Questions fréquentes

Chapter 2: What Is Project NoéMI?

## Chapter 2: What Is Project NoéMI? Project NoéMI is a **global AI fluency accelerator program** delivered by **NewPush®** and academically credentialed by **George Mason University**. But that sentence, while accurate, undersells what it actually is. Here is a better way to think about it: NoéMI trains *organizations* — not individuals — to build, govern, and deploy what we call a **Virtual Workforce**. That Virtual Workforce is made up of AI agents — structured, governed, auditable AI systems that perform repeatable work alongside your human team. It is **not** a prompt engineering course. It is not a "learn ChatGPT in a day" workshop. It is an enterprise AI transformation program that begins with cybersecurity, runs through structured training, and ends with certified, production-ready AI workflows governed by trained human leaders. The strategic aim is concrete: help enterprises convert AI from a labor-displacement fear into a productivity system. In practice, that means: - Enabling teams to produce more with the same people - Shifting skilled staff away from repetitive execution toward review, judgment, and strategy - Building the managerial capability to supervise AI d

A. "Security First, Then Speed"

### A. "Security First, Then Speed" Every building begins with solid foundations, and every AI architecture begins with **Phase 0 Security** — establishing the data perimeter before any AI is deployed. This is the opposite of what most AI programs do. Most start with "let's try prompting." NoéMI starts with "let's make sure your data cannot leak." The reasoning is simple: if you build AI capabilities on top of a compromised or ungoverned foundation, every efficiency gain you achieve also increases your risk surface. You move faster, but you move faster toward a breach. Phase 0 is covered in depth in [Chapter 11](#chapter-11-phase-0-security-the-foundation-you-cannot-skip).

B. "Architecture Over Tools"

### B. "Architecture Over Tools" Tools change monthly. The AI model that is state-of-the-art today will be surpassed next quarter. The workflow automation platform you pick will release breaking changes. The API you integrated will deprecate an endpoint. NoéMI does not teach tools. It teaches **durable mental models** — the 4 Ds framework, governance principles, delegation patterns, discernment skills — that remain valid regardless of which AI model or platform is current. NoéMI teaches how to keep learning and using new tools with ease, and how to design and build agent workflows that can be easily adapted to new tools and models. This reference architecture keeps being updated to reflect the latest changes in the AI landscape, so you don't have to wonder how to adapt to the constant onslaught of changes. This is your guide to stay on the swelling wave. The curriculum achieves this through a **Two-Layer System**: Layer A contains First Principles that never change, and Layer B contains Dynamic Labs that get updated per cohort, per tool release, or per quarter. You learn how to think about AI governance once. You learn the specific buttons to click in the current tool version t

C. "The Replacement Trap Is a Dead End"

### C. "The Replacement Trap Is a Dead End" This is perhaps the most important pillar, and the one many organizations get wrong. Armed with this architecture, you can help your organization steer clear of this trap. The **Replacement Trap** is the strategic dead end where organizations try to use AI simply to cut headcount. The logic seems sound on a spreadsheet: if AI can do the work of five people, fire four and pocket the savings. Here is why that logic fails: if your only competitive advantage becomes "cheaper labor via AI," a competitor with a slightly better algorithm, a slightly cheaper model, or a slightly more automated pipeline puts you out of business in six months. You have not built a moat. You have built a race to the bottom. NoéMI rejects this entirely. Instead, it restructures organizations into what we call the **1:50 Equilibrium** — a model where trained human leaders govern fleets of AI agents. This creates a **Virtual Workforce**: not a smaller workforce, but a *different kind* of workforce. One where AI handles the repeatable, data-intensive, first-pass work, and humans handle the judgment, the exceptions, the creativity, and the governance. Not fewer peop

Phase 0 Security Baseline — Client Guide Before Starting an Advanced AI Project

# Phase 0 Security Baseline — Client Guide Before Starting an Advanced AI Project Project NoéMI treats **Phase 0 Security** as the prerequisite for serious AI adoption. Before an organization launches copilots, agentic workflows, RAG systems, or autonomous integrations, it should first baseline its cybersecurity posture. The goal is not to become "perfect" before using AI. The goal is to understand your real risk, fix critical gaps, and establish safe operating boundaries. This guide is written from the **client's perspective**. It is meant to help leadership teams, IT managers, and operational owners ask better questions of their internal team, their MSP, or their MSSP before they fund an advanced AI initiative. This guide covers the **security half** of that conversation. The paired **AI readiness half** of the initial assessment lives in the [Phase 0 Assessment Kit](phase-zero-assessment/README.md), where security readiness and AI readiness are treated as separate executive questions.

