The M3 Framework: Mount, Monitor, Manage — A Practical Guide to AI Compliance

AI Compliance Is a Lifecycle, Not a Checkbox

Most organisations approach AI compliance the same way they approached IT compliance a decade ago: as a one-time audit exercise. Define controls, pass the assessment, file the certificate, and move on until the next renewal cycle. This worked — barely — when systems were static. But AI agents are not static. They learn, drift, ingest new data, and interact with thousands of users in unpredictable ways every day.

The regulatory landscape reflects this reality. The EU AI Act mandates ongoing risk management and post-market monitoring. ISO 42001 requires continuous improvement cycles. GDPR demands that data processing remains lawful not just at deployment but throughout the system's lifetime. A point-in-time audit cannot satisfy these requirements. What organisations need is a governance framework that treats compliance as a continuous operational discipline — not a project with a finish line.

That is exactly what the M3 Framework® was designed to provide.

What Is the M3 Framework?

The M3 Framework is an open compliance standard created by Julius Gromyko — a PECB-certified ISO 42001 Implementer, ISO 27001 Foundation practitioner, ISO 31000 Risk Manager, and GDPR Data Protection Officer. It is published at m3framework.org and is free for internal use by any organisation.

M3 stands for its three operational phases: Mount, Monitor, Manage. Rather than mapping compliance to a single regulation, the framework provides a unified control structure that spans multiple regulatory regimes simultaneously — GDPR, EU AI Act, ISO 27001, ISO 42001, and NIST AI RMF. This means organisations maintain one operational workflow instead of five separate compliance checklists.

M3 Framework is not a replacement for ISO 42001 or GDPR. It is the operational layer that makes compliance with all of them sustainable, auditable, and continuous.

The framework was born from practical necessity. When building Sinaptic® DROID+ — a platform that deploys AI agents across retail, healthcare, hospitality, beauty, dental, and travel — Gromyko needed a single governance structure that could satisfy European regulators, enterprise procurement teams, and healthcare compliance officers simultaneously. No existing framework provided that unified operational view. M3 was built to fill that gap.

Phase 1: Mount

The Mount phase covers everything that must happen before an AI system goes live. This is where governance begins — not as documentation theatre, but as genuine architectural decision-making that shapes how the system will behave in production.

Define scope and boundaries. Document exactly what the AI agent is intended to do, what it must never do, which user populations it serves, and which data sources it accesses. Ambiguity in scope is the number one cause of compliance failures.
Identify and classify risks. Conduct a structured AI risk assessment aligned with ISO 42001 Annex B categories and the EU AI Act's risk classification tiers. This includes risks of bias, hallucination, data leakage, adversarial manipulation, and unintended autonomous decisions.
Map to regulatory requirements. For each identified risk and control objective, M3 provides a cross-reference matrix that maps to the corresponding requirements in GDPR (Articles 5, 13, 22, 25, 35), EU AI Act (Articles 9, 10, 13, 14, 15), ISO 27001 (Annex A controls), ISO 42001 (Clauses 6–10 and Annex B), and NIST AI RMF (Govern, Map, Measure, Manage functions).
Establish data governance. Document data provenance, processing purposes, retention policies, and access controls. For AI agents using retrieval-augmented generation (RAG), this includes governing which knowledge sources are trusted, how ingestion is validated, and what DLP controls are applied.
Configure human oversight. Define HITL thresholds — which decisions require human review, which can proceed autonomously, and what escalation paths exist. This directly satisfies EU AI Act Article 14 and GDPR Article 22 requirements.
Document stakeholder mapping. Identify data subjects, data controllers, processors, AI system operators, and affected persons. Every integration point is catalogued with its compliance obligations.

The output of the Mount phase is a deployment compliance package — a structured set of documentation that serves as the system's governance baseline. This package is auditable from day one and becomes the reference point for all subsequent monitoring and management activities.

Phase 2: Monitor

The Monitor phase addresses the fundamental gap in traditional compliance: what happens after deployment. AI systems do not stay compliant by default. Models drift. User behaviour changes. Knowledge bases update. New vulnerabilities emerge. Regulations evolve. Without continuous monitoring, the compliance baseline established in the Mount phase degrades within weeks.

