AI System Technical Overview

Overview

This article provides technical transparency into how ISMS Copilot's AI systems are built, tested, and operated. These details demonstrate our commitment to responsible AI development through verifiable implementation practices.

Who This Is For

This article is for:

• Security and compliance teams evaluating AI governance controls
• Auditors assessing AI system implementation against policies
• Risk managers requiring technical transparency for AI systems
• Technical users wanting to understand the AI architecture

ISO 42001 Implementation

ISMS Copilot's AI system is designed and operated in accordance with ISO 42001:2023 (Artificial Intelligence Management System) requirements. Our technical implementation maps to specific ISO 42001 controls:

Architecture Alignment:

• A.4.2 (Context Establishment): Dynamic framework knowledge injection system documented in AI System Design Document (AI-SDD-001)
• A.5 (Impact Assessment): Comprehensive AI Impact Assessment (1.9 Low Risk, EU AI Act Limited Risk classification)
• A.6 (Responsible Development): Secure development lifecycle with regression testing, SAST/DAST scanning, prompt injection testing
• A.7 (Data Management): Zero Data Retention agreements, workspace isolation, user-controlled retention periods
• A.8 (User Interaction): Transparency notices, human-in-the-loop design, verification disclaimers throughout platform
• A.9 (Responsible Use): Purpose limitation, jailbreak prevention, content scope guardrails

Dynamic Framework Knowledge Architecture

ISMS Copilot uses dynamic framework knowledge injection to ground AI responses in verified compliance knowledge. As of version 2.5 (February 2025), this replaces the previous RAG (Retrieval-Augmented Generation) architecture with a more reliable, token-efficient approach.

How Framework Knowledge Injection Works

Architecture Components:

• Framework Detection Layer: Regex-based pattern matching detects framework mentions in user queries (ISO 27001, SOC 2, GDPR, HIPAA, CCPA, NIS 2, DORA, ISO 42001, ISO 27701, EU AI Act)
• Knowledge Injection Layer: Dynamically loads only relevant framework knowledge into AI context based on detected frameworks
• Generation Layer: Large language models (LLMs) from enterprise AI providers receive framework knowledge before generating responses
• Validation Mechanism: Framework knowledge provided to AI ensures responses are grounded in actual compliance requirements, not probabilistic guessing

Why Dynamic Injection Matters for Compliance:

• Eliminates hallucination: AI receives verified framework knowledge before answering, preventing fabricated control numbers and requirements
• Token efficiency: Only relevant frameworks loaded (~1-2K tokens) vs. sending all knowledge (~10K tokens) on every request
• Reliable detection: Regex pattern matching (not AI-based) ensures framework mentions are never missed
• Extensible architecture: New frameworks added with single object definition, no model retraining required
• Multi-framework support: Handles queries mentioning multiple frameworks simultaneously (e.g., \"Map ISO 27001 to SOC 2\")

Technical Implementation

Detection Process:

1. User submits query (e.g., \"What is ISO 27001 Annex A.5.9?\")
2. Framework detection scans query for pattern matches (ISO 27001, GDPR, SOC 2, etc.)
3. Matched frameworks trigger knowledge injection
4. Relevant framework knowledge added to AI system prompt before generation

Supported Frameworks (v2.5):

• ISO 27001:2022 — Information Security Management System
• ISO 42001:2023 — Artificial Intelligence Management System
• ISO 27701:2025 — Privacy Information Management System
• SOC 2 — Service Organization Control (Trust Services Criteria)
• HIPAA — Health Insurance Portability and Accountability Act
• GDPR — General Data Protection Regulation
• CCPA — California Consumer Privacy Act
• NIS 2 — Network and Information Systems Directive
• DORA — Digital Operational Resilience Act
• EU AI Act — European Union Artificial Intelligence Act

Evolution from RAG to Dynamic Injection

Previous Approach (Pre-v2.5): RAG Architecture

• Semantic search retrieved relevant documentation snippets
• Retrieval quality varied based on query phrasing
• All ~10K tokens of knowledge sent on many requests
• Focused primarily on ISO 27001

Current Approach (v2.5+): Dynamic Framework Injection

• Regex-based detection ensures reliable framework identification
• Only relevant frameworks loaded (token-efficient)
• Supports 10 frameworks simultaneously
• Extensible design for rapid framework additions

AI Providers & Data Protection

We use enterprise-grade AI providers with strict data protection agreements.

