Definition · Architecture
What an AI control plane is.
An AI control plane is the infrastructure layer that sits between every application in an organisation and every model it calls, providing centralised governance, observability, cost attribution and policy enforcement.
In short
Not a router: the system of record.
Control plane
Data plane
RequestTokensResponse
Where the term comes from
Borrowed from networking, not invented.
In AI
The data plane
The model call itself: the request body, the tokens, the response.
In AI
The control plane
Everything that governs the call: who may make it, against which provider, under which policy, at what cost, with what audit trail.
A gateway lives on the data plane: it forwards requests. A control plane is a category above. It contains a gateway as one of its components, the way a network's control plane contains routing tables, but it can't be reduced to that one part.
What it includes
Four capabilities define the category.
A cost ledger
Per-request attribution that can be enforced, not just reported. A log says what was spent. A ledger says what was spent, by whom, against which budget, and what happens at the threshold. That is the difference between observability and governance.
Policy enforcement
Guardrails that run inline on every request, not advisory dashboards. A control plane decides whether the call proceeds, not whether to flag it after the fact.
Observability as a system of record
Every request kept as an immutable, signed event: an artefact an auditor can rely on, not a screenshot of a Grafana dashboard.
Translation by construction
Native support for every major SDK dialect in both directions, so the application keeps speaking the SDK it was written against and the control plane handles the wire format the chosen model speaks.
These four aren't features; they're the definition. A system with three of them is in an adjacent category.
How it differs
Gateway, proxy, SDK and control plane.
| An AI gateway | Routes requests from one or many applications to one or many providers. It typically handles fallbacks, basic rate limiting and some translation. Its concern is getting the request to the right place: a data-plane object. |
|---|---|
| A proxy | A thinner thing, usually one application's outbound hop to one provider, with logging bolted on. The team running it owns it as code. Proxies are tactical: they answer a question for one team. |
| An SDK | Client-side: the library the application uses to speak to the gateway, proxy or model. It isn't infrastructure; it's the code-shaped end of the request. |
| A control plane | Sits above all three. It contains the gateway as one of its functions, replaces the proxy (one shared boundary instead of many bespoke ones), and is what the SDK ends up speaking to. Where a gateway forwards, a control plane governs. Where a proxy is a tactical hop, a control plane is a strategic boundary. Where an SDK is client code, a control plane is the operational system that client code reports into. |
The shortest way to put it: a gateway is what makes AI work. A control plane is what makes AI accountable.
Why the distinction matters
A quiet threshold, crossed in 2026.
Before the threshold
Most organisations cross it when they have more than one AI-using application, more than one model provider, and at least one team accountable for cost or compliance. Before that, with one application, one provider and one team, governance and routing are the same question, and a gateway answers both.
Past it
They separate, and routing becomes a small part of the problem. The larger part: which team's budget did that request hit, which policy applied, which auditor will see what record of it, and what happens when the regulation changes next quarter. A gateway has no representation for any of this. A control plane has it as the central object.
Why now
Enterprises whose whole AI strategy was built around a gateway are crossing the threshold and finding they lack the operational layer they need. The control plane is what they end up building anyway, badly, over eighteen months, or what they buy as a category.
How we start together
Start with one team, on your own AI traffic.
- 01
Pick one app or team
Point it at the Control room. One configuration change; nothing else changes.
- 02
Watch for two weeks
Every call priced, checked and recorded. Rules start in observe mode.
- 03
See the evidence
What it cost, what it caught and what it would have stopped, from your own traffic.