AI Control Architecture

An open-source, vendor-agnostic model for governing, securing, assuring, and containing enterprise AI.

The thesis, in one line: the next major AI failure in the enterprise will not come from a model becoming evil, it will come from giving a probabilistic system deterministic authority over data, decisions, or actions without a control architecture. This is how you build that architecture.

The architecture is organized around six control questions:

What AI exists?
What can it see?
What can it decide?
What can it do?
Who is accountable?
How is failure evidenced and contained?

These resolve into ten control pillars and a supporting operating model.

How the documentation is structured

01–06 Foundation what control means and why
07–16 Pillars the ten control domains, in depth
17–25 Operating how to implement, assess, run, and audit it
26 Glossary shared definitions

Foundation (01–06)

#	Document
01	Executive Summary
02	Introduction
03	Core Thesis
04	Architecture Principles
05	Reference Architecture
06	Requirements Catalogue

The ten pillars (07–16)

#	Pillar	Question
07	AI Inventory & Classification	What AI exists?
08	AI Identity & Access Control	Under what identity / access?
09	Data Boundary Control	What can it see?
10	Prompt & Input Control	What shapes it?
11	Output & Decision Control	What can it decide?
12	Tool & Action Control	What can it do?
13	Human Accountability Model	Who is accountable?
14	AI Assurance & Testing	Do the controls hold?
15	Monitoring, Logging & Evidence	Can we reconstruct it?
16	Incident Containment & Recovery	Can we stop and recover?

Operating (17–25)

#	Document
17	Implementation Checklists
18	Control Maturity Model
19	Common AI Control Patterns
20	Common Failure Scenarios
21	Adoption Playbook
22	Governance & Operating Model
23	Metrics & Reporting
24	Assurance & Audit Guide
25	Triage & Minimum Controls

Reference

#	Document
26	Glossary

Supporting assets (repository root)

templates/, reusable intake, risk-tiering, per-pillar control, assurance, evidence, and incident templates.
mappings/, standards crosswalks (NIST AI RMF, ISO/IEC 42001, EU AI Act, SR 11-7, NYDFS Part 500, OWASP LLM & Agentic, US state AI laws).
examples/, worked use-case examples.
../QUICKSTART.md, a single AI use case walked from intake to evidence.

Where to start

New to the architecture → read the foundation in order: 01 → 02 → 03 → 04 → 05, then skim 06.

Implementing controls → start with the Quickstart, then 17 (checklists), 21 (adoption playbook), 22 (governance).

Assessing maturity → 18 (maturity model) + 23 (metrics) + the maturity-assessment template.

Understanding a control domain → the relevant pillar (07–16).

Understanding how AI fails → 20 (failure scenarios).

Testing or auditing → 24 (assurance & audit) + the assurance and evidence templates.

Just need the minimum → 25 (triage & minimum controls).

Contributing, license, and the name

The AI Control Architecture is open and openly licensed under CC BY 4.0: use and adapt it freely, including commercially, with attribution. The name and any conformance marks are reserved, see Trademark & Conformance. To propose changes, see Contributing.

Version 0.3.0 · Architecture licensed under CC BY 4.0 · Stewarded by Neo Control · neocontrol.ai