Reifegradmodell & RoadmapMaturity Model & Roadmap

Zusammenfassung

Die KI-native Organisation wird nicht über Nacht gebaut. Diese Dimension definiert ein 5-Phasen-Reifemodell (Foundation → Structured Assistance → Managed Autonomy → Adaptive Organization → Enterprise), gemappt auf Batesons (1972) Lernniveau-Hierarchie, eine 3-Phasen-Operational-Roadmap (Architect → Trainer → Orchestrator) für die Evolution der Principal-Rolle, und den VCOM Continuous Improvement Cycle mit drei Zeitskalen (operativ, taktisch, strategisch).

Zwei komplementäre Modelle arbeiten zusammen: Das Reifemodell (diagnostisch) beschreibt, wie die Organisation auf jeder Stufe aussieht. Die Roadmap (präskriptiv) beschreibt, was der Principal auf jeder Stufe tut. Verwechslung dieser beiden ist eine dokumentierte Ursache für Implementierungsfehler.

Kontext im VCOM-Framework

Das Reifegradmodell integriert alle anderen Dimensionen: Jede Reifephase impliziert ein Mindestmaß an Entwicklung in Organisation, Governance, Prozessen, Wissen, Technologie und Performance-Messung. Es ist die zeitliche Dimension von VCOM. Zwei häufige Fehlermodi motivieren diese Dimension: Premature Autonomy (L3–L4-Agenten ohne Governance-Fundament) und Governance Rigidity (Phase-1-Kontrollen bei Phase-3-Agentenpopulationen).

Summary

An AI-native organization is not built overnight. This dimension defines a 5-phase maturity model (Foundation → Structured Assistance → Managed Autonomy → Adaptive Organization → Enterprise), mapped to Bateson's (1972) learning level hierarchy, a 3-phase operational roadmap (Architect → Trainer → Orchestrator) for the principal's evolving role, and the VCOM Continuous Improvement Cycle with three timescales (operational, tactical, strategic).

Two complementary models work together: The maturity model (diagnostic) describes what the organization looks like at each level. The roadmap (prescriptive) describes what the principal does at each stage. Confusion between these two is a documented source of implementation failure.

Context within the VCOM Framework

The maturity model integrates all other dimensions: each phase implies a minimum level of development in organization, governance, processes, knowledge, technology, and performance measurement. It is the temporal dimension of VCOM. Two common failure modes motivate this dimension: Premature Autonomy (deploying L3–L4 agents without governance foundations) and Governance Rigidity (maintaining Phase 1 controls at Phase 3 agent populations).

Das 5-Phasen-Reifemodell

Jede Phase entspricht einem qualitativ unterschiedlichen Typ adaptiver Kapazität, fundiert in Batesons (1972) Hierarchie von Lerntypen:

Phase 1: Foundation (Bateson Level 0)

KI als persönliches Produktivitäts-Tool. Einzelne Agenten bei L0–L1 produzieren fixe Antworten ohne erfahrungsbasierte Modifikation. Keine Inter-Agenten-Koordination. Principal reviewt alle Outputs.

Transitionskriterien: 3+ wiederkehrende Tasks identifiziert; >30% Principal-Zeit auf automatisierbare Tasks.

Phase 2: Structured Assistance (Bateson Level I)

Agenten passen operative Parameter basierend auf Erfahrung an, innerhalb fixer Workflow-Strukturen. 5–10 Agenten bei L1–L2. Governance führt Standing und Prohibited Policies über einen Basis-GaaS-Layer ein.

Transitionskriterien: 5+ Workflows bei L2; Goal Accuracy >80%; sinkende Intervention Rate über 3+ Monate.

Phase 3: Managed Autonomy (Level I → II)

Agenten beginnen, Workflow-Strukturen zu verändern, nicht nur Parameter. 10–30 Agenten bei L2–L3. Volle GaaS Policy-Taxonomie, Knowledge Graph mit episodischer und semantischer Schicht, RepuNet Peer-Qualitätssicherung.

Transitionskriterien: 20+ Workflows bei L2–L3; <5% Intervention Rate; Trust Scores treiben Autonomie; Lifecycle Management operativ.

Phase 4: Adaptive Organization (Bateson Level II)

Agenten innovieren systematisch, transferieren Fähigkeiten über Domänen und schaffen neue Fähigkeiten ohne Principal-Spezifikation. 30–100+ Agenten bei L3–L4. Dynamische Team-Formation, strategische Agenten (S4), interne Marktmechanismen.

Phase 5: Enterprise / AI-Native (Level II → III)

Nähert sich der Grenze zwischen Learning II und III, wo die Organisation beginnt, ihr eigenes Betriebsmodell zu transformieren. Aspirational — Stand 2026 von keiner bekannten Organisation erreicht.

