The first episode of WTF Is the Context Layer? opened with a poll: is the context layer just the semantic layer rebranded? A knowledge graph with better PR? Something entirely new? The majority of respondents voted for "the glue between data and AI," but it was clear that our core question resonated: WTF is the context layer anyway?
Atlan Co-founder and Co-CEO Prukalpa Sankar joined me in our inaugural episode to unpack the contextual intelligence equation, the biggest context barriers, and — a question I get constantly — who actually owns the context layer. She acknowledged what we're all living through: the tech is changing so fast, that what was true three or four months ago is already changing shape.
But by answering your burning questions, we can all keep a finger on the pulse of the skills, knowledge, and tools needed to stay a step ahead. Here's a roundup of our favorite Q&As from episode one.
PS — Have a question for our next episode? Submit it here.
Q: Intelligence is getting better every year. Why isn't AI performance improving at the same rate?
You don't need to be an AI expert to see how drastically models have evolved, even in the past year. Gemini 3.1 Pro scored 94% on PhD-level science reasoning tests. Claude 4.6 fixed context-window processing issues that existed in 4.5. A GPT-5.4 Pro user with no advanced training solved a 60-year-old math problem.
Still, 56% of CEOs report zero financial benefit from AI, and only one in five projects creates significant business value. Something is missing.
Prukalpa presented a framework to explain it: Performance = Intelligence × Context. Intelligence is the variable that's accounted for. Context is the gap.
"Cognitive intelligence explains 10% of job performance variance. Would you say your best employee is the one who scored highest on the SATs? Or the one who learns fastest and takes feedback and works the hardest?" Prukalpa asked. "You ask any human this, and it's a very obvious answer as to what it takes to create the best teammate. That's what we believe is the missing layer."
Context breaks down into three components: knowledge (what an agent knows about your business); skills (how it applies that knowledge to real decisions); and tools (what it can act on). Without all three, better models won't deliver better performance.
In fact, Prukalpa warned, it could be the opposite. "I actually think the worst and the most dangerous is high intelligence and low context systems," she said, citing an example of a customer support agent trained to maximize NPS. It learned, on its own and with no human approval, that refunds improved scores.
"What you need is a high intelligence, high context system. And that's really what most performance systems will start looking at."
Q: Where are enterprises getting stuck with context in practice?
Prukalpa acknowledged that the challenge she hears most often from teams starting an agentic journey is how to even begin encoding everything that comes along with context: what it is, who owns it, where it sits, and what the operating model should look like. But once those questions are answered, it's not always smooth sailing.
Three failure modes routinely show up in practice, and they're often hard for teams to foresee until they're in the thick of a deployment.
Context bootstrapping. Building an agent takes five minutes. Giving it enough business context to be trusted takes months. One customer Prukalpa referenced launched 1,000 Databricks Genie rooms across their organization, only to abandon 90% within a month. The technology worked, but users couldn't get 70–80% accuracy on basic questions, so they simply stopped relying on it.
Prukalpa explained how a seemingly simple query — what are my top 10 new customers? — could easily go sideways. If sales is asking, "top" means revenue. If customer success is asking, "top" means adoption. "New" might mean this quarter or this year. Even a highly intelligent agent can't answer this question reliably without context. And when it guesses wrong, users stop trusting it.
Context management. Once agents are deployed, they each develop their own memory and behavior. That raises questions most organizations aren't ready for: which updates should propagate centrally? Who approves a change to brand voice guidelines so it reaches every downstream agent? What's the governance model for context drift?
Context portability. Organizations are deploying dozens of agentic platforms and hundreds of agents, Prukalpa explained. Their AI systems are now running into the same problem that used to plague sales and finance: two departments have two different revenue numbers, and no one can agree on which to use.
One customer told her that their Snowflake Cortex agents, Google Agent Space, and internal CLI tools were giving contradictory answers about their own data. "Now each of these platforms is speaking a slightly different language, and they don't know how to control that."
Q: My organization already has Confluence, SharePoint, and process docs. Why isn't that enough context for AI?
"If you want to understand how a company thinks that it runs, read its process documents," Prukalpa advised. "If you want to understand how it really runs, go understand its systems."
The distinction may seem nuanced, but it's consequential. AI can extract far richer context from operational systems than from documentation. Prukalpa presented a taxonomy comprising systems of record (ERP, CRM), systems of data (data warehouse), systems of semantics (BI tools, business logic), and systems of engagement (Slack, email).
Each of these systems encodes how the business operates. The problem is they're disconnected. Every time data moves between them, context gets lost. In the old BI world, humans were the connective tissue. In today's agentic world, that connective tissue needs to be explicit and machine-readable.
In practice, AI that reads a system of record, checks BI usage patterns, and understands domain context before writing a description will produce dramatically better output than an analyst writing from memory.
"If you can get AI to actually operate at scale and do this, they're going to generate a very high quality description," Prukalpa explained. "We see this now with our customers, where 89% of our customers say that AI is better than a human."
Q: Knowledge graphs, ontologies, semantic layers — which one do I actually need?
