Announcing Collate 2.0: Build Trusted Context, Then Put It to Work
Ask an AI agent any business question ("which regions missed our quarterly revenue plan?"), and it will find a sales table in your data warehouse, write valid SQL, and hand you a number. However, whether that number is the one your finance team would trust enough to sign off on is a separate question. The reliability of AI results depends on business understanding and data context your warehouse doesn't hold, scattered across systems in your data estate and in your team's institutional knowledge, locked away in their heads. That gap usually becomes the platform team's problem to close.
Today we're announcing the release of Collate 2.0, the next generation of our agentic automation and orchestration platform for data teams, to solve this challenge end to end. Collate 2.0 builds that enterprise context and understanding. It carries the context from the question to the answer, into recurring workflows, through data-access decisions, out to the agents your engineers build, and into governance of the AI itself. And it is built on OpenMetadata, the Open Context Layer for AI agents, to ensure that this remains open and interoperable with all your AI and data systems.
The gap between a valid answer and a trusted one
Let's walk through this journey, from data to business understanding to answers, and how it can go astray. Taylor is a business analyst who needs to present at Monday's operating review on which regions missed Q2 revenue plan, and why. It's a high-stakes discussion, and every regional sales and marketing leader will see the answer. When it goes wrong, Sam, the platform engineer supporting her team, has to explain why.
Taylor asks a general-purpose AI agent running on Claude or Codex, connected to the company's warehouse. It takes multiple round trips to query the data warehouse, looking for the correct data and tuning its SQL, consuming tokens and running up warehouse costs. After a number of loops, it creates a chart that says that EMEA missed the plan by 3%, while APAC hit its target.
However, she does not trust it, so she escalates to Sam's team. Sam spends two days chasing the same three facts he chased last quarter: Finance's approved definition of "revenue". FP&A digs out the board deck naming the Q2 plan version the board actually committed to. A colleague remembers territories were redrawn in July and finds the mapping in effect during Q2. On those refined definitions: EMEA beat plan by 1%, APAC missed by 9%. Different answers that have dramatically different impacts on business decisions and success. Taylor gets her number for Monday. Sam gets the same ticket again next quarter.
This is the context AI agents and human data teams need in order to understand the data and answer business questions correctly. However, getting data to be AI- and people-ready is time-consuming, and this manual approach Sam takes can't scale with the questions people ask or the rate at which data estates grow. Multiply this by every AI agent your team deploys, and the problem is no longer just building the context; it's also governing the agents that use it. This is the context challenge that data teams face, and what Collate 2.0 solves.
Introducing Collate 2.0
Today we're introducing Collate 2.0, the AI for Data platform that operationalizes OpenMetadata, the Open Context Layer for AI agents.
OpenMetadata gives people and agents three primitives for working with enterprise data:
- Context defines what data exists. It connects metadata, documents, policies, lineage, quality signals, and access rules relevant to a task.
- Ontology defines what the data means, through the business concepts and typed relationships an agent uses to understand that context.
- Memory preserves approved decisions and corrections, so the next answer starts from what the organization has already settled.
Together, they help agents connect all the pieces of data together, reason correctly over how the data connects to the business, and remember learnings on how to do this analytics correctly. These primitives are designed into a knowledge graph built on the open standards of OWL and RDF, which keeps it portable and lets people, Collate agents, and external AI applications reason over these shared predicates at scale. Additionally, context has become strategic to how organizations organize their data landscape and create value for their AI initiatives, and why companies choose open solutions like Collate to prevent vendor lock-in.
Collate turns those primitives into operationalized workflows for enterprises. Teams build context, use it to answer a business question, turn the answer into a recurring workflow, and make the same context available to the people and agents that need it in the future. Additional Collate capabilities for security, governance, and automation help enterprises productionize this work for data teams and agents, accelerating AI and data initiatives at scale. Let's return to Sam and Taylor's example and how Collate 2.0 can deliver her trusted answers, built on enterprise context.
