--- title: "Is the Data Catalog Finally Dead?" url: "https://atlan.com/context-and-chaos/issue/the-data-catalog-for-humans-is-dying/" description: "The catalog isn't dying. Its reader is changing. Agents now build context from evidence and read it at runtime, and the reading is the harder half." keywords: "Context Engineering, AI Agents, Data Governance, Enterprise AI" --- > Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/ A Context & Chaos PK Editorial by Prukalpa Sankar (Co-founder & Co-CEO, Atlan), published September 8, 2026 (15 min read). Every data catalog so far assumed a human on the other side of the screen. AI is now both producer and consumer of metadata, so the data catalog for humans becomes infrastructure for agents. The catalog answered "where is the data" and "what does it mean"; it is starting to answer a third question: now do the work. ## The three problems data catalogs never solved The catalog waited twice: for someone to create context, then for someone to come looking for it. In [Data Catalog 3.0](https://humansofdata.atlan.com/2021/01/data-catalog-3-0-modern-metadata-for-the-modern-data-stack/) (five years ago) PK argued metadata had become "big data", too vast and dynamic for humans to maintain field by field; that thesis still assumed a human reader. - **Creation.** Someone had to populate descriptions, business terms, domains and owners, and keep them current. The estate always changed faster than humans could document it, and AI needs far more context than people ever did. - **Consumption.** People had to leave the workflow, open the catalog and translate a question into catalog language. An agent reasoning inside Claude or Cursor has no interface into a search box and pages built for human eyes. - **Learning.** The catalog never learned. It was measured by coverage, which can't tell whether context worked. Nothing traced a bad answer back to the gap that caused it. Agents will not stay quiet. The agentic catalog is what happens when all three get answered. ## Phase one: solve the metadata supply problem Agents read evidence already in the estate (lineage, SQL and query history, usage patterns, column metadata) and write a grounded first draft. Humans set standards, inspect evidence, resolve ambiguity and certify what can be trusted. "AI-powered catalog" is too small a frame: the operating model changes. ### Can agents really document a data estate? - Atlan shipped its first AI features in 2023; quality plateaued at roughly 70%. Accuracy needed rich signals stored in systems built for a human interface. - In 2025 Atlan rebuilt the [foundation as a metadata lakehouse](https://atlan.com/know/is-the-data-catalog-finally-dead/): catalog knowledge as open Iceberg tables, queryable by anything that speaks SQL. The agent harness on top was rebuilt more than once and started working early this year. - A three-week cohort with 50 enterprises generated [1.03 million descriptions and 15,000 READMEs](https://www.linkedin.com/posts/atlan-hq_atlan-context-agents-1-million-context-descriptions-activity-7453083620082327552-uraL) (Atlan's own program numbers). Two surprises from participants: - **Agents recover knowledge nobody wrote down.** Query-pattern analysis turned real usage into business-friendly questions, common joins and important filters. Izabela Wilczynska: "The kind of knowledge that usually lives in one person's head or a Slack thread suddenly became visible across thousands of assets." - **Trust came from receipts, not fluency.** Cody Brees generated context across 103,515 assets and was convinced by provenance: "The context traced back to something I could point at rather than arriving as an assertion." An agentic catalog must generate context with receipts, not merely more prose. The cohort proved generation runs at enterprise scale; it did not prove the context is accurate, current and useful when another agent has to decide with it. ### The data steward becomes a context engineer Human expertise moves up the stack: deciding what good context looks like, which sources to trust, how to resolve conflicting definitions, testing agent output, catching exceptions that need business judgment. Leonardo Vinicius Ruivo: "It allows Data Governance professionals to spend more time validating information, improving context, and making decisions that require real knowledge of the business." Documentation becomes a continuously enriched product; metadata people were context engineers before the name existed. ## Phase two: AI becomes the consumer The primary consumer may be an agent inside ChatGPT, Claude, Copilot, an IDE or a custom app. It needs retrieval that is accurate, relevant, fast and cheap in tokens, delivered in the agent's reasoning loop (as argued in [The War of the Catalogs](https://www.linkedin.com/pulse/war-catalogs-prukalpa--liwwc)). An MCP server exposes catalog capabilities as tools under the catalog's existing permissions, but MCP is the interface, not the intelligence. Retrieval has two dimensions: - **Breadth: where should the agent look?