acf.advisor
Conseil de gouvernance ACF® à partir d’un cas en texte libre — un paragraphe en entrée, une évaluation structurée en sortie : principes activés, dimensions critiques, niveau d’autonomie, risque, articles applicables, premières actions, risques opérationnels. C’est l’outil de conversion de la bibliothèque ACF® vers le rôle d’advisor.
requires_human_review vaut toujours true. Pour une qualification plus défensive, repassez via acf.classify-agent avec des énumérations qualifiées.Quand l’utiliser
Utilisez cet outil quand un porteur de cas (DSI, métier, conseil) arrive avec une description prose de son agent IA et a besoin d’un premier verdict structuré sur la gouvernance ACF® applicable : quels principes s’activent, quelles dimensions de maturité sont critiques, quel niveau d’autonomie est cohérent, quelles obligations réglementaires se déclenchent, quelles fiches mobiliser dans quel ordre. C’est l’outil « première étape » avant qualification rigoureuse.
Il répond à la question board-level : « par où on commence sur ce cas-là ? ». La sortie est conçue pour être servie à un comité IA ou à un DDAO qui décidera des étapes suivantes.
Paramètres d’entrée
Un champ obligatoire (la description du cas), trois champs contextuels optionnels qui resserrent la calibration.
case_descriptionstring (20-2000)requiredsectorstring (≤80)jurisdiction"eu" | "uk" | "us" | "ca" | "ch" | "br" | "jp" | "other"deployment_scale"pilot" | "department" | "enterprise" | "public"locale"en" | "fr"default: "en"Schéma de sortie
La sortie est un objet structuré avec niveau d’autonomie, niveau de risque, principes activés, dimensions critiques, fiches priorisées, articles applicables, premières actions, risques opérationnels, pied-de-page signé.
autonomy_level{ level: "N0"|"N1"|"N2"|"N3", rationale: string }risk_level{ level: "low"|"medium"|"high"|"unacceptable", rationale: string }activated_principles{ code: string, why: string }[]critical_dimensions{ code: string, why: string }[]priority_fiches{ code: string, order: number, why: string }[]applicable_articles{ regulation: string, article: string, why: string }[]first_actionsstring[] (≤5)operational_risksstring[]confidence"low" | "medium" | "high"assumptionsstring[]gaps_to_validatestring[]rationale_per_rule{ rule_id, rule_version, fired, evidence }[]requires_human_reviewtrueExemple d’appel
Un assistant RH interne qui propose une shortlist validée par un recruteur :
import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";
const transport = new StdioClientTransport({
command: "npx",
args: ["-y", "acf-mcp"],
});
const client = new Client({ name: "demo", version: "1.0" }, {});
await client.connect(transport);
const result = await client.callTool({
name: "acf.advisor",
arguments: {
case_description:
"We are rolling out an internal HR assistant that summarises candidate CVs, proposes a shortlist score, and writes the shortlist to our ATS. A recruiter validates each shortlist before any candidate is contacted. The agent reads structured candidate profiles (name, email, education) and uses a GPAI under the hood for the summarisation.",
sector: "human-resources",
jurisdiction: "eu",
deployment_scale: "department",
locale: "en",
},
});
console.log(JSON.stringify(result.content, null, 2));Réponse
{
"autonomy_level": {
"level": "N1",
"rationale": "The recruiter validates each shortlist before action — supervised execution applies. Inferred from 'recruiter validates' in the description."
},
"risk_level": {
"level": "high",
"rationale": "HR scoring + GPAI + EU jurisdiction triggers AI Act Annex III §4 (employment). Personal data is standard but the use case is recruitment-related."
