How an Open Context Layers Help Enterprises Build, Govern, & Scale Agentic AI - with Prukalpa Sankar of Atlan

Sankar identifies three critical context gaps preventing AI from reaching production: inability to find data (discovery), lack of business meaning (LLMs don't understand enterprise-specific definitions like 'TAM'), and governance challenges (controlling what data AI agents can access). She explains how $250B spent on data/analytics had 75% failure rates, and AI is worse at 95% failure because the context problem has intensified.

• •First context gap: Data discovery - CIOs have 1000 AI projects but don't know where their data is located
• •Second gap: Business meaning - LLMs understand language but not enterprise-specific definitions (e.g., 'TAM' means total addressable market in your company, not generic definitions)
• •Third gap: Governance - Need to control which AI agents can access specific data (e.g., HR chatbot using payroll data shouldn't expose CEO salary)
• •New complexity: AI sprawl creates additional governance needs around AI agents, applications, lineage, and regulations like EU AI Act

Sankar differentiates context layers from traditional semantic layers and knowledge graphs. Using a real customer example, she demonstrates how an AI data analyst failed because it couldn't interpret 'top 10 customers'—sales teams meant by revenue, success teams meant by NPS ratings. Context layers must be dynamic, continuously updated through feedback loops rather than built as static three-year governance programs.

• •Context layers include data context, business meaning, domain context, and process context—not just semantic definitions
• •Real example: 'Top 10 customers' means different things to sales (revenue) vs customer success (NPS/ratings) teams
• •Dynamic feedback loops: AI asks users to clarify definitions, then codifies those choices back into the context layer for future use
• •Every human interaction should enrich the context layer to help AI act more autonomously over time
• •Traditional governance programs that take 3 years to implement are incompatible with AI deployment timelines

Unlike traditional IT projects that only need business sponsorship, AI requires active business involvement as 'context engineers.' Sankar explains this fundamental shift in operating models and roles, emphasizing that domain expertise resides with business teams, making their participation essential for AI success. This represents an existential opportunity for technical leaders to expand their influence.

• •AI cannot reach production without business teams actively involved in context engineering—sponsorship alone is insufficient
• •Domain expertise lives with business teams, requiring collaboration between technical and business stakeholders
• •This changes fundamental operating models and creates opportunity for technical leaders to take larger strategic roles
• •Business users demand ChatGPT-like experiences on enterprise data, raising the bar for accuracy and usability
• •Change management is critical—people hold AI to higher standards than they hold humans

Sankar frames AI adoption as existential, noting that average S&P 500 company tenure dropped from 75 years (25 years ago) to 15 years today, and may drop to 3 years in the AI era. She positions this as a super-cycle opportunity where leaders will either become transformation heroes or lose their jobs within 3-4 years.

• •S&P 500 company average tenure: 75 years (1990s) → 15 years (2010s) → potentially 3 years (AI era)
• •Foundational business models face existential threats from rapid innovation cycles
• •Customer interaction with technology, websites, and digital storefronts is changing dramatically
• •This is a super-cycle opportunity for great companies; others will be left behind
• •Technical and business leaders who solve this will be heroes; those who don't may not have jobs in 3-4 years

Sankar shares concrete results from enterprises building context layers. Workday improved AI accuracy by 5x by moving from 'perfectly governed BI data' to context-driven approaches, onboarded 6,000 employees, and achieved over 1 million platform views. Mastercard's CDO built an agentic governance framework with over one-third of products now AI-first.

• •Workday: 5x accuracy improvement when moving from traditional BI governance to context layer approach
• •Workday: Onboarded 6,000+ employees with 1M+ platform views through data/context products
• •Mastercard: Over 1/3 of products are now AI-first, supported by agentic governance framework
• •Mastercard CDO: 'We're not a finance company, we're a data company'—reflects fundamental transformation
• •Context layers drive measurable improvements in usability, adoption, and ROI across use cases

Sankar introduces 'context readiness' as the prerequisite for AI readiness, sharing a blueprint operating model. The approach: start with business use cases driving measurable ROI, build reusable context foundations around them, and involve business as context engineers throughout. This avoids the failed approach of building complex graphs before production.

• •Context readiness must come before AI readiness—context binds data (technical) and business together
• •Start with specific AI use cases that drive measurable business ROI, not multi-year graph-building projects
• •Build context foundations that are reusable and reproducible across multiple use cases
• •Embed business teams as context engineers in the development process, not as an afterthought
• •Make context engineering part of the AI use case lifecycle to ensure it's actionable and funded

Sankar details the technical implementation framework: create evaluation benchmarks from existing BI queries and Slack questions, bootstrap with verified SQL queries, implement human-in-the-loop verification, and maintain always-on enrichment. Customers achieve 10x accuracy improvements within one month using this approach.

• •Build evals framework first: Pull top questions from BI tools and Slack to understand what business users actually ask
• •Create golden verified SQL queries from existing BI systems to establish accuracy benchmarks
• •Bootstrap process: Present LLM outputs alongside verified answers, have humans validate and improve
• •Customers achieve 10x accuracy improvement within one month of starting this process
• •Context enrichment is always-on, not one-off—business context continuously changes and must be updated
• •Metadata architecture must be iceberg-native, open, and ready for real-time workloads