DataHub Cloud Context Delivers Self-Improving Accuracy for Enterprise AI Agents

Five-month beta across financial services, retail, and technology companies produces significant new capabilities that let AI agents monitor, prove, and compound their own accuracy to earn sustained trust

PALO ALTO, Calif.--(BUSINESS WIRE)--DataHub, the company behind the most widely deployed open-source metadata and context platform, today announced the General Availability of DataHub Cloud Context with new capabilities that turn context management from a one-time setup into a system that continuously proves and improves agent accuracy. Most enterprises don't trust AI agents today because they routinely deliver confident, wrong answers. This is because they lack the business context to know which table is authoritative, what a metric means, or how a calculation changed last quarter. DataHub Cloud Context, now generally available, gives agents the context they need to earn organizational trust over time.

AI agents pointed at enterprise data warehouses without business context frequently produce accuracy rates of approximately 50%. Agents with auto-generated context are unpredictable, delivering results that range from 60% to 80%, depending on the complexity of the data estate. Teams that follow a sequenced domain-specific approach using DataHub Cloud Context consistently achieve near-100% accuracy on mission-critical metrics such as revenue and forecast, and close to 90% on complex operational and behavioral metrics, sustained over time across Claude, ChatGPT, Snowflake Cortex, Databricks Genie, and custom agent frameworks.

“More than six in ten organizations tell us missing or inconsistent context has caused a model to misinterpret data,” said Herb Blecher, Owner and Lead Analyst, Data and Analytics at Enterprise Management Associates (EMA). “At the same time, 41% report that wrong AI output has caused them to take a materially incorrect action. As businesses evolve, the context AI agents rely on needs to evolve with them. A one-time documentation project starts breaking down on day two. What’s required is continuous evaluation, with feedback loops that catch conflicts as they surface and keep human expertise governing the system.”

DataHub Cloud Context now ships with three new capabilities that form a sequenced approach that delivered uniformly consistent, trustworthy agents in the five-month beta that preceded general availability.

• Domain-scoped Context Management lets administrators group data assets, context documents, and eval questions into focused domains, such as sales reporting or product analytics, and connect agents to only the domains they need. Teams start narrow to their own specialized domains, prove accuracy to stakeholders within weeks, and expand from a measured baseline rather than trying to document an entire enterprise data estate at once.
• Cloud Context Evals give teams a persistent accuracy harness. Administrators define test questions and expected answers for a given data domain, and the platform continuously monitors whether agent responses meet those criteria as underlying data and context change. When a metric definition shifts or a source table is restructured, regressions surface before they reach end users. They can also run before changes to important context documents, preventing the introduction of context that conflicts with, invalidates, or removes existing context and degrades overall evals.
• Agent Feedback, through MCP, lets agents flag missing, invalid, or incorrect context based on real interactions with end users. It also provides a monitoring dashboard for data teams to proactively monitor and improve context, ensuring gaps are closed on ungrounded questions over time. When an agent encounters an ambiguous metric definition or a user corrects a result, that learning persists across agents, sessions, and platforms. Institutional knowledge compounds with every query instead of resetting between sessions.

Together, these three capabilities create a closed loop. Teams scope a domain, prove accuracy with evals, and let agents report feedback from real user conversations back into the context graph. Accuracy increases with use instead of decaying over time.

Customer Deployments

Optiver, a global automated trading firm, deployed DataHub across its petabyte-scale data ecosystem to provide trusted context for quantitative research and exchange connectivity, where a wrong definition carries direct financial consequences.

“Agents change the requirements of the data platform. The context people used to carry in their heads now has to become infrastructure,” said Christophe Godefroy, Global Head of Data Platform, Optiver. “On real researcher questions, governed context lifted our agents' accuracy from 42% to 74%, and DataHub is the layer that helps us connect, curate and serve that context.”

ICA, Sweden's largest food retailer, deployed DataHub across a complex ecosystem spanning retail, banking, and insurance operations to make governance practical across diverse user groups with different terminology and incentives.

“DataHub’s context platform solves the two biggest problems every classic data catalog has: keeping metadata up to date and trustworthy with human expertise, and making it easy for anyone to use and understand,” said Bjorn Barrefors, Metadata Management Lead at ICA. “For ICA, it lets us create custom agents that identify gaps in our context and help our experts manage it, then deliver the right high-quality context to people and AI agents through custom MCPs.”

Building the Context Foundation

DataHub organizes context into three layers. The technical layer captures what exists across the data estate, from tables and columns to dashboards and pipelines. The semantics and metrics layer defines how tier-1 KPIs are calculated, so questions about critical metrics return consistent, correct answers regardless of who is asking or which tool is doing the asking. The knowledge layer explains why those assets and definitions exist and how teams should apply them to business outcomes.

DataHub Cloud Context ingests existing tier-1 definitions and unstructured knowledge as they are, then fills the long tail through Context Intelligence, which auto-generates tier-2 context by extracting semantic signals from Looker LookML, Hex dashboards, dbt metrics and models, and warehouse query logs. Metric definitions, join patterns, filters, and analytical intent that typically live only in analysts' heads become explicit and reusable context, available to both humans and agents. Because Context Intelligence runs continuously, context stays current as data and business definitions evolve. The three new GA capabilities add the measurement, scoping, and learning loops that bring inferred context up to production grade.

"The clearest learning from the beta is that you have to start narrow and prove accuracy before you expand," said Shirshanka Das, co-founder and CTO of DataHub. "Customers who identified a specific data domain, set up eval questions with their subject-matter experts, and then layered in DataHub context saw clear, measurable accuracy gains. That sequence is now built into the product as a guided workflow."

Availability

DataHub Cloud Context is generally available now. Documentation for the guided playbook, which covers domain setup, eval configuration, and context generation, is available at www.datahub.com.

About DataHub

DataHub is the context platform for enterprise AI. It gives AI agents and the people who work with them trusted business context, including metadata, business definitions, and governance policies, across every data platform an enterprise runs. The company was founded in 2021, is headquartered in Palo Alto, CA, and leads the open-source DataHub project, with more than 16,000 contributors. More than 3,000 organizations use DataHub, and DataHub Cloud runs in production at Global 2000 companies including Apple, Block, Chime, Etsy, Foursquare, FIS, Netflix, and Notion. DataHub has raised $65 million from Bessemer Venture Partners, LinkedIn and 8VC. Learn more at www.datahub.com.