*Neo Control unified the functions inside an organization. Opening AI Control Architecture (ACA) is about preventing fragmentation across the industry. The industry does not need ten incompatible ways to control AI. It needs shared building blocks that governance, architecture, engineering, detection and red teams can all use.*
When I launched Neo Control, the objective was simple:
Help organizations adopt AI faster without losing control.
AI governance was becoming separated from AI implementation.
Governance teams wrote policies. Architects designed systems. Engineers built them. Security teams monitored them. Red teams tested them.
Each function saw one part of the problem. Very few organisations had a single view of how AI was being deployed, secured, tested and monitored from design through runtime.
Neo Control was built to bring those disciplines under one roof.
The aim was not more governance. It was end-to-end visibility, oversight and control across the AI lifecycle.
Governance defines the boundary. Architecture translates it into design. Engineering implements the controls. Detection monitors whether the system remains inside them. Red teams test whether those boundaries can be broken. Leadership sees the complete picture.
That is how secure AI adoption becomes faster, not slower.
But as the AI security market accelerated, I began to see the same fragmentation appearing at an industry level.
New products are emerging across AI governance, security, evaluation, guardrails, posture management, agent monitoring, red-teaming and runtime enforcement. Most are solving genuine problems.
But each product is also creating its own control language, risk model, schema, evidence structure and regulatory mapping.
The result is predictable.
A control defined by governance cannot be consumed directly by engineering. Evidence collected by one platform cannot travel easily to another. A test performed by a red-team product cannot always update the organisation's control posture. The same requirement is translated repeatedly for security, architecture, audit, customers and regulators.
We are building more capability, but also more fragmentation.
If every vendor creates a different way to describe the same AI authority and controls, organisations will spend the next decade integrating taxonomies instead of securing systems.
A new product cannot solve that problem. A shared foundation can.
The AI Control Architecture is the foundation beneath Neo Control. It provides a common way to describe what an AI system can see, decide and do, assign controls proportionate to that authority, and prove whether those controls work.
Today, I am making that foundation public and executable.
This release includes:
This is not simply a framework document.
People can read it. Engineers can implement it. Products can consume it. AI agents can call it. Organizations can adopt it. Vendors can build on top of it.
The purpose is to give the industry reusable building blocks instead of another closed vocabulary.
ACA starts with six questions:
These answers establish the operating boundary of the AI system.
A versioned rules engine then converts the structured inputs into one of five risk tiers and selects the required control baseline across ten pillars.
The principle is simple:
Same inputs. Same ACA version. Same risk tier. Same control baseline.
The control depth increases with the authority and potential impact of the system. A drafting assistant using public information should not carry the same burden as an agent that can modify customer records, approve a payment or deploy code.
The model may be identical. The authority is not.
ACA then connects that authority to controls spanning identity, data, inputs, decisions, tools, human accountability, testing, monitoring and containment.
The details are available in the architecture. The important point is that everyone works from the same control model.
A shared architecture allows every discipline to work on the same control boundary.
Governance can define what must be true. Architecture can translate it into system design. Engineering can identify what must be implemented. Detection teams can see what must be monitored. Red teams can test whether the boundary holds. Auditors can inspect the evidence. Runtime systems can enforce the control when the AI attempts to cross it.
The control does not change its identity as it moves between those teams. Neither does the evidence.
That is how we replace disconnected assessments with end-to-end oversight.
ACA expresses control strength through four states:
Declared → Evidenced → Verified → Enforced
A policy can declare that an AI agent must not access restricted customer data. Configuration can provide evidence that the boundary exists. Testing can verify that the restriction holds. Runtime control can enforce it when the agent attempts to cross the boundary.
This is what "prove, do not assert" means in practice.
ACA requirements are crosswalked to:
A crosswalk does not create automatic compliance. It allows the same implemented control and its evidence to support multiple obligations.
Implement once. Test once. Evidence many times.
Because the control language should not belong to Neo Control or any other vendor.
If every product owns its own vocabulary, interoperability will always require translation. Evidence will remain trapped inside platforms. Organizations will be locked into whichever control model they adopted first.
That is not how an AI control ecosystem should develop.
The architecture, machine-readable controls, crosswalks and reference tools are therefore openly available for others to inspect, use and build upon.
The MCP interface matters for the same reason. AI agents should be able to query a common control architecture while a system is being designed, not wait for governance to review it after deployment. A development agent should be able to describe a proposed use case, retrieve the controls and evidence its risk tier requires, and surface them while the system is being built.
Control should become part of the build process. Not a gate at the end of it.
Opening ACA does not mean stepping away from Neo Control. It clarifies what Neo must earn.
ACA is the shared foundation. Neo is an operational implementation that brings assessment, architecture, engineering, evidence, testing, monitoring and runtime control into one platform.
But Neo Control should compete on how effectively it proves, tests and enforces those controls. It should not compete by owning the language used to describe them.
Other organizations should be able to use ACA, extend it and build their own products on top of it. That is how a shared control ecosystem begins.
I am not declaring ACA an industry standard. A standard is earned through adoption, implementation and challenge.
But standards do not emerge from intention alone. Someone has to place a complete, usable foundation into the open and invite others to build on it.
That is what I am doing now.
If you are building an AI governance, security, assurance or runtime product, inspect ACA before creating another private control model.
Use what works. Challenge what does not. Extend it where the architecture is incomplete.
But let us stop rebuilding the same foundations in different languages. AI control should become an interoperable ecosystem, not another collection of walled gardens.
The architecture is public. The controls are machine-readable. The MCP interface is ready. Neo is on AWS Marketplace. The Community Edition is live.
Now we can build on the same ground.
*Notes*
*The AI Control Architecture and its control catalogue are licensed under CC BY 4.0; the ACA Tools (MCP server) code is Apache-2.0. The "AI Control Architecture" name and conformance marks are reserved. Framework and regulation names referenced here are the property of their respective owners.*
*This is version 0.1.0, an early, versioned release intended to be adopted, challenged and improved. Crosswalks indicate how controls map to external obligations; they do not by themselves establish compliance with any law, regulation or framework. Nothing here is legal or regulatory advice.*
*The AI Control Architecture is stewarded by Neo Control (*neocontrol.ai*).*