Part 1: Build the enterprise context an answer needs
Collate connects to more than 130 data sources across warehouses, databases, BI tools, data catalogs and pipelines including Snowflake, Databricks and Tableau. The rebuilt connection and ingestion experience verifies access before a job runs, identifies setup failures, carries service-level scope into ingestion workflows, and shows logs and runtime estimates while jobs execute. This is part of a revamped AI-Native product experience, redesigned from the ground up, to deliver a conversational AI and agent-oriented way of working with your data.
Once connected, metadata, lineage, usage, profiles, quality signals, and classifications become part of the shared knowledge graph. Specialized AI agents automatically organize and enhance this data to make it ready to be used for work. The Documentation agent drafts descriptions, the Data Quality agent proposes tests, and the Tier agent assigns criticality. AutoPilot orchestrates these across the data estate. All these agents are designed for human-in-the-loop reviews and tracking, such as recording agent-proposed descriptions and which person approved them.
Bring structured and unstructured knowledge together
Context Center is the governed home for articles, documents, and memories, including third-party knowledge connected from Confluence and Google Drive. Sam can bring in the same three documents that cost him two days last quarter, once, so nobody chases them down again. Access controls apply when an agent retrieves this material, ensuring that the agent sees only the data and knowledge the person it represents is allowed to use. Memories tagged to data can be tracked and customized here as well, to ensure learnings persist for other team members and agents.
Make relationships explicit for AI agents
Connecting the relationships within and between technical context and business context is essential in order for humans and agents to understand how to use and interpret the data correctly. Knowledge Graph is asset-centric: starting from any table, dashboard, or pipeline, it shows everything connected to that asset. The enhanced Knowledge Graph visualization helps users understand how technical metadata connects with business definitions, ownership, policies, lineage, quality, data products, organizational knowledge, and AI assets.
Additionally, the Ontology Explorer makes the relationships between business concepts navigable and connects them to the underlying data assets that implement them. This way, a "Revenue" business concept can be tied to the database tables behind it. These graphs of relationships, with their different visualizations, are necessary for agents to bridge the span between data and business correctly.
Together, both Knowledge Graph and Ontology Explorer give you both directions of the same graph, from data assets upward into business meaning, and from business concepts downward into the data that implements them.
Curate the context each role needs
Structured context is powerful, but it is also important to focus on what specific parts of it are relevant to the task at hand to manage context bloat, reduce token consumption, and improve performance. Persona-Based Context Curation defines the different context that different users receive. Rules can define the assets a business user, data engineer, or compliance lead can receive, such as filtering by service, domain, glossary term, tag, or type. These can be further refined by the relevant signals needed, such as definitions, metrics, joins, lineage, quality, profiles, and samples. This way, a finance persona (human or agent) receives the approved revenue metric and the planning policy. In contrast, an engineer investigating the same result receives the underlying schemas, lineage, quality failures, and transformation models. Taylor receives the context specific to her needs, without unnecessary clutter.
Retrieve context more quickly
Retrieving this curated context programmatically is necessary for agents to work with the data. The updated Context API and Model Context Protocol (MCP) tooling have been developed for faster context delivery. Calling get_asset_context returns an asset's related context in a single call: schema, transformation logic, business definitions, classifications, lineage, quality, ownership, and relevant knowledge. Similarly, get_persona_context returns the equivalent at the persona level. The relevant passages are pulled from longer documents with semantic chunking, and the embedded MCP server makes the context available to Collate and external agents under existing role-based policies. This tooling is available as Compact Markdown for a model, or structured JSON for an application.
