An AI control plane is the layer that sets policy for an enterprise's AI models and agents and enforces it in the path of each prompt, response and tool call, before the action executes. It writes each decision, with the identity behind it, to a record a third party can verify.
The term comes from networking and Kubernetes, where the control plane decides and the data plane does the work. In AI, the data plane is the traffic itself: prompts, model responses, tool calls and messages between agents. The control plane decides whether that traffic should happen, under whose authority, and what record it leaves.
Forrester announced an evaluation of the agent control plane market in December 2025 and has since published The Agentic Control Plane Solutions Landscape, Q2 2026. Its announcement defines an agent control plane as "an enterprise control plane that inventories, governs, orchestrates, and assures heterogeneous AI agents across vendors and domains." This guide covers what an AI control plane does, where it runs, how it differs from an AI gateway, an AI firewall, observability and MLOps, and what regulated enterprises should require of one.
Updated October 2026.
Control plane vs. data plane in AI
The control plane decides; the data plane executes. Routers have separated routing decisions from packet forwarding for decades, and Kubernetes runs a control plane (the API server, etcd, the scheduler and the controllers) that decides what should run while the nodes run it.
| Control plane | Data plane | |
|---|---|---|
| Job | Decides what may happen and records what did | Carries the work |
| In AI | Policy, identity checks, approvals, budgets, the audit record | Prompts, responses, tool calls, agent-to-agent messages |
| Changes | When policy, people or regulation change | With every request |
| In Kubernetes | API server, etcd, scheduler, controllers | The kubelet and the containers on each node |
One consequence shapes every design decision. Enforcement has to happen in the data plane, at the moment of the call, while the policy behind it has to be the same on every call. Systems that read traffic afterward can only report violations.
What an AI control plane does
Six functions define the category.
| Function | What it means in practice |
|---|---|
| Inventory | Knows which agents, models, tools and MCP servers are in use, because their traffic passes through it or reports to it |
| Identity | Ties each request to the person behind it and, for an agent, to the person whose authority the agent is using |
| Policy | Evaluates rules before the action executes: which models, which data, which tools, which amounts |
| Approvals | Pauses a high-risk action until a person with approval rights decides |
| Evidence | Writes a record of each decision that a third party can verify |
| Cost | Attributes spend to teams, agents and workloads, and checks budgets before spend occurs |
Inventory and identity establish what is running and for whom. Policy and approvals decide what it may do. Evidence and cost show what it did and what it spent.
Our glossary covers the building blocks: runtime governance, policy as code, agent authority scope, human in the loop and the tamper-evident audit log.
Why enterprises need one now
Agents act on delegated authority
An agent calls tools, changes records, moves data and starts payments on someone's behalf. Each of those actions uses authority a person delegated, so every incident review starts with two questions: whose authority was it, and who approved the step that went past it. The CISO's guide to runtime governance and the agent identity guide cover how agents get identities and scopes.
Agents run on more than one platform
Microsoft, Salesforce, AWS, Google, ServiceNow, Glean and CrewAI each ship governance controls with their agent platforms. Our guides to Microsoft Copilot, Agentforce, Amazon Bedrock AgentCore, Glean, CrewAI and Google, ServiceNow, OpenAI and Anthropic map what each one controls. An enterprise with agents on several platforms gets separate controls and separate logs for each. Governing AI agents across platforms sets out where one control point for all of them sits.
Forrester draws the same line. It describes the agent control plane as "the third functional plane in an enterprise agentic architecture," separate from the tools that build agents and the systems that orchestrate them.
Examiners ask for records
An examination asks what happened, for which customer, under whose authority, and who approved it. Records written at the moment of decision, stamped with a trusted time and checkable by a third party, answer that. See the FINRA Rule 3110 supervision guide, the EU AI Act compliance guide and examination readiness.
Where an AI control plane runs: four patterns
| Pattern | Where it sits | Can it stop an action before it executes? | What it misses |
|---|---|---|---|
| Inline gateway or proxy | In the call path between applications or agents and the models and tools they reach | Yes | Traffic that routes around it, unless egress controls make it the only path |
| SDK or library | Inside each application's code | Yes, in the applications that adopt it | Applications that do not adopt it; each policy change needs a release |
| Platform-native | Inside one agent platform | Yes, for that platform's agents | Agents and models on other platforms |
| Out-of-band | Reads logs, traces and configurations after the fact | No. It reports | Anything not sent to it, and anything that already happened |
The patterns combine. The design question is which layer is authoritative: where policy is written once, and which enforcement points apply it. An inline control point gives the strongest guarantee when egress controls make it the only route to models and tools. The enterprise AI gateway guide covers inline deployment in detail.
AI control plane vs. AI gateway, AI firewall, observability and MLOps
| Term | What it is | How it relates to the control plane |
|---|---|---|
| AI gateway | A proxy in the path to models that routes, authenticates, meters and caches | The enforcement point a control plane uses at the model layer |
| MCP gateway | A proxy for agent-to-tool traffic over the Model Context Protocol | The enforcement point at the tool layer |
| AI firewall | Inspection of prompts, responses and tool calls for attacks and data loss | One policy domain inside the control plane |
| Observability | Traces, logs and evaluations of what models and agents did | Evidence and feedback, after the fact |
| MLOps | Tooling to build, test, deploy and monitor the enterprise's own models | Covers the models the enterprise builds; the control plane governs runtime use of any model or agent |
| AI governance platform | Inventories, risk assessments, policies and approvals for an AI program | The source of policy, which the control plane applies at runtime |
| Agent platform governance | Controls built into one agent platform | A control plane for that platform's agents |
More on each: the enterprise AI firewall guide, MCP security and AI guardrails. For how runtime enforcement differs from workflow and identity governance, see the runtime plane vs. the workflow plane.
