# AI GOVERNANCE ASSESSMENT — HIREBOT AI Running AGAF-UFO-001 Lite v1.0 across three phases. --- ## PHASE 1 — SYSTEM PROFILE (UFO Object Typing) ``` SYSTEM PROFILE (UFO Object) ────────────────────────────────────────────────────────────────────── Field | Type | Value | Tag ────────────────────────────────────────────────────────────────────── System name | Quality | HireBot AI | {CONF} Purpose | Mode | Screen applications, rank | {CONF} | | candidates, recommend top | | | applicants for review | Sensitive data | Quality | true | {CONF} Data types | Mode | CVs, LinkedIn profiles, | {CONF} | | cover letters, interview | | | notes | Autonomous? | Quality | partial | {CONF} Human oversight | Mode | HR managers review top 20 | {CONF} | | ranked candidates before | | | interviews are scheduled | Transparency method | Mode | {not stated} | {GAP} Developer | Mode | {not stated} | {GAP} ────────────────────────────────────────────────────────────────────── ``` **Ontological grounding:** HireBot AI is typed as an **Object** — not a process — because it is a persistent, bounded entity with stable properties (name, purpose, data inputs, autonomy level) that endures across individual screening events and can be evaluated, audited, and held accountable independently of any single decision it produces. --- ## PHASE 2 — RISK CLASSIFICATION + DOCUMENT CLASSIFICATION ### 2A. Risk Classification ``` RISK CLASSIFICATION ────────────────────────────────────────────────────────────── Transition: UNCLASSIFIED → HIGH-RISK Reason: The system ranks and recommends candidates for employment — a decision domain explicitly named as high-risk under EU AI Act Annex III and equivalent frameworks; its outputs directly gate which humans proceed to interview, making it a consequential filter on employment access. Confidence: {CONF} Review note: This classification is context-dependent and may change if the regulatory regime or deployment context changes. Jurisdictions without AI-specific employment regulation may apply different thresholds, but the functional risk profile does not change with jurisdiction. ────────────────────────────────────────────────────────────── ``` ### 2B. Document Classification ``` ────────────────────────────────────────────────────────────────────── Document | Class | Type | Weight ────────────────────────────────────────────────────────────────────── Privacy Policy | 2 | Policy / Framework | Medium | | Defines data handling | | obligations. Normative | | force, but not system- | | level enforcement. ────────────────────────────────────────────────────────────────────── AI Recruitment Policy | 2 | Policy / Framework | Medium | | Defines what must happen | | in AI-assisted hiring. | | Normative, not technical. ────────────────────────────────────────────────────────────────────── Candidate Screening | 1 | Awareness Guide | None Guidelines | | Informational. Describes | | how screening works. | | No binding commitments | | confirmed. ────────────────────────────────────────────────────────────────────── Technical Design | 3 | Technical Spec | High Specification | | Specifies how controls | | are implemented in the | | system. Enforcement basis | | exists — content | | unverified. ────────────────────────────────────────────────────────────────────── ``` **Class coverage:** - Class 1 — Present ✓ - Class 2 — Present ✓ - Class 3 — Present ✓ - Class 4 — **ABSENT** → No conformance claim exists. No auditable assertion that the system meets any named standard. `{GAP — MEDIUM priority}` **Notable flag:** The Technical Design Specification (Class 3) is named but its content has not been provided. Its classification as Class 3 is based on the name alone — the enforcement basis it provides is `{UNCONF}` until content is reviewed. A document named "Technical Design Specification" that does not include testable control specifications defaults functionally to Class 1. --- ## PHASE 3 — THEATER DETECTION + REPORT ### 3A. Theater Detection --- **GTD-1: Monitoring as Governance** ``` Status: PARTIAL Evidence: HR managers review the top 20 ranked candidates before interviews are scheduled. This is a review gate, not a monitoring mechanism per se — however, the scope of the gate is the concern: only top-20 candidates are reviewed, meaning the system's decision to exclude candidates below rank 20 is never seen by a human reviewer. The oversight covers the output surface, not the exclusion decisions where harm is most likely to occur. No monitoring of model drift, demographic disparity, or ranking consistency is mentioned. Severity: HIGH ``` --- **GTD-2: Policy Theater** ``` Status: PARTIAL Evidence: Two Class 2 documents exist (Privacy Policy, AI Recruitment Policy). Neither has been provided for content review — enforcement traceability is {UNCONF}. The pattern is flagged as partial because named policies exist (not absent), but without content, assessment cannot confirm that "shall" statements are tied to named owners, escalation paths, or consequences for non-compliance. If the Technical Design Specification does not trace back to specific policy clauses, GTD-2 upgrades to FULL. Severity: MEDIUM ``` --- **GTD-3: Explainability as Story** ``` Status: NOT DETECTED Evidence: No transparency or explanation mechanism has been described — the field is {GAP}. Assessment cannot detect GTD-3 where no explainability claim has been made. This is not a positive finding; it means the system has made no explainability claim to evaluate. The absence of any explanation mechanism is itself a HIGH-priority gap for a HIGH-RISK system. Severity: N/A ``` --- **GTD-4: Probabilistic Shrugging** ``` Status: NOT DETECTED Evidence: No uncertainty or accuracy claims have been made in the inputs provided. No statements of the form "the model sometimes gets it wrong" appear. {GAP} — insufficient input to detect. Severity: N/A ``` --- **GTD-5: Synthetic Authority** ``` Status: NOT DETECTED Evidence: No vendor certifications, third-party endorsements, or credential claims appear in the provided inputs. {GAP} — insufficient input to detect. Severity: N/A ``` --- **Theater Score: 2 / 5** *GTD-1 (partial) + GTD-2 (partial) = 2 detected patterns.