---
title: "Enterprise AI is missing a context layer | Context Conference 2026"
url: "https://atlan.com/context-conference/opening-keynote/"
description: "Opening keynote from Context Conference 2026 with Prukalpa Sankar. Watch the recording on demand."
keywords: "Context Conference, Context Conference 2026, session recording, context layer, Atlan"
---

**Enterprise AI is missing a context layer**: opening keynote from Context Conference 2026 (online, Oct 28, 2026), 21 min, in the [Context Conference recordings](https://atlan.com/context-conference/). On demand from Oct 29, 2026.

## About this session

Join Atlan co-founder Prukalpa Sankar to kick off Context Conference with a look at why context is crucial for getting AI right and what the emerging landscape of context platforms looks like.

## Chapters

- [0:00](https://atlan.com/context-conference/opening-keynote/?t=0) AI returns are going to the few
- [3:59](https://atlan.com/context-conference/opening-keynote/?t=239) Multiplayer AI needs a nervous system
- [7:57](https://atlan.com/context-conference/opening-keynote/?t=477) Three walls: accuracy, collaboration, portability
- [9:20](https://atlan.com/context-conference/opening-keynote/?t=560) How to build context for agents, not humans
- [11:28](https://atlan.com/context-conference/opening-keynote/?t=688) A GitHub for context
- [13:02](https://atlan.com/context-conference/opening-keynote/?t=782) Announcing Atlan Trident, shared context workspace
- [15:45](https://atlan.com/context-conference/opening-keynote/?t=945) Why context needs its Iceberg moment
- [18:30](https://atlan.com/context-conference/opening-keynote/?t=1110) Rent the intelligence, own the context

## Speakers

- **Prukalpa Sankar**, Coined \u201Ccontext is king\u201D before it was cool, Atlan.

## Next steps

- [MIT's first ever study on context](https://atlan.com/resources/mit-report-ai-business-value/): New research from MIT Technology Review Insights on what separates the teams getting returns on AI.

## Up next

[Why invest in context?](https://atlan.com/context-conference/why-invest-in-context/) / Executive decision room.

## Transcript

[0:00] **Prukalpa Sankar:** I'm going to start with something controversial. Everybody says today that context is the hottest topic in AI. I actually think it's underhyped. And today I'm going to make my case. Last year this time, in August, MIT released this report, and it made the headlines for about, uh, a few months. I couldn't go anywhere without hearing ninety-five percent of generative AI pilots at companies are failing. The truth, it's a little different than that. Barely a month later, BCG did a study. They asked companies how they're feeling about their AI initiatives. Sixty percent of the companies said that they're seeing minimal gains of AI despite real investment. Tracks. But five percent, 5% are what BCG calls future-built for AI.

[1:07] Over the last three years, this five percent delivered three point six times the shareholder returns of the laggards. AI returns are happening, but they are going to the few. And these leaders, people at these leaders use AI far more heavily. In January this year, OpenAI study showed that top ten percent of companies, people use AI two point six times as much of an average company. They redid the survey in June, and the gap had widened to eight point three x times. It had tripled. They are pulling away. What are these leaders doing differently? Gartner found that the leaders are investing up to four times more in foundations, data, data quality, governance, people, change management. OpenAI found that their people are six times more likely to be using skills or packaged context that teaches AI about your business.

[2:18] And what's holding the others back? We recently commissioned a study with MIT Technology Review, and we found that sixty-eight percent of organizations rank governance, business understanding, or trustworthy outputs as their biggest AI barrier. But is it going to be worth to get context right? Gartner predicts that organizations that prioritize semantics and AI-ready data, they're going to get up to eighty percent higher agentic AI accuracy and up to sixty percent lower cost. A year ago, Varun and I said context is king. This year it feels really surreal. Twenty twenty-six has really been the year of context. Uh, I have a joke, uh, um, which is that if I ha- I had a shot for every time I heard the word context said in twenty twenty-six, I would be perpetually drunk this year.

[3:18] Uh, knowledge graphs, ontologies, context graphs, metric layers, you know, it's everywhere. So this year the question isn't whether context, it's how context. Every single team I work with asks questions like, "Who owns context in our organization? How do we build a context operating model?" Uh, "Local context versus global context?" No one company can answer these questions alone. But maybe as a community, we can do it together. Welcome to Context Conference, the first ever conference that is dedicated to context management as a discipline inside organizations. Every CIO I meet kind of describes the same story. On one hand, they say, uh, "Everyone wants to be my agent tech platform, but I know that I'm gonna have dozens of platforms and hundreds of agents." At the same time, it's kind of worse that each of these agents, they have their own memory and learning loops.

