Regulatory Compliance
Built for auditors, not against them. Every architectural decision was made with regulatory compliance in mind — not retrofitted.
Privacy by Design
PII Vault · Audit Trail · Execution Rings · Token Tracking
Data Sovereignty
All data stays on your SovereignNode. No byte leaves your network — unless you explicitly allow it.
Privacy Engine
Automatic anonymization of personal data before every external API call.
Mappings are session-bound and not persistently stored
Detection via regex patterns and named entity recognition
Every anonymization is documented in the audit log
Logging
Every agent action is comprehensively logged. The audit trail captures six event types:
{
"id": 42847,
"timestamp": "2026-03-22T14:32:17.445Z",
"session_id": "ses_7f3a2b91",
"agent": "logistics",
"event_type": "TOOL_START",
"tool": "sql_query",
"ring": 1,
"params": {
"connector": "uyumsoft_db",
"query": "SELECT item, qty FROM stock WHERE qty < 10"
},
"user_id": "tg_12345",
"pii_anonymized": false
}{
"event_type": "LLM_USAGE",
"session_id": "ses_7f3a2b91",
"model": "qwen3.5:27b",
"prompt_tokens": 2847,
"completion_tokens": 312,
"total_tokens": 3159,
"latency_ms": 4280,
"cognitive_balance": 3
}Access Control
How execution rings prevent unauthorized actions.
Resource Monitoring
Every LLM call is documented with exact token usage, model name, and latency.
Exact breakdown into prompt_tokens and completion_tokens per request.
Every LLM request is recorded with latency_ms, including model name and session context.
Remaining LLM calls per session are tracked as cognitive_balance.
Regulatory Compliance
GDPR · ISO 9001 & 27001 · GoBD
EU Data Protection
AIMOS meets GDPR not through retrofitting but through architectural decisions — Privacy by Design.
Standards Compliance
AIMOS supports the requirements of both standards through built-in mechanisms.
Bookkeeping
The GoBD (Principles for the proper management and storage of books, records, and documents in electronic form) imposes special requirements on AI-assisted systems.
AIMOS solves this through strict architectural separation:
Agents prepare booking entries (draft). Finalization is performed exclusively by authorized clerks in the target system (ETA V8, DATEV, etc.). No LLM has transaction rights.
Ring 2 actions (write access) require explicit authorization. Booking-relevant operations always require human approval — architecturally enforced, not by policy.
EU AI Act
Risk Classification · Transparency · Human Oversight · Documentation
EU Regulation
The EU AI Act (EU) 2024/1689 takes effect in phases and affects every company that deploys AI systems. AIMOS is architecturally designed for compliance — not retrofitted.
AIMOS agents are assistance systems with human oversight (human-in-the-loop). No autonomous decisions about individuals, no biometric surveillance, no social scoring functions. This places them in the "limited risk" category — not "high-risk".
Art. 50: Users must know they are interacting with AI. Every AIMOS agent identifies itself in its system prompt as an AI assistant. All actions are traceable in the audit trail.
Art. 14: Execution Rings ensure that no agent independently performs critical actions. Bookings, emails, and orders are presented as drafts — the human decides.
Art. 11: System architecture, data flows, model description, performance metrics, and risk assessment are documented and accessible. MIT License: full source code access.
Art. 12: Every agent action is logged — tool calls, LLM requests, token usage, decision chains. Audit-proof trails with session IDs and timestamps.
Art. 10: Training data and processing data do not leave the SovereignNode. PII Vault automatically anonymizes external API calls. No cloud dependency.
Cloud AI services must comply with stricter "General Purpose AI" (GPAI) requirements. AIMOS uses a locally operated open-source model — GPAI obligations fall on the model provider (Alibaba/Qwen), not the operator.
Provability
Comprehensive traceability of every AI-assisted decision.
Every LLM call is documented with exact token usage (prompt + completion), model name, and latency.
From user request through tool calls to response — every step is traceable with timestamp and session ID.
Session-bound anonymization mappings. Traceable which data was anonymized and de-anonymized when.
MIT License: Auditors can review every line of code. No black box, no proprietary logic.