three key questions before they start shopping for the next LLM:
Trace one high-value decision end to end and list every system that held part of its context. Then note how much of that record survived the decision.
For every AI tool in production, establish whether the context it gathers is contractually your asset or someone else’s telemetry.
If a model or platform were replaced tomorrow, ask what learning would carry forward and what would restart cold.
If the answers to these questions come back scattered, start there. Your business already generates the context with every decision it makes. Owning that context is what allows every new model to start smarter than the one it replaces.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