Current Providers

Backend AI Models:

• xAI Grok — Current default path for many paid Fast / Think / Beyond sessions when Advanced Data Protection is off (zero-retention style routing; see Trust Center for the live table)
• Anthropic Claude — Backup / alternate path depending on plan and availability
• OpenAI / other OpenRouter-class models — Economy routing for Free/Essential and some specialized tasks
• Mistral AI — EU path when Advanced Data Protection is on

Routing is plan- and setting-dependent, not a single fixed brand for every user. Automatic failover can move traffic when a provider is unhealthy. When Advanced Data Protection is enabled, AI processing follows the EU Mistral topology described in the ADP docs. All routes use the same specialized compliance knowledge injection. For diligence, always prefer the Trust Center over this narrative summary if they ever differ.

Zero Data Retention Agreements

All AI providers operate under Zero Data Retention (ZDR) agreements:

ZDR Agreement Terms:

• No user data retention beyond request processing
• No model training on customer content
• GDPR-compliant data transfers with Standard Contractual Clauses (SCCs)
• Enterprise security standards enforced

For detailed processor information and data flows, see our Register of Processing Activities.

Development Requirements

Every AI system component is developed against documented requirements defining expected behavior, safety constraints, and performance thresholds.

Functional Requirements

Scope Definition:

• AI provides compliance assistance, not legal advice
• Task boundaries: policy generation, gap analysis, audit preparation, document review
• Constraint enforcement: no internet access, no code execution, no personal data processing beyond platform usage

Performance Requirements

Quality Targets:

• Response accuracy grounded in retrieved sources with citation
• Context window sufficient for multi-document compliance analysis
• Response time optimized for interactive use (target: under 10 seconds)
• Rate limits defined per user tier to ensure system stability

Safety Requirements

Hallucination Mitigation:

• Source grounding: responses must reference retrieved documentation
• Retrieval validation: responses checked against source content
• Confidence scoring: uncertainty acknowledged when sources are ambiguous
• User verification disclaimers: all outputs require human review

Content Filtering:

• Inappropriate content detection and blocking
• Scope boundaries: AI refuses out-of-scope requests (e.g., unrelated topics, medical/legal advice)
• Jailbreak and prompt injection protection

See AI Safety & Responsible Use Overview for detailed safety guardrails.

Data Handling Requirements

Privacy by Design:

• No user data for model training (ZDR agreements enforced)
• Data minimization: only necessary data processed for retrieval and generation
• Temporary processing: no long-term storage of prompts/responses beyond user session logs
• Retention controls: user-configurable data retention periods (1 day to 7 years, or keep forever)
• Transfer controls: GDPR-compliant data transfers with SCCs

For comprehensive data handling practices, see our Privacy Policy.

Verification & Validation Testing

AI systems undergo rigorous testing before deployment. No system goes live without passing requirements-based validation.

Regression Testing

Automated tests run on every code change to ensure existing functionality remains intact.

Test Coverage:

• Retrieval accuracy: Precision and recall against ground truth datasets
• Response grounding: Verification that outputs cite retrieved sources
• Hallucination detection: Comparison against known incorrect responses
• Performance benchmarks: Response time and context handling validation

Security Testing

AI systems undergo the same security validation as all platform components.

Testing Pipeline:

• SAST (Static Application Security Testing): Code-level vulnerability scanning with Semgrep integration
• DAST (Dynamic Application Security Testing): Runtime security validation
• Penetration Testing: Annual third-party security assessments
• Prompt Injection Testing: Validation against adversarial inputs attempting to bypass safety constraints

User Acceptance Testing

Real-world scenario validation with compliance professionals ensures:

• Outputs meet professional quality standards
• Responses are appropriate for compliance use cases
• Limitations are clearly communicated
• Feedback mechanisms are accessible and effective

Deployment Validation Checklist

AI systems are deployed only after meeting documented requirements:

Deployments that fail validation are rolled back until requirements are satisfied.

