1. Executive Summary & Architectural Baseline
Carbonatoz operates under a strict, formally governed "Deterministic Engine Core + Governed AI Advisory Projection" architecture (DEC-065 & DEC-068):
Core Rules Governing AI in Carbonatoz
INV-10)**: All AI recommendations and draft mappings are strictly advisory. The AI is architecturally prohibited from signing, submitting, or sealing statutory compliance filings.INV-AI-01)**: The AI Copilot cannot execute any capability that your authenticated user role cannot execute directly. Every tool invocation runs under your active session JWT (auth.uid()).INV-AI-05)**: All foundation model integrations (Anthropic, OpenAI) enforce contractual Zero Data Retention and strictly prohibit training on enterprise data.2. Dual-Agent Operational Modes
The platform provides two distinct AI assistance environments:
| Mode | Environment | Access & Data Boundary | Primary Use Cases |
|---|---|---|---|
| Public Landing Page Guide | Public marketing & knowledge pages (/help, /calculator, /) | Zero tenant data access. Only reads public statutory reference documents and public knowledge articles. | Explaining regulatory requirements, calculating preliminary public CBAM estimates, navigating documentation. |
| Internal Platform Copilot | Authenticated application dashboard (/dashboard, /sustainability, /icm) | Server-derived tenant context. Reads only the active organization's permitted data via PostgreSQL Row-Level Security (RLS). | Navigating complex workflows, explaining data anomalies, proposing factor mappings, drafting form fields. |
3. The Action Proposal Workflow
When the AI Copilot assists with data modifications (such as suggesting an emission factor or flagging anomalous meter data), it operates through the Action Proposal Pattern (Autonomy Tiers A2/A3):
┌─────────────────────────┐ ┌─────────────────────────┐ ┌─────────────────────────┐
│ 1. AI Copilot Intent │ ──> │ 2. Visual Diff Review │ ──> │ 3. Human Execution │
│ Translates request and │ │ Renders Before vs After│ │ User approves or edits │
│ generates typed payload│ │ with evidence citations│ │ before database write │
└─────────────────────────┘ └─────────────────────────┘ └─────────────────────────┘Operational Steps for Reviewing Action Proposals
approve_activity_data).Current State versus the Proposed State.4. Understanding Decision Provenance & Audit Trails
To satisfy ISO/IEC 42001:2023 and EU AI Act Article 14/50 requirements, Carbonatoz maintains an unbroken, tamper-evident audit trail for every AI interaction.
The Provenance Hash Chain (INV-AI-04)
Every AI-assisted recommendation writes an atomic entry into public.ai_decision_support_logs and public.ai_grounding_sources:
Where to Inspect AI Logs
- Compliance Officers & Lead Auditors: Navigate to Governance \to AI Management (ISO 42001) \to Decision Support Logs.
- Audit Details Displayed:
- Exact model code and version used (e.g.
ANOMALY_DETECTOR_V1). - Source data snapshot and input hash.
- Human review outcome (
accepted,modified, orrejected). - Timestamps, user IDs, and regulatory framework citations.
5. Evidence Strength vs. Elimination of "Fake AI Confidence"
In accordance with DEC-065 §2.4 and DEC-068 §1, Carbonatoz completely eliminates arbitrary single-percentage confidence scores (e.g. "87% confidence").
Instead, the platform evaluates and discloses multidimensional Assurance Metrics:
| Metric | Meaning & Calculation Basis | How to Interpret |
|---|---|---|
| Evidence Completeness | Ratio of primary verified evidence records (invoices, meter logs) to secondary defaults. | Higher is better. 100\% indicates fully primary verified operational telemetry. |
| Data Quality Index | Sensor calibration freshness, variance stability, and peer baseline alignment. | Rated from High (verified direct meter) to Low (uncalibrated estimate). |
| Statutory Authority | Hierarchy of regulatory citation (e.g. Official Journal Implementing Act vs. Industry Benchmark). | Indicates whether the factor is a legally binding default or reference guideline. |
| Human Review State | Status of human-in-the-loop validation (Pending Review, Verified, Overridden). | Mandatory check before including data in statutory quarterly filings. |
6. Frequently Asked Questions (FAQ)
Q1: Can the AI Copilot submit our EU CBAM declaration automatically?
No. Under platform invariant INV-10 and EU AI Act Article 14, automated pipelines and AI systems are strictly forbidden from submitting statutory declarations. The Copilot can assist in checking missing fields and validating data completeness, but the final XML generation and signature requires multi-factor authenticated human execution.
Q2: Is our confidential facility data sent to public AI models?
No. All external AI gateway requests enforce contractual Zero Data Retention (ZDR). Customer telemetry is never used for foundation model training. In addition, sensitive compliance data is strictly routed through approved provider tiers meeting enterprise data protection agreements.
Q3: What happens if I disagree with an AI emission factor suggestion?
You can override or reject any suggestion with a single click. When you select a custom factor, the platform logs your override in public.ai_human_override_records, creating an auditable record of your expert professional judgment for third-party verifiers.