Behavioural observation. Continuous tracking of agent responses against defined scope boundaries. Are responses staying within the intended domain? Are there patterns of hallucination or factual drift? Is the agent handling edge cases as designed?
Drift detection. Automated alerting when key performance metrics — accuracy, relevance, safety scores, response latency — deviate beyond acceptable thresholds. Drift can be gradual (knowledge base staleness) or sudden (adversarial prompt patterns), and both require different response protocols.
Security event logging. Every interaction that triggers a security control — Sinaptic Intent Firewall blocks, DLP interventions, prompt injection attempts, data exfiltration detections — is logged with full context for forensic review and regulatory reporting.
Compliance metric dashboards. Real-time visibility into HITL escalation rates, data subject access request (DSAR) processing times, consent status, and transparency disclosure compliance. These metrics feed directly into management review cycles required by ISO 42001.
Alerting and notification. Configurable thresholds that trigger alerts to compliance officers, security teams, or operations staff when monitoring detects anomalies. The framework defines three alert tiers: informational (log and review), warning (investigate within SLA), and critical (immediate human intervention required).

The Monitor phase transforms compliance from a periodic review into an always-on operational function. It is what makes the difference between an organisation that claims compliance and one that can demonstrate it at any point in time.

Phase 3: Manage

The Manage phase closes the governance loop. Monitoring generates data; management turns that data into decisions, actions, and improvements. This is where compliance becomes a competitive advantage rather than a cost centre.

Remediate. When monitoring identifies drift, violations, or incidents, the Manage phase provides structured remediation workflows. This includes root cause analysis, corrective action plans, and verification testing before returning the system to production.
Retrain and update. AI agents require periodic knowledge base updates, scenario adjustments, and — in some cases — model retraining. The Manage phase governs how these changes are validated, tested, and deployed without introducing new compliance risks. Every update follows the same Mount-phase governance requirements as the initial deployment.
Report. Generate compliance reports for regulators, auditors, management boards, and enterprise clients. M3 provides report templates that map evidence directly to regulatory requirements — eliminating the manual effort of translating operational data into compliance language.
Review and improve. Conduct periodic management reviews as required by ISO 42001 Clause 9.3 and ISO 27001 Clause 9.3. These reviews evaluate the effectiveness of the entire governance system, not just individual agent performance, and drive continuous improvement of controls, thresholds, and processes.
Incident response. When a material compliance incident occurs — a data breach, a harmful AI output, a regulatory inquiry — the Manage phase provides response playbooks that satisfy notification requirements under GDPR Article 33/34 and EU AI Act serious incident reporting obligations.

The Manage phase ensures that governance is not just observed but actively maintained. It is the phase that separates mature AI operations from organisations that are one audit away from discovering their compliance has silently expired.

Why Lifecycle Governance Beats One-Time Audits

The traditional audit model assumes that a system assessed as compliant on Monday will remain compliant on Tuesday. For AI systems, this assumption is fundamentally flawed.

AI agents evolve. A RAG-based agent that ingests product catalogues, medical literature, or regulatory updates is — by design — a system that changes its knowledge and behaviour over time. An audit that assessed the agent's knowledge base in January has no bearing on what the agent knows in March.

Threat landscapes shift. New prompt injection techniques, adversarial patterns, and data exfiltration vectors emerge continuously. Security controls validated during an annual penetration test may be bypassed by techniques developed the following month.

Regulations tighten. The EU AI Act's enforcement phases extend through 2026 and beyond. GDPR enforcement continues to set new precedents. Organisations that treat compliance as a fixed target will find the target has moved by the time their next audit arrives.

The M3 Framework addresses this by making governance continuous by design. The Mount phase establishes the baseline. The Monitor phase detects deviations in real time. The Manage phase corrects course before deviations become violations. The cycle repeats — not annually, but continuously.

How Sinaptic® DROID+ Implements the M3 Framework

Every Sinaptic® DROID+ agent deployment follows the M3 lifecycle natively. The framework is not an optional add-on — it is embedded in the platform's operational architecture.

During the Mount phase, each agent is deployed with a documented scope definition, risk classification, data flow mapping, and stakeholder register. HITL thresholds are configured per scenario — from full autonomy on product FAQ responses to mandatory human approval on healthcare referrals or financial advice. Integration points with client systems (Shopify, HubSpot, Booksy, HL7 FHIR) are catalogued with their specific compliance obligations.

During the Monitor phase, the Sinaptic® DROID+ white-label admin panel provides real-time dashboards for conversation quality, security events, drift metrics, and compliance KPIs. The Sinaptic® Intent Firewall continuously enforces guardrails — blocking prompt injection attempts, preventing data exfiltration, and constraining agent behaviour to defined scope boundaries. Every intervention is logged and available for audit.