Part 2: Answer questions within access controls
With this data and business context structured and organized, humans and agents can now run analytics against it with higher-quality results. Collate AI Analytics takes business questions in natural language chat and routes each request through the metrics, glossary terms, relationships, memories, and policies to understand the question and correctly answer it. It resolves "revenue" to the finance-approved recognized metric definition, checks against territory plan documents, and maps to the correct production data table, all while maintaining role-based access controls and governance. Taylor can review the generated SQL, read the prompt and reasoning history, and trace the answer through lineage to its underlying data sources to validate the accuracy of the analysis and the associated charts generated, all without opening a ticket to Sam's team to get there.
Part 3: Operationalize the answers into repeatable data workflows
The charts created in chat can be promoted to a governed Dashboard for sharing or scheduling through Slack or email across the broader team, each with an owner, role-based access, status, lineage, and build history. If any questions or deeper analytics emerge, the team can easily drill back into the original analysis with collaborative, real-time chats.
These Shared Chats bring together colleagues and AI agents into a single chat thread with mentions, role-based access, and an audit history of every contribution. The APAC regional leader can take the existing AI Analytics chat conversation behind a revenue chart, and ask for a territory drill down. The Collate AI Analytics agent will answer and generate a new widget, without having to redo the original analysis. Taylor follows up in the same thread to confirm the updated SQL is right and the answer is safe to share, without looping Sam in.
AI Automations can further operationalize tasks, by turning a natural language instruction into repeatable data workflows. Since the documentation, classification, tiering, data quality, PII discovery, glossary linking, knowledge import, and reporting all run on the same open context layer, these can all be turned into automatically recurring data workflows at scale, with natural language instructions. Each automation has an owner, permissions, status, live logs, and execution history. Authorized users can inspect it, rerun a failed execution, and audit what it did. Whether it's reviewing PII, linking assets and terms, generating data quality tests, or scheduling analytics work, you can easily automate these tasks across your data estate. In our case, Taylor takes the APAC regional leader's chart and turns it into a weekly report to be sent by email and Slack with a natural language command, without a ticket to Sam's team.
For Sam's manager, administrators, and data leaders who are looking to understand user adoption, data health, and work that still needs review, Collate 2.0 ships new capabilities to give more data about your data. The Admin Analytics dashboards, a weekly impact email, the AI Audit Log, and the redesigned Inbox of pending tasks bring visibility into the data, teams, and AI working across the estate.
Part 4: Grant access with controls
The APAC regional leader wants to go deeper into the data and is looking for the customer-level detail behind the 9% miss, but doesn't have data access. He starts in the Data Marketplace, where he can review the business Domains and Data Products available. He filters by owner, glossary, domain type, tags, or service, and evaluates a candidate data product against its definitions, ownership, lineage, and quality signals — then requests access.
That kicks off a Data Access Request against the table, where the data steward can accept or adjust access, including column-level scope, time-bound expiration, and data masking. Every request captures its business reason and approval activity, and joins a dashboard for approvals, revocations, and audit. From here, Collate's Policy Agent turns that approved request into a scoped grant against the source system, such as Snowflake or Databricks, which is revoked automatically when it expires. Reverse Metadata pairs with this by syncing edits to descriptions, tags, and owners back to the data source, so governance decisions stay consistent across systems.
In this case, the finance data owner approves the request but narrows the scope to a select set of columns and sets it to expire one year out, at the next territory reassignment. This used to mean a ticket to Sam, but not anymore.
Part 5: Build your own agents on the context
While the new dashboard has been automated to run weekly, the finance team now wants to know about a variance as soon as it appears, and wants an agent that watches the data continuously. They'd like a more sophisticated analysis of the data, more sensitive detection of variances, and deeper reporting of details, all while maintaining the necessary data access governance controls.
AI Studio provides a powerful UI-based builder for customizing the existing agents within Collate, or for building your own custom agents. An analyst can assemble the agent from a pre-built persona, assign it skills, plug in data access, and add custom prompt instructions. These agents can leverage Collate's turnkey LLM service, or customers can bring their own LLM and keys across OpenAI, Anthropic, Bedrock, Vertex, and on-premises models, to match each organization's specific security and privacy needs. These AI Studio agents can then be run or scheduled through the AI Automations noted earlier, or through the AI SDK.
