AI for data needs a context layer. See how OpenAI and Anthropic do it.

Anthropic just published a piece on how they enable self-service data analytics with Claude inside their own company. It is worth reading in full, and it maps exactly to what we have been saying.

The blog landed at the same time Bonnie Xu from OpenAI's data team walked through the architecture of Kepler, their internal AI data agent, at Collate Summit '26. At this same summit, we announced Collate 2.0's AI for Data Teams, a native experience of agents, skills, memory, and context.

The industry is consolidating on a single point of view around what it takes to build AI agents and analytics that actually work on enterprise data, and it is not a bigger model. It's about what surrounds the model: the data it reasons over, the meanings it grounds in, the memory it inherits.

This is a pattern we have seen before. Throughout the internet and cloud eras, the most transformative enterprise products started as internal tools built by tech giants to solve their own massive problems. NoSQL databases. Recommendation engines. Stream processing. We built Uber's Kafka platform to scale to 5 trillion events per day before any of it became commodity infrastructure. The frontier proves the architecture. The commercial platform makes it available to everyone else.

We are watching that pattern play out again, in real time, with AI for enterprise data. So how did the frontier labs actually make these agents work and prevent them from hallucinating or pulling from the wrong tables?

The secret isn't a more powerful LLM. It is the context layer.

The Anthropic team is direct about what makes agentic analytics actually work. Before they layered structured skills on top of the data, Claude's analytics accuracy was 21%. After: it consistently exceeded 95%.

The more revealing number is the one most teams overlook. Anthropic tested an obvious shortcut: give the agent grep access to the entire SQL corpus — every dashboard query, every transformation, every notebook ever written. Accuracy moved less than one point. The bottleneck was not access. The bottleneck was structure.

This is the same conclusion OpenAI reached. Their Kepler system serves 3,500+ internal users against 600+ petabytes across 70,000 datasets. They did not build the metadata layer from scratch. They built Kepler on OpenMetadata, the open standard that came from our team's work at Uber. Six layers of context, governed in one place, queried by the agent.

Two frontier labs, two implementations, one architectural conclusion: for AI to work on enterprise data, it needs a rich context and semantic layer. It needs infrastructure that grounds the model in explicit business meaning, so it never has to guess what a metric means.

What everyone is consolidating on

AI for code worked because engineers spent a decade building a context layer for it: the IDE, the type system, the dependency graph, the git history. AI for data has had nothing comparable. Models pattern-match on tables they do not understand.

What Anthropic calls canonical datasets, semantic layer, skills, and validation, we describe as three primitives:

• Context: a unified metadata graph across every source, every connector, every dimension. The agent reasons over the actual shape of your enterprise, not a sample.

• Semantics: an ontology that gives data business meaning, not just structure. "Revenue" means what your CFO says it means, not what the agent guesses from a column name.

• Memory: an auditable record of every correction, every classification, every decision. Captured once, inherited everywhere a future agent or analyst touches the data.

Different vocabularies, same architecture. We had the same finding for our own benchmarks as well. Adding semantic context to enterprise data lifts analytics accuracy roughly 7×, from ~11% baseline to ~76%. Gartner expects organizations that prioritize semantics in AI-ready data to lift agentic AI accuracy by up to 80% and cut costs by up to 60% by 2027. The variable isn't the model. It is the context surrounding it.

Build your own AI for data, or ship what they've already chosen

Anthropic has the engineering depth to build a four-layer agentic stack from scratch, maintain a skills inventory under CI, and run their own adversarial review sub-agents. OpenAI has the depth to build Kepler, with a multi-layer context model and a memory system that learns from every interaction. Most data teams do not. You should not have to be a frontier AI research lab to ship a reliable, comprehensive data agent for your company.

This is the pattern from the last two decades. The largest tech companies build for themselves first; the rest of the industry follows the approach they prove and the open foundations that come out of it. The same is true here. OpenAI's and Anthropic's data teams have already done the architectural work. You don't need to repeat it, you can start where they ended up, on the same open foundation they already chose.

Collate 2.0 runs on OpenMetadata, the same open standard OpenAI's data team selected for Kepler. Context is driven by 130+ ingestion connectors, column-level lineage, automated profiling, and a unified metadata graph. The semantic layer ships with Ontology Explorer, governed metrics, glossary relations, classifications, and a queryable knowledge graph. Memory is a first-class object model: captured by humans or agents, tagged to assets, inherited by every interaction that follows. MCP is built in. Skills, plugins, and AI Automations are out of the box. All the tooling is there for data engineering and governance teams to keep the enterprise context layer structured and grounded, ensuring high-quality analytics.

For analysts and leaders, self-service analytics surfaces as Collate AI Analytics — ask a question in natural language, get charts and queries grounded in governed metric definitions. The agent runs the query against your warehouse with your credentials, not a shared service account. A Thinking Panel shows what tables it picked, what definitions it referenced, what filters it applied. Transparency and provenance built in the same way Anthropic ships theirs.

Our goal is that your users only need to be business literate, not data literate, to harness the full power of your enterprise data.

Every company deserves an Enterprise Brain for Data

AI for code had a context layer. AI for data does too now, and the architecture is no longer an open question. If you have OpenAI's engineering bench, you can build it. If you do not, you can stand on the same architecture that OpenAI's and Anthropic's data teams trust, and ship in days, not quarters.

Either way, the bet is the same: context, semantics, and memory beat tokens and parameters on enterprise data. Let's build the future of AI for data together — open, unified, and grounded in your business. Check out our explainer page for more details on how the industry is evolving.