What is a control plane?

The term came out of networking. It now decides how much AI an enterprise can actually put into production. Read it three times below: once in plain words, once for the people who will run it, and once for the people who will own it.

The definition

One layer between everything that calls a model and everywhere a model can run. It decides where each request runs, under the configuration you set, and signs the record.

At SAPTIVA AI, that layer is frIdA.

  • Decideswhere each request runs
  • Measureswhat it costs, per team
  • Provesevery decision, signed

The tower, not the plane.

Think of the tower at an airport. The planes are not the tower's. It does not fly them, does not own them, could not carry one passenger across the street.

It still runs the airport. It decides which plane lands on which runway. It keeps a log an inspector can read three years later without asking a single pilot to remember anything.

Models are the planes. Vendors will keep shipping faster ones and you will keep swapping them out. The tower is the part you cannot rent from the people who build the planes.

That tower is a control plane.

For engineers

Same split your infrastructure already runs on. In a network, the data plane forwards packets and the control plane decides where packets go. In a cluster, containers do the work and the control plane decides which machine runs them. AI is the third time this shows up: inference is the data plane, and something above it has to own placement, cost and proof.

For business

Same split your business already runs on. Outside firms do the work, and procurement decides which firm gets which job. Vendors send the invoices, and your books charge each one to a cost center. Models are suppliers now. The control plane is the part you own: which model runs each request and where, the cost per team, and a signed record of every decision.

frIdA CONTROL PLANE Decides · Measures · Proves a979·45b9 ✓ signed Cloud On-premise Air-gapped Hybrid YOUR ENVIRONMENTS · YOUR MODELS · YOUR DATA
The control plane sits above the places where work actually runs. It never becomes the model, and the model never becomes the control plane.

Three jobs.
Nobody sits in front of it.

A control plane is not an application your team opens in the morning. It stands in the path of every request, whether anyone is watching or not.

Decides

Every request has to run somewhere: a public cloud, your own datacenter, a sealed room with no cable to the outside. The control plane makes that call per request, under the configuration you set once, instead of leaving it to whoever wired the first integration.

Routing at call time. Model, environment and cost chosen per request. Swap a model or a region without touching application code.

Measures

Every request is priced as it runs and attributed to the workload and the team that sent it. AI spend stops arriving as one invoice at month end and becomes a number somebody owns.

Cost metered per request, per workload, per team. The number finance reads is the same one engineering sees.

Proves

Each decision is written down and signed as it happens. When an auditor asks what the AI did in March, somebody reads the answer. Nobody reconstructs it.

Append-only record. Every entry references the one before it, so any edit breaks the chain. Read-only view for audit, exportable to your SIEM, stored where the workload ran.

Inside for what must stay inside.
Cloud for everything else.

A typical hybrid deployment: sensitive data classes run inside your perimeter, everything else on cloud capacity. One control plane decides, measures and proves across both.

WorkloadDataRuns onWhy
Customer identity checksKYC documentsOn-premiseSet once as a sensitive data class. Runs inside your perimeter.
Customer-facing agentsData not set as sensitiveCloudEverything else routes to the cheapest capable compute.
Internal knowledge systemNon-sensitive corporate documentsCloud burstBursting compute needs. Same configuration, same audit trail.

Inside the perimeter or out, one control plane: one configuration applied on every route, one signed record, and every request priced and attributed to the team that sent it.

Deloitte, Aug 2026, on the hybrid paradigm: a central control plane governs identity, audit and cost while execution stays distributed. MultiMoney (financial services) runs hybrid today: on-premise and private cloud. Six countries. One configuration.

Three different things carry the same name.

Most products sold as AI control planes are one of the first two columns. The difference is not marketing. It is who ends up owning the decision.

AI gateway Governance proxy Control plane
What it is A checkpoint in the request path A watchtower over models you rent The authority over where AI runs
What it decides Which vendor endpoint answers Whether a request is allowed through Which model, which environment, at what cost, per request
Where the models live Wherever the vendor hosts them Wherever the vendor hosts them Your cloud, your datacenter, air-gapped, hybrid
What it leaves behind Traffic logs Alerts and reports A signed record of every decision
If the vendor disappears The checkpoint goes with it The oversight goes with it It keeps running inside your perimeter

A gateway is a checkpoint. A control plane is the authority.