[4:22] Uh, I had a, uh, head of data and AI tell me, "CRR, Writer, Google Agents, Pacemaker, Cortex. I've deployed all of these, and now they all have slightly different versions of the truth." This is the oldest joke in data. We always used to say, "Ask sales and finance for the revenue number and you're gonna get two different numbers." Now, it's just happening at agent speed. Enterprise AI is multi-harness. People use multiple tools, and AI is going to be enabled across all of these tools. It's multi-model. Every company is going to have a token mix. They're going to send certain tokens to the frontier models, certain tokens to the open weight models, and certain tokens to maybe train their own models.

[5:14] And third, multiplayer. Multiplayer is the most underrated aspect of enterprise AI. Today, even the best AI tools that we have, they're built for one person. And so you see one person starting to compound. You start hearing about 100X engineers. But my assistant's memory, it's my memory. Uh, it's not my teammates, even if we're on the same project. This is why individuals are winning and projects and companies aren't. Because companies don't run on individuals, they run on teams. Think about the best team you've ever been on, right? You passed work back to each other, you riffed off each other, you shared a sense of what good looks like, you learned together. It's a dream team, and every great company runs on them, and we have not yet been able to crack this for AI.

[6:14] So what does a multi-harness, multi-model, multiplayer future need? We believe it needs a nervous system. A nervous system that connects three very important nodes. The first, actors. Actors inside organizations do the work. This is your people and now agents working together. The second, artifacts. This is the context that your actors depend on. This could be skills, tools, metrics, ontologies. In the AI world, artifacts are production dependencies, and they need to be managed almost like software was managed, production software was managed. The third node, activity. Activity is what actually happened. Every conversation today with an AI agent, every action of an agent, uh, it's leaving a trace, it's leaving output. We have access to human feedback, whether someone said, "That's right," or, "Fix it." And without activity connected to actors, you can never improve the quality of AI.

[7:32] For example, the actual outputs of a debt creation skill. How does the owner of an agent improve the quality of a debt creation skill unless they understand how it's actually used? This is what we call the central nervous systems, connecting these three nodes across every harness, every team, every model. A new type of system of record. Every time I share this, uh, the room typically nods. You know, someone, you know, says, "Yeah, that makes sense." Uh, and then someone asks me, "I want this. How do I start?" And when they try, every company, they hit the same three problems. The first, accuracy. Can I trust it? A Fortune 100 bank, they launched hun- thousands of analytics agents in one quarter, and within a few months, 90% had been abandoned because business didn't trust it.

[8:32] They said it was easier to build an agent than to give it context that made it trustworthy. Then, as you go a little further along, collaboration starts hitting. I had a customer who said, "Marketing changed positioning last quarter, but our sales development agent is still pitching the old version, and nobody even knows what else depends on it." And then third, portability. Portability hits when you start adding second platforms or be- better models come out. The CIO of a global automaker told me, "I don't want my business IP locked into one platform." So how do you go about solving for this? Three principles. First, context artifacts that are purpose-built for agent retrieval, not for human consumption, but agent retrieval.

[9:37] This is a non-trivial problem. Studies say that agents don't really retrieve the way you would expect them to retrieve context. In one study, agents were asked, uh, across 18 setups to just open the skill that it was given, just open the skill that it was given. The range, 46% to 99%. In another study, an agent was given exactly the right skills that it needed, and we tested, does it load every skill it needs? 49% of the time, it did. The most interesting part, more context made it actually worse. When you added more skills, like a few distracting skills, we went down to 31%, and the agent task solution rate went down from 51% to 44%.

[10:36] And does the format of the context matter? Yes, it does. Vercel did a study where it took context and give it, give it in the format of a skill. The agent retrieved it 53% of the time. Then it translated it into an index. It went up to 100%. The same context, just formatted differently. This is why context artifacts need to be built for agent retrieval. This is our biggest learning. Multiple types of artifacts, skills, certified metrics, ontologies, tools, agent operating procedures. They're all needed for different kinds of use cases and different kinds of consumption patterns. Uh, how do you treat and maintain these context artifacts, like code, for agent retrieval? That's the first principle. Second principle, managing context like code.

[11:31] We already established that co- context now is a production dependency. Today, we need to manage it the way we do code. We need a GitHub for context, collaborative authoring, versioning, dependency management, testing, evals. Uh, all of these are code changes that we need to be able to do to manage our context across our organization, solving collaboration. Third, context is a compounding asset. Today, traces are the most amazing source of context about your organization. How do you harness them to turn them into a compounding asset? For example, let's say an end user asks a question, "What is our Q3 net recurring revenue?" And the agent answers something, and the human says, "I actually wanted it by segment." That's context.

