The Context Flywheel: Why AI Agents Fail Without Institutional Memory

Prukalpa Sankar of Atlan argues that the main barrier to AI adoption is not model intelligence, but a lack of contextual intelligence. While companies focus on the raw power of LLMs, they overlook the institutional knowledge--semantics, processes, and organizational nuance--that allows human employees to function. The result is a cycle of testing and abandoned AI pilots. To gain a competitive advantage, organizations must stop treating AI as a standalone tool and start building a context flywheel that allows agents to learn, adapt, and operate with the depth of a veteran employee. This requires moving from human-in-the-loop governance to human-on-the-loop oversight, creating a systemic advantage for teams that prioritize durable, portable context over quick-fix automation.

The Hidden Cost of Smart Interns

Most organizations treat AI agents like highly capable interns: they are brilliant but lack the institutional memory required to be useful. Sankar notes that 56% of CEOs report zero financial benefit from AI, largely because these agents are built from scratch without access to the company internal logic.

Companies in the real world care about performance. Performance is a function of I like to think intelligence, but also context. If you think about human job performance, less than 10% of human job performances explained by cognitive intelligence.

-- Prukalpa Sankar

The immediate solution is to deploy agents and hope they learn on the job. However, the downstream effect is a massive accumulation of technical debt and hallucination-prone systems. By failing to provide explicit context, such as business definitions and operational workflows, teams end up in a cycle of abandonment where 90% of agents are scrapped after a month because they lack the trust required for production.

Why the Obvious Fix Makes Things Worse

Conventional wisdom suggests that data catalogs, business glossaries, and master data management are the solutions to this fragmentation. But as Sankar points out, these projects historically fail because the effort required to populate them is too high for humans to maintain.

The non-obvious shift is using the AI itself to build the context layer. Instead of humans manually documenting columns, modern systems can use AI to reverse-construct business logic by analyzing query logs, BI tools, and data lineage. This creates a feedback loop: high-quality context leads to better agent performance, which generates more accurate traces, further improving the context.

Today we are at a point where our customers tell us, 89% of our customers say that it is as good or better than humans from the accuracy perspective. This was a tipping point we got to this tipping point this quarter.

-- Prukalpa Sankar

The competitive advantage goes to companies that treat context as a living, portable asset--a company brain--rather than a static documentation project.

The 18-Month Payoff: From Human-in-the-Loop to Human-on-the-Loop

The most critical change is the shift in governance. In the early stages of AI adoption, humans are involved in every decision. But as the system matures, the volume of context generation becomes too vast for manual review. Sankar argues that the goal is to move to human-on-the-loop, where AI does the heavy lifting and humans act as governors, approving or rejecting systemic changes like metric definitions rather than performing the grunt work of data entry.

This requires a change in how we view the data practitioner. They are no longer just data people; they are context engineers. By building a context repo--a version-controlled, portable unit of institutional knowledge--teams can ensure that when their business logic changes, the impact propagates correctly through all downstream agents. This prevents the heterogeneity trap, where different agents in the same company begin speaking different versions of the truth.

Key Action Items

• Audit your Context Debt: Identify where your AI agents are failing due to lack of institutional knowledge, such as misinterpreting new customer definitions. Immediate priority.
• Implement AI-First Context Generation: Stop manual documentation. Use AI to scan existing query logs and BI metadata to reverse-engineer your business semantics. Over the next quarter.
• Establish a Context Repo: Treat your organizational context like software code. Create a repository that acts as the single, portable source of truth for your agents, version-controlled and auditable. Over the next 3-6 months.
• Shift to Human-on-the-Loop Governance: Move your senior analysts away from manual data maintenance and toward governing the AI decision-making logic. This creates long-term leverage.
• Prioritize Interoperability: Avoid vendor lock-in for your context layer. Ensure your company brain is open and accessible, allowing you to swap model ecosystems as the technology evolves. This pays off in 12-18 months.
• Build for Change: Accept that your AI stack will be obsolete in six months. Invest in modular, open architectures that allow you to rebuild and change along with times. Ongoing.