3-Phasen-Operational-Roadmap

• Phase I: The Architect (Monate 1–3): System designen. VSM-Struktur aufbauen. LangGraph-Backbone und Knowledge Graph. L2-Agenten für Kernprozesse. Basis-GaaS. Deliverables: Organisationsverfassung, 3–5 Produktions-Workflows, Cockpit mit Basis-Observability.
• Phase II: The Trainer (Monate 4–6): Feedback-Schleifen und Kalibrierungstests. System-Prompt-Verfeinerung. Spezialisten-Cluster. KPI-Framework und Trust Scores. Deliverables: Performance-Baselines, Trust-Score-getriebene Autonomie operativ, Knowledge Graph durch Gardener gewartet.
• Phase III: The Orchestrator (Monat 6+): L3/L4-Autonomie. Marktmechanismen. Strategische Agenten (S4). >80% Principal-Zeit auf S4/S5-Aktivitäten. Deliverables: Agenten bei L3+ für Routinebetrieb, strategische Agenten generieren umsetzbare Insights.

Failure Modes bei Phasen-Transitionen

Transition	Fehlermodus	Symptom	Prävention
1 → 2	Premature Workflow-Formalisierung	Starre, minderwertige Prozesse	≥ 10 manuelle Ausführungen vor Formalisierung
2 → 3	Governance Gap	L3-Agenten ohne adäquate GaaS-Abdeckung	100% Policy-Abdeckung für Agent-Domäne vor L3
2 → 3	Knowledge-Graph-Vernachlässigung	Kein geteiltes Organisationsgedächtnis	Knowledge-Graph-Health in Transitionskriterien
3 → 4	Unzureichende Attribution	Multi-Agent-Workflows ohne Outcome-Attribution	Stage-based Attribution vor Phase 4
3 → 4	Überabhängigkeit vom Principal	Principal bleibt primärer QA-Mechanismus	Funktionales RepuNet und Peer Review vor Phase 4

Phase Regression & Recovery

Reife-Regression ist eine normale organisationale Reaktion auf Störungen, kein Versagen. Drei Typen:

• Model-Change Regression: Großes Modell-Upgrade verursacht Verhaltens-Drift. Organisation regrediert temporär. Recovery: Autonomie auf L1–L2 reduzieren, Kalibrierungstests wiederholen, Baselines neu erstellen. 1–4 Wochen.
• Governance-Failure Regression: Signifikanter Governance-Breach. Recovery: Alle Agenten sofort auf L1 reduzieren, Audit durchführen, Governance-Lücke schließen. 2–6 Wochen.
• Market-Shift Regression: Fundamentale Umfeldänderung invalidiert Workflows und Strategien. Recovery: Autonomie nicht reduzieren. Domänen-Strategien aktualisieren, Workflows revidieren, Agenten auf neuem Kontext re-trainieren.

VCOM Continuous Improvement Cycle

Vereint cross-dimensionale Feedback-Loops in einer dreiskaligen Architektur (Argyris, 1991):

• Operativer Loop (Per-Task): Automatisch, <5% Overhead. Assign → Execute → Measure → Evaluate → Record → Feedback.
• Taktischer Loop (Zweimal monatlich): Single-Loop Learning. Review → Analyze → Diagnose → Plan → Adapt → Capture. Hier wird das Empirische Lernprotokoll (Dim 09) am häufigsten invokiert.
• Strategischer Loop (Vierteljährlich): Double-Loop Learning. Erfordert Principal-Teilnahme. Environmental Scan → Portfolio Review → Strategy Adjustment → Maturity Assessment → Constitution Review → Roadmap Update.

Autonomer Operationsmodus

Bei Abwesenheit des Principals:

• Strategischer Loop: Suspendiert. Keine Strategieänderungen. Environmental Scanning läuft weiter; Ergebnisse für Rückkehr akkumuliert.
• Taktischer Loop: Reduzierter Scope. Prozessverbesserungen innerhalb bestehender Strukturen erlaubt. Budget-Reallokation auf ±10% beschränkt.
• Operativer Loop: Normalbetrieb innerhalb bestehender Einschränkungen.

Conservation Protocol (ab Tag 7 / Gelb→Orange): Risikoappetit eine Stufe konservativer, Explorations-Budgets −50%, alle L3+-Aktionen erfordern Manager-Bestätigung.

