The details come from Collibra's own press release, product blog posts and documentation. The release was syndicated through PR Newswire and republished by AiThority. We found no independent hands-on reporting. This article treats capability claims as vendor descriptions, and treats the documentation as the most reliable record of what actually runs today.
Collibra Maestro brings no-code governance agents into the Collibra Platform
Maestro gathers Collibra's existing AI features into one suite and adds a new agent-building layer. Collibra says it is built on the company's enterprise Knowledge Graph. It has three parts:
- Maestro Studio is a visual, no-code builder where governance teams create their own agents.
- Maestro Assistant is the chat interface where end users talk to those agents in natural language.
- Maestro Actions (formerly branded Collibra AI) are built-in AI utilities, such as writing help for text fields and semantic mapping, that work without a chat.
According to Collibra's product blog, agents come in three kinds. Collibra manages system agents such as Business Definitions, Data and Analytics Discovery, and Collibra Documentation. Ready-made templates can be used as they are or adapted. Teams can also build fully custom agents in Studio. A builder gives the agent a name, a description and a unique @handle, then writes instructions that set its role, tone and guardrails. The builder then picks the agent's knowledge base by selecting specific views of organizational data, and chooses which tools it may use. Tools let an agent do things such as create an asset.
The key design choice is that tools are human-in-the-loop, meaning a person signs off before anything happens. The agent summarizes what it plans to do and waits for explicit approval. Before an agent goes live, the builder tests it in a live preview, sets who can access and own it, and submits it to a reviewer for approval. Published agents come with usage analytics. Agent configurations can also be imported from a YAML file.
The templates cover common governance chores. One turns a cataloged table into a data product with ports and a data contract. Others check glossary definitions for conflicts, convert discovered data problems into tracked issue assets, and detect duplicates in a glossary. Collibra says the Data Product Agent is built on Collibra Maestro and available as a skill through MCP Server, where MCP refers to the Model Context Protocol. It pulls columns, lineage, semantic-layer data and sample data from the catalog and drafts the full package, and then stops and shows you the full proposal before anything is written to Collibra.
Collibra's headline claim that Maestro turns "weeks of manual work into minutes" rests on the company's own examples. One describes a governance team cutting an afternoon of reviewing AI use cases to under ten minutes. These are illustrative scenarios, not measured customer results.
Maestro Assistant is cloud-only, and access depends on three permission layers
The documentation adds limits that the launch release leaves out. Maestro Assistant is labeled "in preview" and is a cloud-only feature. It isn't supported in Collibra Platform for Government or Collibra Platform Self-Hosted (CPSH) environments. That rules out regulated and air-gapped customers for now, and those are often the organizations that most want tight AI governance.
A user needs to clear three checks to use an agent:
- An administrator has enabled Maestro Assistant in the environment.
- The user holds a global role with the Maestro Assistant global permission.
- For custom agents, the user has access to that agent and view permission on the data it references.
That third check does real work. The documentation says an agent suggests assets only when it can reach the information and the user has View permission on them. A custom agent therefore can't show a user anything that user couldn't already see.
Collibra's Settings documentation says configuring agents requires a global role with the Manage All Agents global permission. Agents listed there can have a status of Preview or Generally available.
The Assistant accepts prompts of up to 5,000 characters in English, Spanish, French, German and Dutch, and it replies in the language of the question. It uses the full conversation history across every agent called in the same chat. Collibra advises starting a new chat so answers aren't shaped by earlier context. Users can switch agents in the middle of a conversation, but they can't send one question to more than one agent at a time.
Existing customers face a deadline. The Settings documentation says Collibra AI Copilot will be upgraded to Collibra Maestro with the 2026.10 release. If you are using Collibra AI Copilot, we recommend enabling Collibra Maestro before October 11.
Live Map targets the document context agents keep rebuilding
Live Map comes from Deasy Labs by Collibra, a unit that specializes in unstructured data. Collibra says it gives agents access to a context graph capturing what entities exist across a company's documents, how they relate, and how those relationships evolve.
The idea is to do the work once. Collibra says Live Map means agents no longer have to repeatedly reconstruct those relationships for every question. The graph is built ahead of time and shared, so agents don't each work out, query by query, how the entities in a document collection connect. Collibra's release ties this to a survey finding: 87% of decision-makers said their teams regularly re-verify that the context their agents rely on is still accurate and current.
The limits are clear. Live Map opens this quarter only to a small group of design partners. The release names no general-availability date, supported document sources, integration list or pricing. It also specifies "curated" unstructured content, which means someone still has to pick which documents feed the graph. No performance figures back the "more performant agentic retrieval" claim.
Agent Contracts and Guardian Agents aim at the agents Collibra doesn't own
Collibra CEO Felix Van de Maele said Maestro handles governance work inside Collibra, but that "The much bigger challenge is everything else. Agents your teams bought, built on other platforms, or stood up without governance in mind are all subject to the hallucination tax." Agent Contracts and Guardian Agents are Collibra's answer to that problem.