For organizations looking to build their own agent infrastructure, the AI SDK and Enterprise-grade MCP Server give platform engineers the tooling needed to build on the Collate Platform. Python, Java, and TypeScript SDKs support code-first development, and the natively embedded MCP server can be used to invoke the same context from outside Collate, with high scalability and with the same user-inherited governance model. The Context API and MCP tooling discussed earlier can be used for easily pulling out the complete and relevant context for agents, applications, and third-party services like Databricks Genie and Snowflake Cortex.
Sam's newly created agent escalates APAC's 9% miss on its first run, though Finance already knows this is due to one large customer's renewal slipping by three weeks, and they don't want it escalated every morning. A memory is logged back in the Context Center described earlier, so that the next time Sam's agent runs, the rule is applied to ensure the agent's alerts remain actionable.
Part 6: Govern the AI you deploy
By the end of a quarter, a dozen more variance agents have been set up to monitor and report on other areas of business. However, when the audit committee asks the AI governance lead for the list of AI systems touching financial data, who approved each one, and what each can reach, the team finds it has quickly outgrown its manual spreadsheet tracking.
AI Governance Studio enters Release Preview with Collate 2.0. It brings AI applications, LLMs, MCP servers, and agents into the same graph as the data and context they use. The AI Asset Registry tracks AI across your estate as governed assets with owners and stated use, which also enables their discovery by other teams to prevent duplicate agent creation. AI lineage can also be traced back to the underlying context it is built from. New AI assets can be routed through a risk council approval process for review before production release, and policies can be defined for AI governance requirements. These policies roll into a framework dashboard to ensure compliance across multiple regulatory requirements, such as the EU AI Act, NIST AI RMF, and ISO/IEC 42001. And when an auditor asks for evidence of these controls, machine-readable audit packs can be easily exported with the latest state of governance. Read the full AI Governance Studio announcement for more feature details and Release Preview information.
OpenMetadata or Collate
OpenMetadata 2.0 is also being released, providing the open context layer under Apache 2.0 licensing. Context Center, Context APIs, MCP Server, AI SDKs, Memory, Knowledge Graph, Ontology Explorer, Domains, Data Marketplace, and Data Products are all included, among other feature improvements.
Collate 2.0 combines a managed OpenMetadata runtime with additional enterprise and AI capabilities, including: AI Analytics, Shared Chats, AI Studio, AI Automations, Reverse Metadata, Admin Analytics, and AI Governance Studio.
Pick the right approach based on the requirements of your data estate and data team. For either solution, your context remains open, portable, and extensible. We believe deeply in open source and open standards, with support for Data Catalog Vocabulary (DCAT), Data Product Ontology (DPROD), Resource Description Framework (RDF), Simple Knowledge Organization System (SKOS), PROV Ontology (PROV-O), Web Ontology Language (OWL), Schema.org, Open Data Contract Standard (ODCS), Open Semantic Interchange (OSI), OpenLineage, MCP, JSON, and more.