Self-Reflection Checkpoints (S3 Common Sense Framework)

Meta-Level-Organisationsgesundheits-Assessment über zehn diagnostische Dimensionen (Bockelbrink et al., 2022):

1. Purpose Clarity — geteiltes Verständnis des Organisationszwecks
2. Strategy Coherence — Initiativen aligned mit Strategic Themes
3. Value Delivery — Outputs dienen gesteckten Zielen
4. Sense-and-Respond — Spannungen erkannt und effektiv geroutet
5. Experimentation Culture — Empirisches Lernprotokoll genutzt
6. Autonomy & Accountability — angemessene Level mit Verantwortung
7. Collaboration Effectiveness — Cross-Agent-Workflows funktionieren
8. Learning Orientation — neues Wissen erfasst und angewandt
9. Cultural Coherence — Verhaltensnormen konsistent mit Verfassung
10. Shared Mental Models — konsistente Domänenannahmen über Agenten

Jede Dimension 0.0–1.0 gescored. Checkpoints wöchentlich (Normalbetrieb) und täglich (Principal-Abwesenheit). Unter 0.5 oder sinkender Trend über 3 Checkpoints löst Spannungssignal aus.

The 5-Phase Maturity Model

Each phase corresponds to a qualitatively different type of adaptive capacity, grounded in Bateson's (1972) hierarchy of learning types:

Phase 1: Foundation (Bateson Level 0)

AI as a personal productivity tool. Individual agents at L0–L1 produce fixed responses with no experience-based modification. No inter-agent coordination. Principal reviews all outputs.

Transition criteria: 3+ recurring tasks identified; >30% principal time on automatable tasks.

Phase 2: Structured Assistance (Bateson Level I)

Agents adjust operational parameters based on experience, within fixed workflow structures. 5–10 agents at L1–L2. Governance introduces standing and prohibited policies through a basic GaaS layer.

Transition criteria: 5+ workflows at L2; goal accuracy >80%; declining intervention rate over 3+ months.

Phase 3: Managed Autonomy (Level I → II)

Agents begin to modify workflow structures, not merely parameters. 10–30 agents at L2–L3. Full GaaS policy taxonomy, knowledge graph with episodic and semantic layers, RepuNet peer quality assurance.

Transition criteria: 20+ workflows at L2–L3; <5% intervention rate; trust scores driving autonomy; lifecycle management operational.

Phase 4: Adaptive Organization (Bateson Level II)

Agents systematically innovate, transfer capabilities across domains, and create new capabilities without principal specification. 30–100+ agents at L3–L4. Dynamic team formation, strategic agents (S4), internal market mechanisms.

Phase 5: Enterprise / AI-Native (Level II → III)

Approaches the boundary between Learning II and III, where the organization begins to transform its own operating model. Aspirational — no known organization has achieved this as of 2026.

3-Phase Operational Roadmap

• Phase I: The Architect (Months 1–3): Design the system. Build VSM structure, LangGraph backbone, and knowledge graph. L2 agents for core processes. Basic GaaS. Deliverables: Organizational constitution, 3–5 production workflows, Cockpit with basic observability.
• Phase II: The Trainer (Months 4–6): Feedback loops and calibration tests. System prompt refinement. Specialist clusters. KPI framework and trust scores. Deliverables: Performance baselines, trust-score-driven autonomy operational, Knowledge Graph maintained by Gardener.
• Phase III: The Orchestrator (Month 6+): L3/L4 autonomy. Market mechanisms. Strategic agents (S4). >80% principal time on S4/S5 activities. Deliverables: Agents at L3+ for routine operations, strategic agents generating actionable insights.

Failure Modes at Phase Transitions

Transition	Failure Mode	Symptom	Prevention
1 → 2	Premature workflow formalization	Rigid, poor-quality processes	≥ 10 manual executions before formalizing
2 → 3	Governance gap	L3 agents without adequate GaaS coverage	100% policy coverage for agent domain before L3
2 → 3	Knowledge graph neglect	No shared organizational memory	Include knowledge graph health in transition criteria
3 → 4	Insufficient attribution	Multi-agent workflows without outcome attribution	Implement stage-based attribution before Phase 4
3 → 4	Over-reliance on principal	Principal remains primary QA mechanism	Require functional RepuNet and peer review before Phase 4

Phase Regression & Recovery

Maturity regression is a normal organizational response to disruption, not failure. Three types:

• Model-Change Regression: Major model upgrade causes behavioral drift. Organization temporarily regresses. Recovery: Reduce autonomy to L1–L2, re-run calibration, re-establish baselines. 1–4 weeks.
• Governance-Failure Regression: Significant governance breach. Recovery: Immediately reduce all agents to L1, audit, close governance gap. 2–6 weeks.
• Market-Shift Regression: Fundamental environmental change invalidates workflows. Recovery: Do not reduce autonomy. Update domain strategies, revise workflows, re-train agents on updated context.

