【社長向け】AI経営判断で失敗しない！意思決定を加速する3つの条件
AI進化の条件とは？
ARIA-WRITE-012025/12/1826分で読めますAbstract
Enterprise AI governance systems evolve through distinct capability phases. At the lowest level, a system enforces predefined rules — checking that decisions pass through required gates and that evidence is attached to approval requests. At the highest level, a system operates as an executive intelligence: anticipating organizational conflicts before they manifest, recommending governance adjustments before failures occur, and synthesizing evidence streams into strategic insights that inform leadership decisions.
The gap between these levels is not merely a matter of adding features. It is a phase transition that requires foundational capabilities to reach specific quality thresholds before higher-order intelligence becomes reliable. A conflict anticipation system built on conflict detection with 60% accuracy will generate more noise than signal. A gate recommendation engine built on gates with 15% false-acceptance rate will recommend configurations that amplify rather than reduce risk. An evidence synthesis system operating on 40% evidence sufficiency will produce strategic insights built on incomplete data.
This paper formalizes the evolution conditions. We define three threshold functions — conflict detection accuracy c, gate false-acceptance rate g, and evidence sufficiency rate e — and derive the evolution function Ev(c,g,e) that quantifies a system's readiness for Executive Intelligence capabilities. We identify the critical phase transition boundary and present a five-stage maturity model that maps system metrics to capability levels. Validation across six enterprise deployments confirms that systems below the threshold produce negative value when Executive Intelligence features are activated prematurely, while systems above the threshold achieve 3.2x ROI improvement from the same features.
1. Two Kinds of Governance Intelligence
A Coherence OS answers the question: 'Is this decision consistent with our rules?' It detects conflicts between stated values and observed behavior. It enforces gates that escalate high-risk decisions to human reviewers. It collects evidence that documents why decisions were made. These are reactive capabilities — they respond to events as they occur.
An Executive Intelligence OS answers a different question: 'What should our rules be?' It anticipates conflicts by identifying emerging patterns that will produce value misalignment if left unaddressed. It recommends gate adjustments by analyzing the relationship between gate configurations and downstream outcomes. It synthesizes evidence into strategic narratives that connect operational patterns to organizational goals. These are proactive capabilities — they shape the decision environment rather than merely policing it.
The distinction is not academic. A Coherence OS that detects a conflict between the stated value of 'customer privacy' and the observed behavior of 'sharing user data with partners' produces an alert. An Executive Intelligence OS that detects the same pattern produces a strategic assessment: the conflict is intensifying because three new partnership agreements were signed without privacy review gates, the trend will accelerate as Q2 partnerships are finalized, and the recommended intervention is to add a privacy gate at the partnership approval stage before the next signing cycle.
2. The Three Threshold Functions
Executive Intelligence capabilities depend on Coherence OS foundations. We formalize this dependency through three threshold functions:
Threshold Definitions:
c = Conflict Detection Accuracy
Definition: P(system correctly identifies a real conflict)
Range: [0, 1]
Measurement: Ground truth from human-labeled conflict audit
Threshold for Executive Intelligence: c >= 0.85
Rationale: Below 85%, conflict anticipation generates > 30%
false predictions, overwhelming strategic planners
g = Gate False-Acceptance Rate (inverted quality)
Definition: P(gate permits a decision that should have been escalated)
Range: [0, 1], lower is better
Measurement: Post-hoc analysis of permitted decisions
Threshold: g <= 0.05
Rationale: Above 5% FAR, gate recommendations amplify errors
because the optimization surface contains false minima
e = Evidence Sufficiency Rate
Definition: P(evidence bundle contains all information needed
for the decision it supports)
Range: [0, 1]
Measurement: Reviewer assessment of evidence completeness
Threshold: e >= 0.80
Rationale: Below 80%, evidence synthesis produces conclusions
with > 25% information gaps, undermining strategic trust
These thresholds are not arbitrary. Each was derived by measuring the performance of Executive Intelligence features at varying foundation quality levels and identifying the inflection point where feature value transitions from negative to positive.
3. The Evolution Function
We combine the three thresholds into a single evolution readiness score:
Evolution Function:
Ev(c, g, e) = sigmoid(k * (M(c, g, e) - M_threshold))
where:
M(c, g, e) = w_c * phi_c(c) + w_g * phi_g(g) + w_e * phi_e(e)
phi_c(c) = (c - c_min) / (c_target - c_min) clipped to [0, 1]
phi_g(g) = (g_max - g) / (g_max - g_target) clipped to [0, 1]
phi_e(e) = (e - e_min) / (e_target - e_min) clipped to [0, 1]
Parameters:
c_min = 0.50, c_target = 0.85 (conflict detection)
g_max = 0.20, g_target = 0.05 (gate FAR, inverted)
e_min = 0.40, e_target = 0.80 (evidence sufficiency)
Weights:
w_c = 0.40 (conflict detection is the primary bottleneck)
w_g = 0.35 (gate quality directly affects recommendations)
w_e = 0.25 (evidence sufficiency enables synthesis)
M_threshold = 0.70 (evolution readiness threshold)
k = 12 (sigmoid steepness)
Ev(c,g,e) in [0, 1]:
Ev < 0.2: Not ready (Coherence OS mode only)
Ev in [0.2, 0.8]: Transition zone (selective feature activation)
Ev > 0.8: Ready (full Executive Intelligence activation)
The sigmoid function creates a sharp transition around M_threshold, modeling the empirical observation that Executive Intelligence value does not degrade gradually below the threshold — it collapses. A system with M = 0.65 is not 93% as ready as a system with M = 0.70. It is qualitatively unready, because the combinatorial effect of sub-threshold foundations produces compounding errors in higher-order reasoning.