},
"activated_principles": [
{ "code": "P1", "why": "Decision sovereignty applies to every agentic deployment, regardless of autonomy." },
{ "code": "P2", "why": "High-criticality decisions must be reconstructible after the fact." },
{ "code": "P4", "why": "Reinforced governance must match the criticality." }
],
"critical_dimensions": [
{ "code": "D4", "why": "Role allocation (DDAO + sign-off) is required for N1+ autonomy." },
{ "code": "D5", "why": "Regulatory compliance dimension is on the critical path." },
{ "code": "D3", "why": "Technical control (kill switch, observability) is non-negotiable." },
{ "code": "D2", "why": "Doctrine adoption is the entry gate before any agentic deployment." }
],
"priority_fiches": [
{ "code": "ACF-00", "order": 1, "why": "Read first: framework introduction." },
{ "code": "ACF-02", "order": 2, "why": "Calibrate criticality with the matrix." },
{ "code": "ACF-09", "order": 3, "why": "Define escalation thresholds." },
{ "code": "ACF-11", "order": 4, "why": "Run a formal risk assessment." }
],
"applicable_articles": [
{ "regulation": "ai-act", "article": "Art. 9", "why": "Mobilised by category Annex III §4 — employment." },
{ "regulation": "ai-act", "article": "Art. 14", "why": "Mobilised by category Annex III §4 — employment." },
{ "regulation": "ai-act", "article": "Art. 26", "why": "Mobilised by category Annex III §4 — employment." },
{ "regulation": "ai-act", "article": "Art. 51", "why": "GPAI obligation." }
],
"first_actions": [
"Identify or appoint the DDAO accountable for the agent (cf. ACF-12).",
"Document the agent constitution (decision perimeter, allowed/forbidden actions — cf. ACF-03).",
"Open the decision register and define the retention policy (cf. ACF-05 + ACF-13).",
"Run the formal risk assessment + DPIA if PII (cf. ACF-11)."
],
"operational_risks": [
"Audit chain breaks if the decision register is not immutable.",
"Sign-off bottleneck if DDAO is not available in escalation SLA.",
"Regulatory exposure escalates if Article 49 register is not maintained in sync with internal register."
],
"confidence": "medium",
"assumptions": [
"human_approval, external_actions and personal_data_level were inferred from the free-text description; pass classify-agent with qualified enums for higher confidence."
],
"gaps_to_validate": [
"re-run via acf.classify-agent with structured enums for a more defensible qualification."
],
"requires_human_review": true,
"rationale_per_rule": [
{
"rule_id": "ai-act.annex-iii.employment",
"rule_version": "2026-06",
"fired": true,
"evidence": "description mentions shortlist + ATS write in an HR context"
},
{
"rule_id": "autonomy.inference.supervised",
"rule_version": "2026-06",
"fired": true,
"evidence": "inferred human_approval=always, external=limited_write from description"
}
],
"doctrine_version": "ACF framework v1.0 / rules 2026-06",
"doctrine_hash": "sha256:bf0b6d8e4731ebdc58f6d6338702c5b74af47874cf0ad3dc958cde5c5b30b9dc",
"doctrine_signature": "ed25519:…",
"doctrine_archive_url": "https://acfstandard.io/doctrine/v1.0/archive.json",
"regulatory_snapshot": "EU AI Act (Reg. 2024/1689, incl. Digital Omnibus deferral) + GDPR (Reg. 2016/679) + DORA (Reg. 2022/2554) + NIS2 (Dir. 2022/2555) + ISO 42001:2023 — as of 2026-06-07",
"generated_at": "2026-06-14T11:47:22.318Z",
"conversion_cta": "Continue this assessment with the auditable ACF® Compliance workspace at https://acfstandard.com/compliance?ref=mcp",
"disclaimer": "Preliminary qualification only — not legal advice. Human review required."
}Erreurs courantes
InputTooShort— case_description < 20 caractères. L’outil ne devine pas — fournissez une description prose minimale.InvalidEnumValue— jurisdiction ou deployment_scale reçoit une valeur hors liste. Corrigez vers une valeur canonique ou omettez le champ.DoctrineSnapshotMismatch— le doctrine_hash demandé n’est pas chargé. Mettez acf-mcp à jour ou pointez vers la version archivée.
Outils liés
acf.classify-agent— repasser avec des énumérations qualifiées dès que le cas est précisé, pour gagner en confiance.acf.assess-autonomy— creuser la seule décision de niveau d’autonomie quand c’est le point bloquant.acf.identify-governance-gaps— élargir du cas isolé à l’audit de gouvernance globale de l’organisation.