During the Manage phase, platform operators conduct periodic reviews using M3 report templates. Knowledge base updates follow governed RAG pipelines with Sinaptic DLP enforcement. Incident response procedures are pre-configured, and corrective actions are tracked through to verification. The white-label admin panel includes full audit logs, RBAC controls, and compliance reporting that clients can present directly to their regulators.

Sinaptic® DROID+ does not just comply with standards — it operationalises them through the M3 Framework, turning compliance from a periodic burden into a continuous operational advantage.

M3 and Sinaptic: Governance Meets Enforcement

The M3 Framework defines what must be governed. Sinaptic® — the AI security and compliance platform also developed by Julius Gromyko — provides the how. Together, they form a complete governance-and-enforcement stack.

Sinaptic's Intent Firewall enforces the controls that M3 specifies. When M3 requires that an agent must not disclose personal data in responses, Sinaptic's DLP layer detects and blocks such disclosures in real time. When M3 requires that prompt injection attempts are logged and reported, Sinaptic's security engine identifies the attack pattern, blocks the interaction, and generates the audit trail that the M3 Manage phase uses for incident reporting.

This separation of concerns — M3 for governance definition, Sinaptic for runtime enforcement — means organisations can adopt the M3 Framework independently of any specific technology platform. The framework is open and technology-agnostic. But organisations that deploy on Sinaptic® DROID+ get the enforcement layer built in, with zero additional integration effort.

Which Regulations Does M3 Cover?

The M3 Framework provides cross-reference mappings to five major regulatory and standards regimes:

GDPR — data protection by design (Art. 25), lawful basis for processing (Art. 6), automated decision-making rights (Art. 22), data protection impact assessments (Art. 35), breach notification (Art. 33/34).
EU AI Act — risk classification (Art. 6), risk management systems (Art. 9), data governance (Art. 10), transparency obligations (Art. 13), human oversight (Art. 14), accuracy and robustness (Art. 15), serious incident reporting.
ISO 27001 — information security management system controls (Annex A), risk assessment (Clause 6.1), monitoring and measurement (Clause 9.1), management review (Clause 9.3), continual improvement (Clause 10.1).
ISO 42001 — AI management system (Clauses 4–10), AI-specific risk controls (Annex B), AI impact assessment, data quality management, transparency, and human oversight requirements.
NIST AI RMF — Govern (establish AI governance), Map (contextualise risks), Measure (assess and track risks), Manage (prioritise and act on risks).

For each M3 control objective, the cross-reference matrix identifies which specific clauses, articles, or controls in each regime are satisfied. This eliminates the duplicated effort of maintaining separate compliance programmes for overlapping requirements — a problem that costs enterprise organisations thousands of hours annually.

Getting Started with the M3 Framework

The M3 Framework is designed for practical adoption, not theoretical elegance. Here is how to begin:

Visit m3framework.org. The framework documentation, control mappings, and templates are published openly and free for internal use. No licence fees, no registration walls.
Start with Mount. Pick one AI system in your organisation — ideally the highest-risk one — and work through the Mount phase: define scope, classify risks, map to regulations, establish data governance, and configure human oversight. This exercise alone will reveal governance gaps you did not know existed.
Instrument monitoring. Once the Mount baseline is established, implement the monitoring controls. Even basic logging and alerting will dramatically improve your compliance posture compared to periodic manual reviews.
Close the loop with Manage. Schedule your first management review. Use the data from monitoring to drive concrete improvements. Document everything — this documentation is what auditors and regulators will evaluate.
Scale across systems. Once the M3 lifecycle is proven on one system, extend it to all AI deployments. The framework is designed to scale from a single chatbot to an enterprise portfolio of AI agents.

The Bottom Line

AI compliance is not a destination — it is an operational discipline. Regulations will continue to evolve. AI systems will continue to learn and change. Threat landscapes will continue to shift. The only sustainable approach is a governance framework that treats compliance as a continuous lifecycle: Mount the system with proper controls. Monitor it in real time. Manage it through structured review and improvement cycles.

The M3 Framework provides that lifecycle. It is open, it is free for internal use, and it maps across the five regulatory regimes that matter most for enterprise AI. Whether you deploy on Sinaptic® DROID+ or build your own stack, M3 gives your organisation a single operational framework that replaces fragmented compliance checklists with continuous, auditable governance.

The organisations that will lead in AI are not the ones with the most advanced models. They are the ones with the most mature governance. M3 is how you get there.