What regulated enterprises should require
Banks, insurers, healthcare providers and defense contractors answer to examiners and auditors. Seven requirements separate a control plane that stands up in an examination from one that does not.
- Policy decided before execution. Deterministic rules applied in the path, before the model or tool receives the call.
- A person on every record. Each request ties to the person behind it. For an agent, that is the person whose authority it is using, and the agent's own scope stays inside that person's.
- A person decides exceptions. A high-risk action pauses until someone with approval rights decides, and the decision becomes part of the record. The approver can be required to re-authenticate first.
- Evidence a third party can verify. Hash-chained records with trusted timestamps under RFC 3161, checkable without the vendor's software.
- Deployment inside your environment. On premises or in your own cloud account, so the policy engine, logs and audit records stay with your other regulated data. Prompts still leave when routed to an external model or embedding service. For air-gapped deployment, confirm support for a local model, local embeddings and your own timestamp authority.
- More than one model provider. Routing and ordered failover across providers, so one provider's outage or policy change does not stop the business.
- Records mapped to your frameworks. Evidence organized against the rules you report under, such as FINRA Rule 3110, NYDFS Part 500, the EU AI Act and ISO/IEC 42001.
Seven questions to ask a vendor
- Does it govern MCP tool calls and agent-to-agent messages as well as model calls?
- Does it govern agents built on platforms you do not control?
- Where does it run, and which components call outside services?
- Are budgets enforced before spend, and attributed to teams and agents?
- How are policies versioned, tested and rolled back?
- If the control point is unavailable, does traffic stop or pass unchecked, and who chooses?
- Which frameworks does it map evidence to, and who keeps those mappings current?
How APERION builds it
Smartflow is APERION's AI control plane for regulated enterprises. It runs on premises in the call path between applications, agents and models, and enforces policy on each prompt, response and MCP tool call that passes through it, before the action executes. It validates the OIDC ID token that Microsoft Entra ID or Okta issued and carries that identity into the record. With agent identity switched on, each agent's scope is constrained to a subset of the delegating person's. Critical violations are blocked; high-risk actions pause as approval tickets until a person with approval rights decides. Every decision is sealed into a hash-chained record with an RFC 3161 timestamp from a public authority, DigiCert by default, or the customer's own, and the customer can verify the record independently. See Smartflow or read the documentation.
Frequently asked questions
What is an AI control plane?
An AI control plane is the layer that sets policy for an enterprise's AI models and agents and enforces it in the path of each prompt, response and tool call, before the action executes. It writes each decision, with the identity behind it, to a record a third party can verify.
What is the difference between an AI control plane and a data plane?
The data plane carries the work: prompts, responses, tool calls and messages between agents. The control plane decides whether each of those should happen and under whose authority, and records the decision. Enforcement runs in the data plane at the moment of the call. The policy and the record live in the control plane.
Is an AI gateway an AI control plane?
Not on its own. An AI gateway enforces policy on model calls. A control plane also governs tool calls, agent identity, approvals and evidence, and uses gateways as its enforcement points.
What is the difference between an AI control plane and an MCP gateway?
An MCP gateway is the enforcement point for agent-to-tool traffic. A control plane sets its policy, applies the same policy to model calls, and keeps one record across both.
How does an AI control plane differ from MLOps?
MLOps manages how an enterprise builds, tests, deploys and monitors its own models. An AI control plane governs runtime use of any model or agent, including hosted models the enterprise did not build: who may call it, with which data and tools, under whose authority, and with what record.
What is an agent control plane?
Forrester uses "agent control plane" and "agentic control plane" for the market. Its announcement defines an agent control plane as "an enterprise control plane that inventories, governs, orchestrates, and assures heterogeneous AI agents across vendors and domains." The terms overlap. An AI control plane also covers model calls from applications that are not agents.
Do I need an AI control plane if my agent platform has governance built in?
If every agent runs on one platform, start with that platform's controls. An enterprise with agents on several platforms and frameworks gets separate controls and separate logs for each. A control plane applies one policy and keeps one record across all of them.
Can an AI control plane run on premises?
Yes. Inline control planes can run on premises or in the customer's own cloud account, which keeps the policy engine, logs and audit records inside the enterprise. Prompts still leave when routed to an external model or embedding service. An isolated network needs a local model, local embeddings and the customer's own timestamp authority.
What evidence should an AI control plane produce for an examination?
A record for each decision: who made the request, which agent and whose authority it used, which policy version applied, the outcome, who approved any exception, and a trusted timestamp. The records should chain together so that a third party can check whether any record was altered or removed.
Sources
- Forrester, Announcing Our Evaluation Of The Agent Control Plane Market, December 2025
- Forrester, Agent Control Planes Still Need A Robust Standards Stack
- Forrester, The Agentic Control Plane Solutions Landscape, Q2 2026 (subscription report)
- Kubernetes documentation, Cluster Architecture
- IETF, RFC 3161: Internet X.509 Public Key Infrastructure Time-Stamp Protocol (TSP)
Related reading
Put this in the path of your own agents.
Policy enforced inline between your agents and every model and tool they reach, with a record bound to the human who owns it.
Request a Demo Read the docs