Framework · Open specification

AI Control Architecture

A ten-pillar framework for provable AI governance.

The AI Control Architecture (ACA) is a framework for governing artificial-intelligence systems by proving that controls are in place rather than asserting that they should be. It organises the full surface of AI risk into ten control pillars, each crosswalked to the prevailing governance standards, and pairs them with a structured assessment methodology that takes an AI use case from a plain-language description to a defensible, evidence-backed governance decision. Its organising principle is a single sentence: we do not ask whether a control is in place — we prove it.

The six questions it answers

For any AI use case — a copilot, an agent, embedded vendor AI, a RAG system, an internal LLM application, or a customer-facing model — the architecture answers six practical questions:

The ten control pillars

The framework organises all AI controls into ten pillars. Each is a coherent domain of risk with its own objectives, controls, and evidence. Together they span the lifecycle from knowing what AI exists to recovering when an AI system fails.

1
AI Inventory & Classification
Maintain a complete, current inventory of AI systems, models, datasets and dependencies, and classify each by purpose, capability and risk. Nothing can be governed that is not known.
2
AI Identity & Access Control
Govern the identities under which AI operates and the access it holds. AI actors require least-privilege, attributable and revocable access, like any other identity.
3
Data Boundary Control
Define and enforce the boundaries of the data an AI may read, retain and reveal — including retrieval, embeddings and memory — so it cannot surface or leak beyond its sanctioned scope.
4
Prompt & Input Control
Govern what enters the model: instructions, context and untrusted input. Addresses injection, manipulation and the integrity of the inputs that shape behaviour.
5
Output & Decision Control
Govern what the model produces and the weight its determinations carry — filtering, validation and the conditions under which an AI's output may be acted upon.
6
Tool & Action Control
Govern the actions an AI may take in the world — the tools, functions and integrations it can invoke, and the authority, constraints and approval conditions for AI-initiated action.
7
Human Accountability Model
Establish who is accountable for each AI system and decision, and where human judgement is required. Accountability is named, not assumed.
8
AI Assurance & Testing
Verify, through test plans with explicit pass criteria, that controls perform as intended. Assurance means tested, not asserted.
9
Monitoring, Logging & Evidence
Continuously observe AI behaviour and retain tamper-evident evidence of what the system did and how its controls performed. This pillar makes the others auditable.
10
Incident Containment & Recovery
Detect, contain and recover from AI failures and compromises, and restore a known-good state. Governance must assume that some controls will fail.

Crosswalked to the standards you already answer to

Each pillar and its controls are crosswalked to the principal AI-governance standards, so work performed once satisfies several frameworks simultaneously. A control is never expressed in isolation — always with its mapping — so evidence is portable across regulatory and audit contexts.

StandardRole in the crosswalk
NIST AI Risk Management FrameworkGovern / Map / Measure / Manage functions mapped per pillar.
ISO/IEC 42001AI management-system clauses and Annex A controls.
EU AI ActRisk-tier obligations for high-risk and GPAI systems.
OWASP (LLM & Agentic)Technical failure modes — injection, excessive agency, data exposure.

Sector crosswalks extend the same evidence to SR 11-7 (model risk), NYDFS Part 500, and emerging US state AI laws.

From plain language to a decision on the record

The architecture pairs the pillars with a structured assessment methodology. A use case starts as a plain-language description; it is assigned a risk tier; a proportionate and complete set of controls is selected across the ten pillars by tier and behavioural pattern; each control is expressed with its standards crosswalk; evidence shows the control exists and assurance (a test plan with explicit pass criteria) shows it works. The output is a defensible, evidence-backed governance decision — with scope-lock so risk cannot drift silently, re-assessment when the system or its dependencies change, and portfolio governance that rolls posture up across the whole AI estate.

How it relates to control planes and control matrices

An AI control plane enforces policy at runtime; an AI controls matrix (such as the CSA AICM) catalogs control objectives. The AI Control Architecture is the connective methodology between them: it decides which controls a specific use case actually needs, proves they work, and produces a decision on the record. Its enforcement counterpart — the Recursive Cyber Defense Framework (RCDF) — governs how those controls are observed and, where warranted, enforced in the path of an AI's actions at machine speed, under human authority.

Run it on the Neo platform Read RCDF →

The AI Control Architecture is a framework originated by Ankush Chowdhary and published by Neo (Neo Control Private Limited), which operationalises it as a platform. · neocontrol.ai