Agent Contracts establish layered, machine-readable controls that scale across the agent fleet, with global rules and specialized requirements based on factors such as function, risk classification or operating context. The structure is familiar to anyone who has built layered policy: organization-wide rules, with stricter overlays for, say, a high-risk finance agent. Collibra's platform page says contracts are generated automatically from the policies, glossary and compliance documents already in Collibra. A Collibra product post says contracts are meant as an open standard that doesn't depend on any particular runtime. Collibra hasn't published a specification, so readers can't yet check that claim.
Guardian Agents enforce those rules while agents run. According to Collibra, Guardians read them at runtime to evaluate every material action. When an agent steps outside its boundaries, Guardians can respond anywhere from flagging and escalating the issue to blocking the action. Collibra's related product post says blocking depends on which enforcement mode is turned on. Organizations will likely start in monitor-and-flag mode and only later allow Guardians to stop agent actions, though that rollout path is our inference.
The announcement doesn't answer the question an IT architect will ask first: which agent frameworks, clouds and runtimes can Guardians actually watch? Collibra hasn't published a compatibility list. Its October 2025 blog on the AI agent registry offers one earlier clue. That registry was described as tool-agnostic, with initial integration in place for Azure AI Foundry. That is context, not confirmation that Guardian Agents will support Foundry-built agents at launch.
Both components are scheduled to arrive through Collibra's AI Command Center in October 2026. Collibra plans demos at its AI Summit in New York on October 6 and London on October 21.
Harris Poll figures put numbers on the "hallucination tax"
Collibra's case rests on its 2026 Hallucination Tax Report, which The Harris Poll conducted for Collibra and published on September 16. According to Collibra's survey release, 306 U.S. adults took part online between August 5 and 11, 2026. All were full-time data management, privacy or AI decision-makers at director level or above. The findings:
| Finding | Share of respondents |
|---|---|
| Hit critical roadblocks moving agents from pilot to production in the past 12 months | 76% |
| Teams regularly re-verify agents' context is accurate and current | 87% |
| Spend significant staff hours manually reviewing and correcting agent outputs before go-live | 51% |
| Same as above, at organizations with $100 million or more in revenue | 64% |
| Feel AI initiatives today are falling short | 72% |
The survey is vendor-sponsored and limited to U.S. respondents, and each product line is launched on the back of the statistic that suits it. The figures are still useful as a snapshot of where agent programs stall. Review and re-verification hours are the costs Collibra means by "hallucination tax." That is a label for staff time, not a measured dollar amount.
AIUC-1 certification covers Maestro Assistant, not the whole Collibra Platform
Collibra also says Maestro Assistant has earned AIUC-1 certification. AIUC-1 is a standard from the Artificial Intelligence Underwriting Company with requirements covering data and privacy, security, unauthorized actions, monitoring and human oversight. Collibra, and AIUC co-founder Rajiv Dattani in the release, call it the first data-and-AI governance company to have one of its agents certified. That "first" claim belongs to the company and the certifier.
The certification's scope is narrow. It covers one agent, Maestro Assistant. It doesn't cover the full Maestro suite, custom agents customers build in Studio, or the upcoming Contracts and Guardians. Collibra adds that AI Command Center has included a built-in AIUC-1 assessment since May 2026. The assessment helps organizations measure their own AI against the standard and find gaps. The announcement doesn't say that passing the assessment leads to certification, so treat it as gap analysis only.
What this means for Collibra admins and AI governance teams
If your organization already runs Collibra in the cloud, act on the Copilot-to-Maestro change now. If you're weighing Collibra for broader agent governance, wait for the October release and its integration details. Maestro is usable today, in preview. The cross-platform enforcement pieces are announcements, and Live Map is effectively a pilot.
- Collibra AI Copilot customers should enable Maestro before October 11, ahead of the 2026.10 upgrade that converts Copilot to Maestro.
- Collibra Platform for Government and Collibra Platform Self-Hosted customers can't use Maestro Assistant, because it is cloud-only.
- Admins should map the Maestro Assistant global permission and the Manage All Agents permission to roles on purpose. Custom agents inherit users' view permissions but still need reviewer approval before they publish.
- Teams evaluating Agent Contracts and Guardian Agents should ask Collibra for a list of supported agent runtimes and an explanation of which enforcement modes can block actions, rather than only flag them.
- Risk teams can use the AI Command Center AIUC-1 assessment for gap analysis. Maestro Assistant's certification doesn't extend to agents your organization builds.
Collibra is pitching itself as the control layer for AI agents it never built, and so far it has shipped only the part that governs its own. The October AI Command Center release and the New York and London demos will show whether Agent Contracts work as a runtime-neutral standard or mainly cover platforms Collibra already integrates with. Until then, Maestro in preview is the only part customers can actually run, and governance teams can already try its approval-gated agents on real backlog work.