01
The definition
An AI governance control plane is the governing layer between every AI agent in an organization and everything those agents are allowed to touch. It gives each agent an identity and an owner, applies policy to each session, and keeps one audit trail across all of them.
The word agent is used broadly on purpose. A custom GPT a sales rep built in an afternoon is an agent. A Claude project with the finance spreadsheet attached is an agent. A LangGraph service that answers support tickets is an agent. A coding assistant with an MCP server connected to GitHub is an agent. The control plane has to see all of them, or the inventory is a guess.
02
Why it starts in the browser
Most AI use in a company happens in a browser tab: ChatGPT, Claude, Gemini, Copilot. Most of the people doing it are not developers and will never install an SDK. The gateways and proxies that work for code-built agents do not see a paste into a chat box.
A browser-first control plane puts a sensor where the agent acts. On a managed browser, an extension rolled out by policy observes what is typed, pasted and uploaded, classifies it on the device, and records a content-free finding. The company learns which agents exist and what reaches them without a single prompt leaving the device.
- Chrome and Edge cover most of the desktop; Firefox and Safari are available on the Enterprise plan. Rollout through Google Admin, Intune or Jamf takes about an hour.
- Mobile AI apps stay outside coverage; on the Enterprise plan, the desktop collector covers the ChatGPT, Claude and Copilot desktop apps. A control plane that is honest about this shows the gap as a number.
- Developers connect code agents through an SDK, an MCP server and a scanner, into the same inventory.
03
The four functions
Every control plane, whatever it is built around, ends up with the same four functions. They share one identity model, one policy model and one trail.
- Connect: agent identity. Every agent carries a persistent identity and a named owner. Observed agents are attributed to a device and a person until someone claims them.
- Secure: agent security. Every violation becomes a finding: personal data sent to an AI app, a limit crossed, a red-team check failed. Findings carry a lifecycle and are content-free by default.
- Govern: agent governance. A limit per agent in six dimensions, versioned, evaluated on every session, with a plain-language explanation when it is crossed.
- Observe: agent observability. Sessions, findings, cost and activity in one trail, exported as alerts, webhooks and an evidence package.
04
Deployment in an hour
A control plane that takes a quarter to deploy governs nothing in the meantime. The browser-first path is short: publish the extension policy, let devices enroll with an enrollment token, watch agents appear with owners as they are used. Nothing changes in the agents themselves.
Developers join the same day. The SDK instruments a framework in a few lines. The MCP server registers the coding assistant that spawned it. The scanner needs only an endpoint the team is authorized to test.
05
What leaves the device
Governance that reads employee prompts creates a privacy problem of its own. The right default is content-free: categories, counts, destinations, channels and HMAC fingerprints leave the device; prompt text, file contents and file names do not, unless a company opts in to redacted capture.
When a company needs more, capture should be opt-in, redacted on the device before it is sent, tied to a written lawful basis and a retention period, encrypted at rest and audited on every read. Employees should be able to see on their own machine whether capture is on.
06
How it maps to compliance
Frameworks ask the same three questions in different words: what AI is in use, what can it reach, and who approved each action. A control plane answers them as a by-product of governing agents.
- GDPR: lawful basis before any content capture, retention limits, audited access and content-free evidence by default.
- EU AI Act: support for risk classification, an inventory with an owner for every agent, logging and human oversight evidence for each agent.
- ISO/IEC 42001: a governed inventory, declared limits and a decision history for the AI management system.
- NIST AI RMF: govern, map, measure and manage, with fleet-wide visibility and runtime alerts.
- AIUC-1: evidence across the six domains, from data and security to accountability, prepared for an independent assessment.
- LGPD: the same privacy controls as GDPR, from lawful basis and retention limits to audited access, with Brazilian national identifiers such as CPF and CNPJ classified on the device.
07
How to evaluate one
Nine questions to ask any vendor, including us.
- Does it see the agents non-developers create, or only the ones engineers instrument?
- What exactly leaves the device, and can an employee see it?
- How long from zero to the first agent in the inventory?
- Can an agent have an owner before anyone declares it?
- Are limits per agent, versioned, and evaluated on every session?
- When a limit is crossed, does the alert explain why in one plain sentence?
- Which frameworks does the evidence package map to, and does the vendor claim certifications it does not hold?
- What is outside coverage, and is it shown as a number?
- Can you start free, before a procurement cycle?