The Control Plane Is Not the Data Layer - The Difference Will Define Your AI Governance Strategy

In the past two weeks, four major enterprise platforms announced that they are "the control plane for your AI." Snowflake. Databricks. Microsoft Fabric. Palantir. Each framing is directionally consistent: unified data, semantic layer, governed context for your agents. Each one means the same thing - grounding agents in clean, permissioned, structured data so they know the right things before they act.

That is real infrastructure. It matters. No serious AI governance argument dismisses it.

It is not the control plane.

The data layer tells an agent what the world looks like. The control plane determines what the agent is allowed to do about it, and produces a record proving it did only that. Forrester's formal category definition is precise on this point: an agent control plane provides governance that is independent of individual runtimes, sitting above both the build layer and the orchestration layer to enforce policy, maintain a live agent inventory, and produce auditable outcomes. As Forbes describes it: "The control plane is about whether agents are allowed to run, under what constraints and with what evidence."

The market has saturated the data governance conversation. The action layer remains ungoverned.

Concrete example. An AI agent detects a churn risk in your CRM - a customer flagged at 74% likelihood to cancel within 30 days. Here is every decision the data layer has already made correctly: the customer record is clean, the risk score is accurate, the agent has permission to read it.

Now the control plane takes over.

Permission check: Is this agent authorized to send a retention offer to this customer segment? The business rule says yes, but only for discounts below 15%. The request is for 20%.

Escalation routing: The agent doesn't proceed. It routes to a human approver - the account manager or a revenue operations lead with context, recommendation, and a decision record attached.

Execution gate: The human approves a modified 12% offer. The agent executes only the approved action. No freelancing. No probabilistic drift.

Audit trail: A signed, timestamped record of the intent, the rule, the human decision, the action taken, and the outcome is written per decision, not per session.

That sequence: permission, escalation, execution, signed record is what a control plane governs. The data platform told the agent the customer was at risk. The control plane decided everything that happened next. UiPath Maestro has built toward this architecture with its explicit framing that business rules define what "the agent is not allowed to do" deterministic guardrails as the governance layer validating the category demand, if not the cross-system scope.

PIF's position is sharper: your business rules are your IP. They live in your control plane not in a vendor's model, and not in a platform you don't control.

As one editorial observation in the CIO space has crystallized it: no enterprise should outsource its control plane to anyone.

The EU AI Act's full enforcement window opened August 2, 2026. This is not a soft launch. The AI Office now holds full enforcement powers, and non-compliance with Article 5's prohibited AI practices carries fines up to €35 million or 7% of global annual turnover the highest regulatory ceiling in the world, exceeding GDPR's 4% cap.

Article 14 is the provision every enterprise with EU operations needs to read carefully. It requires that high-risk AI systems be designed to enable human oversight during operation specifically, the ability to monitor, understand, disregard, override, and stop agent outputs in real time. More demanding still: the compliance record must be reconstructable per decision, on demand.

The practical test is this: can your organization produce a governance record for a specific agent decision not a log export, not a session summary on demand, without a third-party request? If the answer involves your data platform, you are looking at the wrong layer.

Legal analysis from Latham Watkins frames the current window as - an extension of time, not a material relaxation - the obligations have not changed, only the runway to meet them has shortened. CNIL, France's data protection authority, has been the only European DPA actively publishing operational guidance in-window, signaling that formal enforcement scrutiny is building.

The enforcement clock is running. Governance-in-theory does not satisfy Article 14. Governance-in-the-runtime does.

Gartner's warning is precise: enterprises applying uniform governance across all agents will see 40% of those agents decommissioned or demoted by 2027. The prescription is equally precise, match governance to autonomy level, not to a binary trusted/untrusted switch.

This is what Gartner calls "differentiated governance," and it is the emerging operational requirement.

A reporting agent that summarizes pipeline data is not the same governance problem as a customer-action agent that can execute a discount, modify a contract term, or send a communication on your behalf. Different risk profiles. Different escalation thresholds. Different audit obligations. Different regulatory treatment under Article 14.

The four tiers Gartner has defined: observe, advise, act with approval, fully autonomous. Each carry different governance weight. An agent classified as "act with approval" that holds write credentials is, in practice, a fully autonomous agent wearing the wrong badge. The audit trail and escalation rules need to reflect what the agent can do, not what it usually does.

Uniform governance fails in both directions: it over-governs low-risk agents and kills productivity, or it under-governs high-risk ones and creates liability. The AI control planemust enforce different rules per agent type, per use case, per regulatory context without requiring your team to rebuild every workflow from scratch.

The EU AI Act compliance obligation is not an architecture problem you solve once. It is an operational posture you maintain per decision, in perpetuity.

Put It Forward's AI control plane governs what agents do in the runtime, per decision, per business rule, per regulatory requirement - across every system you already use, without replacing any of it. It sits above the agent layer and above the orchestration layer, exactly where Forrester defines the category should sit. It applies differentiated governance, different rules per agent type, per use case, per risk profile without requiring manual configuration per workflow.

The data foundation you've built is not wasted. It becomes more valuable when the action layer is governed. Agents grounded in clean data, operating under a runtime control plane, produce outcomes that are auditable, defensible, and compliant. The enforcement clock is running. The question is whether your governance is.