VCOM Continuous Improvement Cycle

Unifies cross-dimensional feedback loops into a three-timescale architecture (Argyris, 1991):

• Operational Loop (Per-Task): Automatic, <5% overhead. Assign → Execute → Measure → Evaluate → Record → Feedback.
• Tactical Loop (Bi-weekly): Single-loop learning. Review → Analyze → Diagnose → Plan → Adapt → Capture. Most common invocation point for the Empirical Learning Protocol (Dim 09).
• Strategic Loop (Quarterly): Double-loop learning. Requires principal participation. Environmental Scan → Portfolio Review → Strategy Adjustment → Maturity Assessment → Constitution Review → Roadmap Update.

Autonomous Operation Mode

When the principal is unavailable:

• Strategic Loop: Suspended. No strategy changes. Environmental scanning continues; findings accumulated for return.
• Tactical Loop: Reduced scope. Process improvements within existing structures permitted. Budget reallocation restricted to ±10%.
• Operational Loop: Normal operation within existing constraints.

Conservation Protocol (after day 7 / Yellow→Orange): Risk appetite one level conservative, exploration budgets −50%, all L3+ actions require Manager confirmation.

Self-Reflection Checkpoints (S3 Common Sense Framework)

Meta-level organizational health assessment across ten diagnostic dimensions (Bockelbrink et al., 2022):

1. Purpose clarity — shared understanding of organizational purpose
2. Strategy coherence — initiatives aligned with Strategic Themes
3. Value delivery — outputs serving stated objectives
4. Sense-and-respond capability — tensions detected and routed effectively
5. Experimentation culture — Empirical Learning Protocol invoked when appropriate
6. Autonomy and accountability — appropriate levels with corresponding accountability
7. Collaboration effectiveness — cross-agent workflows completing without excessive friction
8. Learning orientation — new knowledge captured and applied
9. Cultural coherence — behavioral norms consistent with constitution
10. Shared mental models — consistent domain assumptions across agents

Each dimension scored 0.0–1.0. Checkpoints run weekly (normal) and daily (principal absence). Below 0.5 or declining trend across 3 checkpoints triggers tension signal.

Maturity Assessment Matrix

Selbstbewertungs-Tool: Pro Dimension wird der aktuelle Reifegrad bestimmt.

Maturity Assessment Matrix

Self-assessment tool: current maturity level determined per dimension.

Phase-Transition-Checkliste

Phase Transition Checklist

phase_transitions:
  phase_1_to_2:
    criteria:
      - "3+ recurring tasks identified"
      - "Principal spending >30% on automatable tasks"
      - "Basic agent capability understanding"
    gate: "All criteria met"

  phase_2_to_3:
    criteria:
      - "5+ workflows running at L2"
      - "Goal accuracy >80%"
      - "Human intervention rate declining 3+ months"
      - "GaaS layer operational with policy-as-code"
    gate: "All criteria met"
    failure_check: "If 3+ dimensions lag, transition not recommended"

  phase_3_to_4:
    criteria:
      - "20+ workflows at L2-L3, <5% intervention"
      - "Trust score system producing reliable adjustments"
      - "Knowledge graph actively used by agents"
      - "Agent lifecycle management operational"
      - "RepuNet and peer review functional"
    gate: "All criteria met"

  phase_4_to_5:
    criteria:
      - "Agents demonstrate L4 strategic judgment"
      - "Value generated in non-designed domains"
      - "Governance robust through stress tests"
    gate: "All criteria met + principal assessment"
    note: "Aspirational as of Feb 2026"

Phasengerechte Metriken

Verschiedene Metriken sind in verschiedenen Reifephasen relevant. Phase-4-Metriken in Phase 1 erzeugen nur Rauschen.

Phase-Appropriate Metrics

Different metrics matter at different maturity phases. Tracking Phase 4 metrics during Phase 1 creates noise.

Pull-System für Veränderung

S3-basierter Ansatz (Bockelbrink et al., 2022): Veränderung wird dort eingeführt, wo nachgewiesener Bedarf (Spannung) besteht — nicht top-down als Gesamttransformation:

1. Mit einer einzigen Domäne starten: Ein Agent Manifest für den schmerzhaftesten wiederkehrenden Task
2. Unter Consent operieren: Algorithmisches Consent-Protokoll validieren
3. Fraktal wachsen: Jede neue Domäne ist eigenständig, verbunden durch Double-Linking
4. Spannungen treiben Expansion: Neue Domänen als Reaktion auf beobachtete Bedürfnisse, nicht theoretische Vollständigkeit

Pull System for Change

S3-based approach (Bockelbrink et al., 2022): change is introduced where there is demonstrated need (tension) — not pushed top-down as total transformation:

1. Start with a single domain: one Agent Manifest for the most painful recurring task
2. Operate under consent: validate the Algorithmic Consent protocol works
3. Grow fractally: each new domain is self-contained, connected through double-linking
4. Let tensions drive expansion: new domains created in response to observed needs, not theoretical completeness