Monitoring & Continuous Improvement

Post-deployment, we monitor AI system behavior to detect degradation, emerging issues, or misuse.

Monitoring Metrics

What We Track:

• Hallucination rate: Tracked through user reports and automated detection
• Response accuracy: Sampled validation against ground truth compliance standards
• Usage patterns: Detection of out-of-scope or inappropriate use
• Performance metrics: Response time, retrieval precision, error rates
• User feedback: Adverse impact reports, support tickets, feature requests

Continuous Improvement Cycle

Monitoring data informs iterative improvements:

Feedback Loops:

• User feedback and adverse impact reports → model updates and retrieval tuning
• Security testing results → safety enhancements and control updates
• Regulatory changes and best practices → documentation and framework updates
• Performance monitoring → accuracy improvements and response optimization

Incident Response

We notify users of AI-related incidents to maintain transparency and trust.

Notification Channels:

• Email alerts for critical incidents affecting AI functionality
• Slack notifications for subscribed teams
• Status page updates with incident timelines and resolutions
• NIS2-compliant early warning notifications (24-hour reporting for significant cybersecurity incidents)

Known Limitations

AI systems have inherent limitations that users must understand to use them responsibly.

Technical Limitations

Current Constraints:

• Probabilistic nature: AI generates responses based on statistical patterns, not deterministic logic
• No internet access: AI cannot retrieve real-time information or access external websites
• No code execution: AI cannot run calculations, execute scripts, or validate technical implementations
• Knowledge cutoff: AI model knowledge is limited to training data cutoff dates (varies by provider)
• Context limits: Maximum context window constrains the amount of information processed in a single request
• Domain boundaries: AI is trained for compliance/security; performance in other domains is not guaranteed

For detailed limitations and workarounds, see our Known Issues page.

User Verification Responsibility

ISMS Copilot is designed to assist, not replace, professional judgment:

• Cross-reference AI suggestions with official standards
• Validate critical information before submission to auditors
• Use the AI as a consultant's assistant, not a replacement for expertise
• Exercise professional judgment in applying AI recommendations

See How to Use ISMS Copilot Responsibly for verification best practices.

Reporting & Feedback

User feedback is critical for AI system improvement. We provide multiple mechanisms for reporting issues, inaccuracies, or unexpected behavior.

How to Report Issues

Adverse Impacts or Hallucinations:

1. Navigate to user menu (top right) > Help Center > Contact Support
2. Include prompt, response, and screenshots in your report
3. Expect response within 48 hours

In-Platform Reporting:

• Use "Report Issue" button available throughout the platform to flag specific AI responses

What Happens After You Report

1. Immediate review (within 48 hours): Support team assesses severity and impact
2. Investigation: Technical team analyzes the issue, reproduces the problem, and identifies root cause
3. Response: You receive an update on findings and planned actions
4. Remediation: Issues addressed through model updates, retrieval tuning, code fixes, or documentation improvements
5. Continuous improvement: Lessons learned integrated into testing and monitoring processes

See AI Safety & Responsible Use Overview for detailed reporting instructions.

Documentation Updates

Technical specifications are updated when:

• AI providers change (new models, deprecated APIs)
• Architecture evolves (new components, validation methods)
• Requirements are revised (new safety constraints, performance targets)
• Testing practices expand (new validation techniques, security tools)

Updates are communicated through release notes and this documentation page. Subscribe to our status page for change notifications.

What's Next

• Review how ISMS Copilot implements ISO 42001
• Learn about AI safety guardrails and responsible use practices
• Understand AI hallucinations and how to prevent them
• Follow best practices for using ISMS Copilot responsibly
• Review our comprehensive Security Policies

Getting Help

For technical questions about AI system specifications or to request additional documentation:

• Contact support through the Help Center menu
• Report safety concerns immediately for investigation
• Review the Trust Center for detailed AI governance information
• Check the Status Page for known issues