* > **Verdict: Governance theater present — address before deployment or audit sign-off.** The two detected patterns are not incidental. For a HIGH-RISK employment system, a review gate that never sees exclusion decisions (GTD-1) and policies whose enforcement traceability is unverified (GTD-2) represent the exact failure modes most likely to cause harm and least likely to be caught by routine operation. --- ### 3B. Final Report ``` AI GOVERNANCE ASSESSMENT — HIREBOT AI Date: 27 June 2026 Framework: AGAF-UFO-001 Lite v1.0 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ SUMMARY ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Risk tier: HIGH-RISK Theater score: 2/5 Document classes: Class 1 ✓ | Class 2 ✓ | Class 3 ✓ (unconf) Class 4 ABSENT {GAP — MEDIUM priority} Overall verdict: CONDITIONAL ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ TOP FINDINGS ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1. OVERSIGHT BLIND SPOT (GTD-1, HIGH): Human review is scoped only to the top 20 ranked candidates. The system's exclusion decisions — which determine who never enters the pipeline — receive no human review. For a HIGH-RISK employment system, exclusion decisions carry equal or greater legal exposure than inclusion decisions. 2. POLICY ENFORCEABILITY UNVERIFIED (GTD-2, MEDIUM): Two named policy documents exist but their content has not been assessed. Enforcement traceability — named owners, escalation paths, consequences for non-compliance — is {UNCONF}. The Class 3 document exists by name only; its status as a genuine technical specification is unverified. 3. NO EXPLAINABILITY MECHANISM IDENTIFIED ({GAP, HIGH}): No transparency method for individual ranking decisions has been described. For candidates who are excluded below rank 20, there is no stated mechanism to explain why. This is a legal exposure under GDPR Article 22 (automated decision-making) and equivalent frameworks, and a functional requirement for any HIGH-RISK employment AI. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ IMMEDIATE ACTIONS ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1. [GTD-1] Extend human oversight to exclusion decisions: implement a sampling protocol where a defined percentage of sub-rank-20 exclusions are reviewed each cycle by a named HR reviewer. Log reviewer identity, date, and finding. Do not treat "top 20 review" as complete oversight. 2. [GTD-2] Audit the AI Recruitment Policy for enforcement completeness: every "shall" statement must name an owner role, a verification mechanism, and a consequence for non-compliance. Map each policy clause to a corresponding control in the Technical Design Specification. Where no control exists, the clause is unenforceable — name it and remediate or remove it. 3. [GAP — Explainability] Define and implement a candidate-facing explanation mechanism for exclusion decisions: at minimum, the system must be able to produce a ranked list of features that influenced a given candidate's score. Log this per decision. "Available on request" is not sufficient — the mechanism must be automatic and auditable. 4. [GAP — Class 4] Commission or produce a conformance claim document: identify the specific standard(s) the system is intended to meet (EU AI Act, ISO/IEC 42001, internal baseline), and produce an auditable assertion of conformance with evidence references. Without this, there is no verifiable governance baseline. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ WHAT THIS ASSESSMENT DOES NOT COVER ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ This assessment covers system profiling, risk tier classification, document classification, and governance theater detection. It does not cover: specification grading of individual controls, accountability node mapping, recursive loop review, human-in-the-loop trigger completeness, or conformance verification against specific regulations. A full governance assessment covers all of these. Specific unassessed exposures relevant to HireBot AI include: bias and fairness testing methodology, demographic disparity monitoring, data retention and deletion obligations under GDPR, model drift detection cadence, and supplier/vendor accountability if HireBot AI is a third-party system. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ``` --- **Assessor note:** The document inventory for HireBot AI is stronger than most systems at this tier — four named documents including a Class 3 specification is a reasonable starting position. The critical vulnerabilities are not in what documents exist but in what the oversight mechanism is actually designed to see. A system that screens hundreds of candidates but only shows a human reviewer the top 20 has built governance around its successes, not its risks. That inversion is the core finding.