Enterprise AI control plane architecture

The failures that create demand for a control plane tend to look similar across organizations:

• AI tools get adopted without IT or governance involvement, because the barrier to spinning one up is low and the barrier to getting it approved feels high
• There's no complete inventory of what's actually running, so "how many AI agents do we have in production" doesn't have a confident answer
• Nobody has a clear record of what data a given agent can reach, which means nobody can say with certainty what it can't reach either

Per-application governance doesn't scale once more than a handful of teams are shipping their own AI systems. Reviewing each one individually might work for the first five. It doesn't work for fifty, and it becomes actively dangerous once agents are approving transactions, routing customer requests or touching regulated data without a consistent layer of control across all of them.

The core argument is simple: Policy written in a document is not policy until something enforces it at runtime. A written rule that says an agent shouldn't access certain customer data is a good intention right up until an agent actually queries that data and nothing stops it. The control plane is the thing that stops it.

A working control plane is built from six components. Each one does something specific, and each one leaves a specific gap when it's missing.

• AI system and agent registry. What it does: Maintains a complete, current inventory of every model and agent in production. What breaks without it: Shadow AI proliferates, and the organization can't govern what it doesn't know exists.
• Governed data and business context layer. What it does: supplies AI systems with trusted definitions, current data and applicable policy. What breaks without it: Agents guess at meaning and act on stale or ambiguous data.
• Policy, identity and access controls. What it does: Defines what each AI system is allowed to reach and do, tied to its own identity rather than a shared credential. What breaks without it: Access can't be scoped, revoked or attributed to a specific system.
• Risk assessment and approval workflows. What it does: Classifies AI use cases by risk and routes higher-risk actions to the right approval path. What breaks without it: Every use case gets the same level of scrutiny, which means high-risk ones get too little and low-risk ones get too much.
• Runtime monitoring and observability. What it does: Tracks what AI systems are actually doing once they're live, not just what they were approved to do. What breaks without it: Drift and unexpected behavior go unnoticed until something fails visibly.
• Audit evidence and reporting. What it does: Produces a defensible, continuous record of what was approved, what was accessed and what happened. What breaks without it: Every regulatory question or incident review becomes a forensic reconstruction project.

Rather than a narrative, this is a practical sequence a request moves through:

1. Register the AI system or agent. Nothing operates without being recorded in the inventory first.
2. Connect it to approved data, models and context. The system draws only from sources it's been explicitly connected to.
3. Assess risk and regulatory scope. The use case gets classified before it's allowed to act, not after.
4. Enforce policy and approvals at runtime. Every action gets checked against policy at the moment it happens.
5. Monitor behavior, drift and cost in production. The system is watched continuously once it's live, not just approved once and left alone.
6. Produce audit-ready evidence. Every step above generates a record automatically, as a byproduct of normal operation.

These terms overlap enough in casual use that it's worth being precise about what each one actually does.

The distinction worth holding onto is the same one that separates a network's control plane from its data plane.

• The control plane decides what's allowed and routes accordingly
• The data plane is where the actual traffic moves
• MLOps and LLMOps manage how models get built and shipped, not what they're allowed to do once they're live
• AI governance platforms can define excellent policy without ever touching the runtime moment where that policy either gets enforced or doesn't

One clarification worth adding: an AI management system under ISO/IEC 42001 defines the organizational processes and policies for how AI is developed and used responsibly. A control plane is the runtime layer that enforces those policies once systems are in production. They're complementary, not competing, and an organization can have a mature AI management system on paper and still have no way to enforce it if there's no control plane underneath it.

The control plane is where three disciplines converge in operation:

• Context governance produces the governed context the Control Plane depends on: data that arrives at the AI system already carrying its meaning, trust and policy. Without context governance, the Control Plane has nothing trustworthy to route.
• AI governance sets the enterprise-wide operating model: which use cases are approved, who owns them, what risk tier applies, what policies attach. The Control Plane is where those policies actually execute.
• AI agent governance enforces authorization, traceability, accountability and permissions at the moment an agent acts. The Control Plane is where authorization decisions get made and evidence gets recorded.

None of these disciplines works in isolation for enterprise AI. Context governance without a Control Plane produces trustworthy data that never gets enforced. AI governance without a Control Plane produces policy that stays on paper. Agent governance without governed context produces authorization decisions based on ungoverned inputs. The Control Plane is the architecture that connects the three.

A working evaluation checklist for any control plane, whether built or bought:

• Delivers both context and control from one connected platform, not two separate systems that require integration to work together
• Works across clouds, models and agent frameworks, not just one vendor's stack
• Discovers AI systems automatically rather than relying on manual registration
• Carries business context and definitions into the AI systems that consume it
• Enforces policy at runtime, not only at review time
• Monitors models and agents continuously once they're in production
• Generates audit evidence automatically, as a byproduct of operation
• Maps its controls to regulatory requirements as they apply
• Doesn't require moving data or locking the enterprise into one platform

• Governing AI agents and copilots in production
• Reducing shadow AI risk
• Standing up an operating model and clear ownership
• Governing AI across multiple clouds and platforms
• Proving compliance and audit readiness
• Extending data governance investments to serve AI

Collibra is the enterprise AI control plane. Built over eighteen years of ontology engineering, it governs context and control across any data, any model, any agent, for AI leaders who need certainty at the speed of AI.

Most AI governance failures trace back to one of two gaps:

• The context gap: Agents don't have governed access to what data means, whether it's current and what they're allowed to do with it, so they guess.
• The control gap: Once agents are in production, no one can see what's deployed, trace what it did or prove what happened if something goes wrong.

Closing one without the other leaves half the problem unsolved. Collibra addresses both.

• Collibra Platform delivers context governance. Governed data products carry their meaning, lineage, quality signals and policy constraints with them, so an agent's core questions about the data it touches are answered before it ever retrieves anything.
• AI Command Center delivers control at runtime. Every agent deployed is visible in real time. Every decision is traceable. Every drift can be caught before it becomes an incident. Assessment templates align to NIST and EU AI Act requirements, so the audit trail is a byproduct of how the platform runs, not a separate project.
• Deasy Labs, by Collibra extends governed context into the roughly 80% of enterprise data that's unstructured (contracts, tickets, transcripts, email threads), building the metadata and governed slices agents need instead of leaving them to parse raw content.

Because no enterprise runs on a single platform, Collibra is built to sit above the stack (governing across Databricks, Snowflake, SAP, Salesforce, AWS, Azure, Google Cloud and on-premise systems) rather than replacing the governance each of those platforms provides for its own environment. A definition of "customer" or "revenue" isn't owned by any single platform. It's owned by the enterprise, and it has to travel with the data wherever an agent reaches it.

• AI operating model for the enterprise AI control plane
• AI agent governance for safe production deployment
• Cross-platform governance, trust and business value
• Enterprise AI governance