Enterprise AI governance is moving closer to the point where software actually takes action, as Collibra adds controls designed not simply to catalogue models and data but to supervise autonomous agents while they operate across business systems.

Collibra has unveiled three related capabilities — Maestro, Live Map, and Guardian Agents with Agent Contracts — aimed at organisations moving agentic AI from limited pilots into production. Together, they address the construction of governance agents, the business context available to those agents, and the rules applied to AI systems created both inside and outside Collibra.

Maestro brings the company’s existing AI utilities together with a visual environment for configuring agents that perform governance work. Administrators can define an agent’s permissions, tools, and approved information sources, while employees interact with the resulting systems through natural-language interfaces.

Live Map addresses a different problem, since the context supplied to AI systems can become stale or fragmented as documents, entities, and business relationships change. Collibra says the product creates a context graph over curated unstructured information so agents can retrieve relationships without reconstructing them independently for every request.

Agent Contracts and Guardian Agents move further into operational control. The company says organisations will be able to express policies in machine-readable form and apply them to external agents, while Guardian Agents monitor behaviour and can flag, escalate, or block actions that fall outside the permitted boundaries.

Governance changes when software can act

Traditional data governance developed around comparatively static questions: who owns a dataset, what information it contains, who may access it, how long it should be retained, and which processes depend on it. Model governance added questions around training data, evaluation, explainability, and risk, but many systems still produced an answer or prediction for a person to act upon.

Agents change that relationship because software may be given credentials and tools that allow it to alter records, contact customers, generate orders, move information, trigger workflows, or interact with other applications. A wrong answer can consequently become a wrong action before an employee sees it, making permissions and runtime supervision as important as the accuracy of the model producing the decision.

Collibra’s new products are designed around that shift. Agent Contracts are intended to make operating rules explicit and machine readable, while the Guardian layer attempts to apply those rules during execution rather than relying entirely on an assessment completed before deployment.

The company frames the wider problem as a “hallucination tax”, although that phrase is Collibra’s own description rather than an established industry measure. A Harris Poll survey commissioned by the company found that 76% of participating data and AI decision-makers reported significant obstacles when moving agents from pilots into production during the previous year, while 87% said teams regularly rechecked whether agent context remained accurate and current.

The same research found that 51% reported significant employee time being spent manually reviewing and correcting autonomous-agent outputs before launch. Those figures are useful evidence of the problem Collibra is targeting, but the survey was commissioned by the vendor and should not be treated as an independent measure of the entire enterprise market.

Agent fleets create a control problem

The difficulty becomes greater when organisations use agents from several suppliers. A company may deploy AI inside productivity software, customer-service platforms, developer tools, analytics systems, finance applications, and internally built workflows without those agents sharing a common policy framework.

That fragmentation can leave governance teams trying to understand permissions and behaviour one application at a time, while the number of systems capable of taking autonomous actions grows faster than the number of people available to review them. A central policy layer therefore becomes attractive for the same reason identity and access management emerged as a distinct enterprise function: controls are harder to maintain when every application implements them differently.

Collibra is trying to use its existing position in data governance as the point from which those agent rules are administered. Its AI Command Center already provides inventories and assessments around AI systems, while Agent Contracts and Guardian Agents extend the product further into operational enforcement.

Availability remains uneven across the portfolio. Maestro is available now, including previews of Maestro Studio and Maestro Assistant; Agent Contracts and Guardian Agents are scheduled through the AI Command Center in October, while Live Map is initially opening to a limited group of design partners.

Collibra has also obtained AIUC-1 certification for Maestro Assistant. The standard covers areas including data and privacy, security, unauthorised actions, monitoring, and human oversight, although certification of one assistant does not certify every customer-created agent or third-party system later governed through the platform.

Runtime control becomes the harder test

The practical evidence will emerge when organisations apply these controls to agents carrying out consequential work across several applications. A governance product can define a policy, but its usefulness depends on whether it can observe enough of an agent’s operating context to enforce that policy without creating so many interruptions that employees or developers look for ways around the controls.

False positives would create one failure mode, particularly if a Guardian Agent blocks routine activity that technically resembles a restricted action. Incomplete visibility creates another: an enforcement layer cannot reliably supervise a decision if parts of the agent’s tool chain, identity context, or data flow remain outside its view.

There is also a commercial question around interoperability. Enterprise customers are unlikely to operate one homogeneous fleet of AI systems, leaving governance suppliers to support agents built through different model providers, application platforms, orchestration tools, and internal frameworks if they want to become a common control layer.

Collibra’s launch therefore represents a product direction rather than evidence that runtime agent governance has been solved. Yet the underlying requirement is becoming clearer as AI systems acquire permission to do more than produce text: organisations need controls that establish what autonomous software is allowed to do, recognise when it leaves those limits, and retain an auditable record of the actions it attempted on the organisation’s behalf.