| Invest in OpenMetadata | Evaluate Collate fit |
|---|---|
| Your estate is stable and small enough to maintain | Your asset count is growing faster than the team maintaining it |
| You have engineers who can own upgrades, connectors, and incident response, and you want them there | You need your platform engineers spending more of their time on the data instead of the platform |
| Discovery, lineage, and quality cover what you need, without needing automations or agents built on top. | You need answers, automations, or agents built on the context, operated for you rather than by your team. |
| Governance decisions can live in the platform and be applied by your team | A governance decision has to reach Snowflake or Databricks as an enforced grant |
| No AI systems are in production against your live data and you don't need governance for them | Production agents are running and you need to list them, their owners, or what they can reach |
Collate 2.0 at a glance
This table provides a deeper dive into the capabilities and updates of 2.0, including additional enhancements not mentioned above:
| Capability | What changed in 2.0 | OpenMetadata or Collate | Availability |
|---|---|---|---|
| Rebuilt connection and ingestion, 130+ connectors | Enhanced | Both | GA |
| Documentation, Data Quality, Tier, and Classification agents, AutoPilot | Enhanced | Collate | GA |
| AI-Native experience, persona-adaptive landing pages, Explore, Inbox | New | Collate | GA |
| Context Center: articles, documents, memories, Confluence and Google Drive | New | Both | GA |
| Organizational Memory: attribution, editing, reuse | New | Both | GA |
| Memory automation for automatic generation and application to tasks | New | Collate | GA |
| Persona-Based Context Curation with filtering, context preview, token cost | New | Both | GA |
| Knowledge Graph visualization of asset relationships | Enhanced | Both | GA |
| Ontology Explorer visualization of business concepts relationships | Enhanced | Both | GA |
Context API get_asset_context and get_persona_context and MCP tooling with semantic chunking, Compact Markdown and JSON output formats | New | Both | GA |
| Embedded MCP server, RBAC-aligned retrieval | Enhanced | Both | GA |
| Data Marketplace, Domains, Data Products with input and output ports | Carried forward | Both | GA |
| Data Access Requests: scoped, time-bound requests, collaborative approval | New | Collate | GA |
| Reverse Metadata: sync decisions written back to source systems | Carried forward | Collate | GA |
| Custom Intake Forms for Data Products, Domains, Glossary Terms | New | Both | GA |
| Bulk Metrics Import, Export, and Edit | New | Both | GA |
| Desktop Application for more convenient access for active users | New | Collate | GA |
| Temporal lineage to see how lineage changes over time | New | Both | GA |
| AI Analytics, chart-level inspection, prompt and reasoning history | New | Collate | GA |
| Dashboard with owner, status, lineage, build history | New | Collate | GA |
| Shared Chats, scheduling, Slack and email delivery, chat-to-automation | New | Collate | GA |
| AI Automations: ownership, permissions, logs, execution history, reruns | Enhanced | Collate | GA |
| AI Studio: Workers, Personas, Plugins, natural language planner and templates | New | Collate | GA |
| AI SDK: Python, Java, and TypeScript SDKs | Enhanced | Both | GA |
| Bring your own LLM and keys: OpenAI, Anthropic, Bedrock, Vertex, on-premises | Enhanced | Collate | GA |
| Admin Analytics, usage, billing, weekly impact email | New | Collate | GA |
| AI Audit Log of humans and agent activities | Carried forward | Both | GA |
| Managed runtime: 99.9% SLA, SOC 2 Type II, 2M+ assets, SaaS, Hybrid Runner, BYOC, on-premises | Carried forward | Collate | GA |
| New Connectors: Omni, Prefect, SAP BW/4HANA, SAP BW/4HANA Pipeline | New | Both: Omni, Prefect · Collate: SAP BW/4HANA, SAP BW/4HANA Pipeline | GA |
| AI Governance Studio: AI Asset Registry across the data landscape | New | Collate | Release Preview, design partners |
| AI Governance Studio: AI Lineage from agent to context to governed data | New | Collate | Release Preview, design partners |
| AI Governance Studio: risk council review, approvals | New | Collate | Release Preview, design partners |
| AI Governance Studio: policies, breaches, and drift tracking | New | Collate | Release Preview, design partners |
| AI Governance Studio: cross-framework checks against policies | New | Collate | Release Preview, design partners |
| AI Governance Studio: machine-readable audit pack export | New | Collate | Release Preview, design partners |
See Collate 2.0 run end to end
We've covered a lot of material in this blog, due to the scale and scope of this monumental release for Collate. Come see these features live in our three-part Collate 2.0 product demo series; we've had to spread it across multiple sessions given the number of capabilities to showcase. Sign up for the demo sessions. And if you'd like to discuss your use case, contact us to meet with a Collate expert.