** The same word can be a metric, glossary term, dashboard or entity. Atlan's first attempt, wrapping the existing API surface as MCP, failed: agents think in tasks, not endpoints, and people search by association (nothing is called "footfall", yet something must map that question to a metric). - **Depth: how much should it retrieve?** Too little omits a column, lineage edge or policy; too much floods the context window. One Atlan search call carried roughly 2,500 tokens of overhead before returning anything useful. The fix is progressive disclosure: a small, ranked, high-confidence result, visible true scope, honest truncation notices, then drill-down. Every tool is scored against an evaluation suite (right tool, right arguments). Those scores are good but don't measure whether the person got their answer, which is the number that matters. ## Phase three: the loop An agentic catalog is measured by outcomes: can an agent answer a real question correctly, explain why, and know what it doesn't know? Rather than waiting for real users, manufacture questions: generate representative personas, ask what they would ask, record every retrieval decision, and attribute each failure to the part of the system that produced it. Traced, three identical-looking failures separate, each with a different owner: context was missing; context existed but the underlying metric was never curated; or everything was there and the agent reached for the wrong thing. The compounding loop: agents generate context, agents retrieve it, real and simulated use exposes gaps, humans and agents fix the right layer, the next answer improves. A catalog is agentic when those halves close on each other. ## The agentic data catalog is not the context layer It is the data and metadata foundation a broader context layer is built on. Enterprise agents also need agentic context: which tools, which sequence, which constraints, when to ask for approval, what a good outcome looks like. | Stage | What it contains | The problem it solves | |---|---|---| | 1. AI-ready data | Metadata, descriptions, ownership, lineage, quality, usage, policy | Reliable context about the data estate. Agents solve the supply problem. | | 2. Agent-ready data | Governed semantics, metrics, business entities, relationships, ontology | Consistent meaning, retrievable at runtime. | | 3. Agentic context | Skills, tools, procedures, constraints, evaluations, operating knowledge | Teaches agents how to perform work, not merely how to understand data. | The agentic data catalog lives in stages 1 and 2; stage 3 extends into the full context layer and eventually a registry for agents themselves. ## The next chapter belongs to the people who built the last one AI makes missing context visible immediately: wrong answers, wrong definitions, no explanation. The fix begins with a remade metadata foundation, not a smarter model. The old human sequence (document, search, inspect, decide) breaks; coverage is no longer enough. Context needs evidence, retrieval needs judgment, and every wrong answer needs a trace to the missing definition, relationship, policy or skill. The foundation will be built and consumed by agents but cannot be entrusted to agents alone. **The catalog for humans is dying. The people who built it are what make the next one work.** - Longer version with the build log and numbers: https://atlan.com/know/is-the-data-catalog-finally-dead/ - Live session with practitioners, September 10, 11am ET, one hour (as listed on the page): https://atlan.com/context-in-practice-can-ai-document-your-data/ ## About Context & Chaos A community newsletter where practitioners, builders and thinkers share lessons on context engineering, governance, architecture, discovery and the human side of data and AI work. Contribute: https://atlan.com/context-and-chaos/contribute/ Related reads: - [Your AI Context Layer Is Being Built on Stale Metadata](https://atlan.com/context-and-chaos/issue/your-ai-context-layer-is-being-built-on-stale-metadata/) - [Context Graphs as AI Evaluation Infrastructure](https://atlan.com/context-and-chaos/issue/context-graphs-as-ai-evaluation-infrastructure/) - [The Onboarding Gap Is Killing Your AI Agents](https://atlan.com/context-and-chaos/issue/the-onboarding-gap-is-killing-your-ai-agents/) - [BI-Ready Is Not AI-Ready](https://atlan.com/context-and-chaos/issue/bi-ready-is-not-ai-ready/) - [Browse all Context & Chaos articles](https://atlan.com/context-and-chaos/)