Both stand in the path. Only one of them decides where the work runs, and only one of them outlives the company that sold it to you.

Where the money shows up.

Control is not the point. The point is AI that reaches the P&L. Each of the three jobs above turns into a line somebody in finance can read.

95%

of enterprise AI pilots never reach the P&L.

Source · MIT
20%

of workloads are moving from global providers to local ones. Every move is a routing decision.

Source · Gartner
3weeks to 5minutes

Savant in production. Sector data and the bank's own book, in one question.

Banco Invex
TodayWith a control plane
A platform program runs three quarters before anything touches a customer.
One real workload goes live in weeks, and the next one reuses the same layer.
AI spend arrives as an invoice at month end, with nobody's name on it.
Every request is priced and attributed to the team that sent it.
An audit request pulls engineers off the roadmap to reconstruct what happened.
The evidence is already written and signed. Answering is a read, not a project.
A better model appears and the integration work starts over.
Swapping the model is a routing change. The configuration and the record stay where they are.

A pilot that never ships is not a cheap experiment. It is a paid line item with nothing on the other side of it.

Four questions.
No technical answer required.

Any executive can answer these about their own institution today. An auditor or a board member will ask them first.

01

Can you say where each AI request ran last Tuesday, and why it ran there?

02

Can you move one workload from a public cloud to your own hardware without rebuilding it?

03

Can your auditor read what the AI decided without asking your team for a screenshot?

04

Can you name the cost of a single AI decision, attributed to a single team?

Four times no is not a tooling gap. It is the reason the pilot never left the lab.

In production at RAPPI MULTIMONEY BANCO INVEX IBERO ACMES

Questions people ask.

What is an AI control plane?
One layer between everything that calls a model and everywhere a model can run. It decides where each request runs, under the configuration you set, and signs a record of what happened. It does not produce answers. It governs the systems that do.
Is an AI control plane the same thing as an AI gateway?
No. A gateway stands in the request path and forwards calls to model endpoints, almost always endpoints somebody else hosts. That is data plane work. A control plane decides which environment a workload runs in at all, measures what it costs, and leaves signed proof behind. A gateway is a checkpoint. A control plane is the authority.
What is the difference between a control plane and a data plane?
The data plane does the work: it moves the packets, runs the containers, executes the inference. The control plane decides what the data plane is allowed to do and where. The split is decades old in networking and standard in cloud infrastructure. AI inherited both the split and the vocabulary.
Do I need one if everything already runs in a single cloud?
Today, maybe not. The question is how long that stays true. A workload arrives carrying data that cannot leave the building. A cost line becomes visible to the CFO. Or the model you standardized on gets beaten by one your provider does not host. A control plane turns each of those into a routing decision instead of a rebuild.
How does a control plane impact my P&L?
As AI adoption grows, so do the requests, the tokens and the spend. A control plane turns that spend from one invoice into lines finance can read. Every request is priced as it runs and attributed to the workload and the team that sent it, so the growth has a name on it. Decisions are signed as they happen, which keeps audits off the engineering roadmap. The next workload reuses the same layer, with the cost metering and the signed record already there.
Does a control plane replace my models?
The opposite. It assumes models keep changing and makes the change cheap. Open weights, commercial APIs, models your own team fine-tuned: all of them sit underneath, and the configuration that routes them stays in one place. Models are the part that moves. The switch is the part that does not.
Where does the record live, and who gets to read it?
Where the workload ran, inside your perimeter. It is append-only, and each entry references the previous one, so any edit breaks the chain and the break is visible. Auditors get a read-only view they can export, without going through your engineering team for evidence.
How long does it take to put one into production?
Weeks, not quarters, when the scope is one real workload instead of a platform program. Banco Invex runs Savant in production: sector data and the bank's own book in one question, 3 weeks to 5 minutes. The usual first door is a five-day Discovery on a single workflow that is stuck.

Bring the pilot that never reached production.

Five days, your data, inside your perimeter, on frIdA. You leave with a signed record of every decision it made, and it keeps running whether you hire us or not.

Start a five-day Discovery → An engineer, not a rep, replies within 48 hours.