Start with the use case, not the technology.
The terminology gets confusing because vendors map everything to their own stack. Graph database companies see every problem as a graph. Semantic layer vendors see every problem as a metrics definition. The honest answer is that the right technology depends on what you're trying to do.
Conceptually, relationships matter most. AI needs a map of your organization that covers not just definitions, but also the paths between them. Which escalation process applies to this type of customer complaint? Which metric definition is relevant for this persona's question? Without that traversal capability, "AI is effectively just a really, really smart intern," Prukalpa said.
On the technology side, graph databases don't yet scale well across hundreds of millions of assets. Graph engines, which apply relationships without moving data, are an emerging alternative. Ontology standards like RDF are still evolving. Prukalpa's advice: "Figure out how to best identify, solidify, and codify relationships. Don't get caught up in what to call the construct."
What about semantic layers? They solve the meaning-context problem: what does "customer" mean here, how is it calculated in the data layer? But for conversational analytics, that's only one of four layers you need. To truly execute conversational analytics that's reliable, you need to give AI user context (who's asking and why), knowledge context (what "new" means in your business), semantic context (how the term is defined), and data context (how it's calculated).
"There's pre-AI world semantic layers, and now there's also post-AI world," Prukalpa explained. "Today, it's clear that definitions are not necessarily enough."
Q: What are skills, and why do they matter for the context layer?
Skills are the mechanism for encoding tacit knowledge. They're the things your best analysts know from experience, but haven't written down.
Prukalpa used an example of a human analyst running a "what if" analysis who knows to always check for seasonality. At Atlan specifically, Q4 sales data spikes at the end of the quarter, while Q1 customer adoption spikes at the start (because customers sign deals and begin onboarding). A new analyst takes time to learn this, and each new analyst must learn the same lesson one by one. But a skill can encode it immediately, making knowledge available to every agent that runs that analysis type.
The implications are significant. Prukalpa described being able to bootstrap skills by analyzing how high performers already behave, like how they write emails, structure analyses, and handle escalations. That behavior can then be reverse-engineered into a reusable skill. As a result, agentic deployment time goes from months to days.
Skills are also shifting the role of the semantic layer. Six months ago, the recommended path to a conversational analytics deployment was: bootstrap context → build a semantic view → publish downstream to agents. Now, for many use cases, Prukalpa says that path bypasses the rigid semantic model. Teams can simply build a context layer with metrics, an ontology, and skills, and keep moving.
Q: Why is context IP, and how do I protect it from vendor lock-in?
In a world where every company has access to the same models, context is the differentiator. The customer support agent at American Express operates differently from the one at Amazon, not because of the underlying model, but because each encodes a different culture, norms, and decision rules.
"As you start thinking about context from that perspective, it becomes your culture. It becomes your business idea. It's how you do business, and how you do business that's different from someone else," Prukalpa explained. "The more you're able to encode that, the more you're able to start creating amazing products and services for your customers."
But the thing that makes context so important also makes it a target.
Every major AI platform, from cloud data warehouses and agentic frameworks to foundation model providers, is now competing to capture enterprise context. The cost to switch from one model to another is near zero. But the cost to move away from a platform that owns your context is high.
Prukalpa was direct: "You should not trust any vendor, including us. Ask whether your context is open, whether it's interoperable, and what happens if you walk away."
All of Atlan's context is stored in Apache Iceberg, an open file format designed to bring compute to the data rather than move the data to the compute. Context is packaged into portable, version-controlled "context repos" — analogous to GitHub code repos — which convert natively to skills for Claude, push semantic models to Snowflake Cortex or Databricks, and bridge gaps between disconnected platforms like Looker and Gemini Enterprise.
"It essentially just makes this a portable ecosystem downstream. Today, it's MCP and there's a new protocol called A2A, or agent to agent, that's coming up," said Prukalpa. "We believe these concepts are likely to change and we want to keep supporting the open standards that come along in the industry."
Q: Who should own the context layer?
This is the pressing question that most companies are asking, but haven't answered yet. Among forward-thinking organizations, the emerging pattern is to implement a central context platform, with federated context ownership by business domain.
The structure is familiar, following the path of data platforms. Platforms landed in a central function, while the data itself and use-case ownership were distributed to business teams. Prukalpa observed that whoever owns the AI architecture diagram typically assumes responsibility for the context layer. It is usually a joint effort between data and AI functions, depending on how the organization is structured.
Prukalpa shared what that means in practice at Atlan: she is the only person who can update the core brand voice document. When she does, every downstream agent, across the website, outbound campaigns, landing pages, and any other external touchpoint, inherits the update. The exception is for the social media lead, who maintains her own version, scoped to her own use cases, connected to the root.
Context management, like data governance, requires both centralized standards and distributed ownership. That balance is where most enterprises are still figuring out where to draw the line. Looking to the architecture, Prukalpa advises, is the most reliable place to start.
Join us for Episode 2 of WTF is the Context Layer? featuring AI advisor and former Gartner analyst Sanjeev Mohan, and get his expert perspective on how enterprises are navigating the context challenge. Register here.