[12:29] That's feedback. Ideally, that goes back to the owner of that metric. "Hey, users keep asking for a segment breakdown, but our metric doesn't really have a segment dimension." They diagnose it. They fix it. Now, it ships to everyone inside the organization. That becomes the flow for building context as a compounding asset. To be able to solve for this, you need to be able to build an experience that is purpose-built for agents. And so today, I am very, very excited to announce Atlan Trident, a shared context workspace built, purpose-built for your humans and agents. I'm going to play a quick video to showcase what Atlan Trident is all about. I saw some jokes about the video going too fast, um, for agents that take screenshots.

[15:31] Um, we will later today, fear not, have humans demo parts of this and showcase how you can actually start making agentic workloads inside your organization go live. The third wall is about portability. Context needs an open ecosystem, and no one company alone, including us, can build that open ecosystem. The good news, the open ecosystem is starting to form. Just in the last year, we had Anthropic open up skills, OpenAI come out... And others come out with plugins. Google opened up open knowledge format. Uh, Apache Orc, uh, is becoming a semantic interchange. The open ecosystem is starting to form. But one layer is still missing. We have learnings about this from the data world. Uh, in the data world, we had these open formats for data, Parquet, ORC, and then you had these tools and engines that read them, right?

[16:37] Spark and Trino and Flink, and so on. We're in a similar space in the AI world. We have these open formats and ecosystems starting to get formed, Agent Skills, AgentMD. Um, and then there's a bunch of harnesses that need to read them, Cloud Code, Codex, Copilot, so on. But in the data world, for a very long time, despite the open formats, interoperability was incredibly hard. Until Apache Iceberg. Apache Iceberg came in and created this common spec, uh, that made data truly interoperable across these engines. This is missing today in the AI layer. Two kinds of problems, both distribution and interoperability, need to be solved. And so with that, I am really, really excited to announce today AI Registry, an open source, open standard for AI registries, uh, built on Apache 2.0.

[17:46] Uh, our team's been working really, really hard on this. We're taking inspiration from what NPM solved for in code distribution and what data, uh, Apache Iceberg solved for interoperability in data, and we're hoping that AI Registry makes distribution and interoperability of AI context a standard across the organizations. What is it really? It's an open standard for AI registries. It's a spec. Uh, it's open format, open storage. It, you know, opens up across any target. It works across your entire ecosystem. Uh, later today, Ben will go deeper into what we really mean by the spec and how you can start using it starting today. So that's all I had for announcements. I want to end today by sharing why I care about this so much.

[18:36] I opened by saying that context is under-hyped. Here's what I meant. I truly believe in the power of AI and what it can do for the world, um, what it's already doing. In Sweden, this year AI helped radiologists find twenty-nine percent more breast cancers. In Nigeria, six weeks with an AI tutor gave kids up to two years of learning and, you know, opened up a new paradigm of personalized learning. This is personal for me. Before Atlan, Varun and I started a company around one question: Um, can data create positive social impact? We worked on government programs that touched close to a billion people, and that's where I learned that data without context, it doesn't just give you a wrong number.

[19:27] It gives a real person a wrong answer, a wrong outcome. AI is raising the stakes. A man whose grandmother had just died asked an airlines chatbot about bereavement fares. It made up a policy. The model wasn't stupid. It just was missing context. That is why the work of the people in this room is about to become some of the most important work in every company, and maybe in our world. Governance, catalogs, ontologies, knowledge management, data management. Uh, the work that we all used to call boring, um, in this new world, these humans become the people that decide what AI is allowed to believe. Uh, you write the playbooks that agents follow. And jobs are coming that don't even have names yet, context engineers, context modelers.

[20:25] The models are going to keep getting smarter. Whether that intelligence get har-harnessed for good, that's up to us. It's up to everyone in this room. Here's one thing that I know for sure. Tomorrow, you and your biggest competitor can rent the same intelligence, same model, same price, but what makes every single company unique, different, it isn't the model. It isn't the intelligence. It's what makes your company unique. It's how you do business. It's how you treat your customers. That is your context. It's how work gets done. Context is IP. So rent the intelligence but own the context. Thank you. That's all I had. Austin, back to you.

[21:17] **Austin Kronz:** Awesome. Thank you, Prukalpa, so much for getting us started today.