3. DISCERNMENT — Critically evaluating AI outputs

### 3. DISCERNMENT — Critically evaluating AI outputs AI produces confident-sounding output regardless of whether it is correct. Discernment is the ability to evaluate what comes back — to catch hallucinations, spot reasoning failures, identify when the AI followed the letter of your instructions but missed the spirit. - **Product Discernment:** Evaluating output quality, relevance, strengths, and weaknesses - **Process Discernment:** Assessing whether the human-AI collaborative dynamic is actually productive - **Performance Discernment:** Evaluating independent AI behaviors over time; gathering feedback to improve

4. DILIGENCE — Taking responsibility for AI-assisted work

### 4. DILIGENCE — Taking responsibility for AI-assisted work You used AI to produce something. Your name is on it. Diligence is the ethical, legal, and operational discipline of taking full responsibility for AI-assisted outputs — including the parts the AI got wrong. - **Creation Diligence:** Ethical and legal best practices; awareness of biases, flaws, and stakeholder impacts - **Transparency Diligence:** Understanding expectations around disclosure; communicating AI involvement - **Deployment Diligence:** Verifying outputs; implementing safety checks; assuming responsibility for what ships For deeper methodology details, see [`docs/METHODOLOGY.md`](METHODOLOGY.md) and the lifecycle documents in [`docs/lifecycle/`](lifecycle/). ---

1. DELEGATION — Knowing WHEN and WHETHER to use AI

### 1. DELEGATION — Knowing WHEN and WHETHER to use AI This is where most organizations fail first. They either delegate everything to AI (dangerous) or nothing (wasteful). Delegation is the skill of looking at a task and determining: should this be done by AI, by a human, or by both in collaboration? - **Goal & Task Awareness:** Understanding task requirements; deconstructing work into AI, human, and collaborative components - **Platform Awareness:** Knowledge of AI capabilities and limitations; evaluating tools based on requirements and compliance - **Task Delegation:** Balancing AI and human capabilities; assigning tasks optimally The Refusal Principle lives here. The most important delegation decision is sometimes "no."

2. DESCRIPTION — Communicating requirements to AI effectively

### 2. DESCRIPTION — Communicating requirements to AI effectively Once you have decided to use AI, you need to tell it what to do — and how to do it — with precision. This is more than "prompt engineering." It is the discipline of translating human intent into machine-actionable instructions. - **Product Description:** Defining the desired output; translating vision into AI-understandable terms - **Process Description:** Dialogic prompting for iterative collaboration — working *with* AI, not just *at* it - **Performance Description:** Directive prompting to define future AI behaviors — configuring agents, not just chatbots

2. The Prepared Digital Environment (Montessori Approach)

## 2. The Prepared Digital Environment (Montessori Approach) Agent development within Project NoeMI is treated as a "synthetic workspace" applying Montessori principles: * **Uninterrupted Flow:** Digital labs and workspaces must be designed for deep, uninterrupted focus. * **Self-Correcting Tools:** AI feedback loops (e.g., automated test suites, MCP validation tools) should provide judgment-free error correction during the development phase. * **Intrinsic Motives:** Agent development projects should be driven by specific, real-world organizational challenges rather than abstract exercises. * **Collective Community:** Foster mixed-competency cohorts where natural mentorship occurs, bringing together domain experts and AI architects.

Producer–Reviewer Separation

### Producer–Reviewer Separation The rule that the identity which creates work cannot be the identity which approves it. Enforced technically rather than by policy: separate machine accounts, and a reviewer credential that cannot write code at all. The reviewer is not *asked* to refrain from editing — it is *unable* to.

Shadow AI

### Shadow AI The unauthorized, ungoverned use of public AI tools by employees who paste proprietary data — the organization's **Secret Sauce** — into unvetted platforms. Phase 0 exists specifically to eliminate Shadow AI.

The 1:50 Equilibrium

### The 1:50 Equilibrium The organizational ratio where **1 highly trained Accelerator (Pilot)** can effectively govern and stabilize the AI usage of **50 Explorers (Passengers)**. The name is inspired by Steve Jobs' 1995 observation about the 50:1 productivity ratio between great and average software engineers. In NoéMI, this ratio is not about individual talent — it is about architecture. You do not need 50 AI engineers. You need 1 Pilot and a structured system.

1. Gartner AI TRiSM Alignment

## 1. Gartner AI TRiSM Alignment To ensure institutional-grade AI orchestration, every agent must be evaluated against the following pillars:

The Refusal Principle

### The Refusal Principle The most important delegation skill is knowing when **not** to use AI. Some tasks, due to data sensitivity, ethical concerns, or the need for human empathy, must remain human. The ability to refuse — to say "this is not an AI task" — is not a limitation. It is a governance capability.

The Feynman Requirement

### The Feynman Requirement Named after Richard Feynman's famous principle: "If you can't explain it simply, you don't understand it." Accelerators cannot earn their badge unless they can teach a core concept to Explorers. This ensures true mastery and creates a self-sustaining ecosystem of trainers within the organization.

Skill: governed-loop

--- name: governed-loop description: "Run an agent loop that is aligned to a standing mission without becoming ungovernable: every iteration has a machine-checkable done-condition, a bounded budget, an escalation path, and an audit trail, and the loop's own outputs can never re-enter it as inputs. Use when the task matches this skill's Purpose and Inputs." license: FSL-1.1-Apache-2.0 metadata: author: project-noemi governance: "NoéMI 4D" --- > **Governance: NoéMI 4D** — this skill ships with Refusal Criteria, hard > `Ask First` / `Never` gates, and an audit-log contract, and passed > cross-model review before publication. > > **Generated file — do not edit.** Built from [`skills/orchestration/governed-loop.md`](https://github.com/project-noemi/agents/blob/main/skills/orchestration/governed-loop.md) > in [project-noemi/agents](https://github.com/project-noemi/agents) by `node scripts/generate_all.js`. > > **License:** Functional Source License, Version 1.1, Apache 2.0 Future > License (FSL-1.1-Apache-2.0) — see [LICENSE](https://github.com/project-noemi/agents/blob/main/LICENSE) > before redistribution or commercial use.