Prukalpa Sankar is the co-founder and co-chief executive of Atlan, where she is building the enterprise context layer that helps AI agents understand how businesses actually operate. Her premise is that powerful models cannot make reliable decisions without access to an organization’s definitions, expertise, institutional memory, and operating norms.
Sankar studied engineering and entrepreneurship at Nanyang Technological University before co-founding SocialCops with Varun Banka in 2013. The company tackled public-interest data problems, contributing to India’s National Data Platform and tools tracking the United Nations’ Sustainable Development Goals.
That experience exposed the organizational confusion created by fragmented data. When figures on a government dashboard unexpectedly doubled, resolving the discrepancy required reconstructing the technical systems and human decisions behind the numbers. Sankar and Banka subsequently built Atlan to help enterprise teams discover, interpret, govern, and collaborate around their data; Sankar recounts that transition in her history of Atlan’s origins.
By 2022, Atlan was developing active metadata infrastructure that connected information across enterprise data systems. Generative AI sharpened the same challenge: agents needed the business knowledge that experienced employees accumulate on the job. Sankar also created Context and Chaos, a publication exploring knowledge graphs, enterprise semantics, governance, and AI infrastructure.
What dependable agents actually need
- Situated knowledge makes model intelligence useful. An agent investigating slower restaurant drive-through service must understand how the company defines the metric, which reporting period applies, whether seasonal patterns matter, and whether recent product changes offer a plausible explanation. Accurate analysis requires business facts, diagnostic judgment, and organizational norms.
- A shared company brain prevents agent silos. Atlan initially built specialized agents around customer-experience tasks, but disconnected memories produced stale messaging and inconsistent assumptions. Sankar advocates a shared company brain containing business definitions, data relationships, organizational entities, and reusable domain skills accessible across agent platforms through interfaces including Model Context Protocol, SQL, and vector search.
- Context requires software-grade operational discipline. Changes to competitive intelligence can cascade into positioning and sales materials. Sankar’s approach to context lifecycle management incorporates ownership, versioning, dependency tracking, security, governance, and protection against skill drift; her account of AI accountability assigns data leaders an important stewardship role.
- Agent interactions should improve institutional knowledge. Execution traces can expose recurring mistakes and suggest improvements that human maintainers review. Connecting business applications and data warehouses also helps reconstruct otherwise fragmented organizational context.
For Sankar, context is organizational intellectual property: when competitors use comparable models, their advantage depends on encoding the knowledge, practices, and judgment that distinguish their businesses.