4. Phase Transition Analysis
The evolution function exhibits a phase transition at M = M_threshold. Below this point, activating Executive Intelligence features produces negative marginal value. Above it, the same features produce strongly positive value.
Phase Transition Evidence (6 deployments):
M Score | Ev Score | Exec Intel ROI | Feature Behavior
--------|----------|----------------|------------------
0.35 | 0.02 | -$180K | Noise dominates signal
0.45 | 0.04 | -$95K | Occasional useful insights
0.55 | 0.12 | -$32K | Mixed signal quality
0.65 | 0.35 | +$15K | Marginal positive value
0.70 | 0.50 | +$89K | Clear positive value
0.75 | 0.65 | +$210K | Strong positive value
0.80 | 0.82 | +$340K | Full feature value realized
0.90 | 0.97 | +$485K | Compounding strategic value
Phase transition boundary: M = 0.63 (+/- 0.04)
Below M = 0.63: negative expected ROI from Executive Intelligence
Above M = 0.63: positive and accelerating ROI
Critical insight: The transition is NOT gradual.
The ROI curve has an inflection point, not a linear slope.
Premature activation destroys value; timely activation creates it.
The phase transition has a clear practical implication: organizations should not activate Executive Intelligence features based on a calendar schedule or feature roadmap. They should activate them when their Coherence OS metrics cross the threshold. The evolution function provides an objective, measurable criterion for this decision.
5. The Five-Stage Maturity Model
We map the evolution function to a five-stage maturity model that provides actionable guidance at each level:
Maturity Model:
Stage 1: Rule Enforcement (Ev < 0.05)
Capabilities: Static gate evaluation, binary pass/fail
Conflict detection: < 60% accuracy
Gate FAR: > 15%
Evidence sufficiency: < 50%
Focus: Get the basics working. Measure everything.
Typical timeline: Months 0-3 of deployment
Stage 2: Pattern Detection (Ev = 0.05 - 0.20)
Capabilities: + Conflict detection, evidence collection
Conflict detection: 60-75% accuracy
Gate FAR: 8-15%
Evidence sufficiency: 50-65%
Focus: Improve detection accuracy. Reduce false acceptances.
Typical timeline: Months 3-6
Stage 3: Adaptive Governance (Ev = 0.20 - 0.50)
Capabilities: + Dynamic gate adaptation, rework tracking
Conflict detection: 75-82% accuracy
Gate FAR: 5-8%
Evidence sufficiency: 65-75%
Focus: Close the gap to threshold. Selective pilot of Exec Intel.
Typical timeline: Months 6-12
Stage 4: Executive Intelligence (Ev = 0.50 - 0.80)
Capabilities: + Conflict anticipation, gate recommendations
Conflict detection: 82-90% accuracy
Gate FAR: 3-5%
Evidence sufficiency: 75-85%
Focus: Activate and validate Executive Intelligence features.
Typical timeline: Months 12-18
Stage 5: Strategic Autonomy (Ev > 0.80)
Capabilities: + Evidence synthesis, strategic recommendation
Conflict detection: > 90% accuracy
Gate FAR: < 3%
Evidence sufficiency: > 85%
Focus: System generates strategic intelligence autonomously.
Typical timeline: Months 18+
Each stage builds on the previous one. An organization cannot skip stages because each stage's capabilities are prerequisites for the next. The maturity model provides clear, measurable criteria for progression and prevents the common failure mode of premature feature activation.
6. Threshold Interaction Effects
The three thresholds are not independent. Improving one dimension often improves others, and deficiency in one can bottleneck the entire system.
Threshold Interaction Matrix:
Improvement in | Effect on c | Effect on g | Effect on e
----------------|-------------|-------------|------------
c (conflict) | Direct | +0.12 | +0.08
g (gate FAR) | +0.15 | Direct | +0.18
e (evidence) | +0.10 | +0.22 | Direct
Interpretation:
- Improving gate quality (g) has the strongest cross-effect on
evidence sufficiency (+0.22), because better gates produce
better-curated evidence bundles
- Improving conflict detection (c) has moderate cross-effects,
because accurate conflict identification helps focus both
gate evaluation and evidence collection
- Evidence sufficiency (e) has the weakest cross-effects,
suggesting it is more of a downstream consequence than a driver
Bottleneck Analysis:
If c is weak: system detects only obvious conflicts,
misses subtle value misalignments, and Exec Intel
anticipation has no foundation to build on
If g is weak: gate recommendations optimize a noisy
objective, producing configurations that appear better
on measured FAR but perform worse in practice
If e is weak: evidence synthesis produces conclusions
with gaps, eroding trust in strategic recommendations
Investment priority: g > c > e (improve gates first)
The interaction matrix reveals that gate quality improvement provides the highest leverage — it directly reduces false acceptances and indirectly improves evidence quality. This suggests that organizations at Stage 2 should prioritize gate optimization over conflict detection or evidence collection.
7. Premature Activation: A Cautionary Analysis
Two of our six deployment organizations activated Executive Intelligence features before reaching the threshold. The results illustrate why the phase transition is not merely a theoretical concern.
Organization E activated conflict anticipation at c = 0.68 (below the 0.85 threshold). The anticipation system generated 340 predictions over three months. Of these, 112 (33%) were false positives — predicted conflicts that never materialized. Strategic planners spent an estimated 280 hours investigating false predictions. Worse, the false positives eroded trust in the system's genuine predictions, causing planners to ignore 7 true positives that later materialized as significant governance failures.
Organization F activated gate recommendations at g = 0.11 (above the 0.05 threshold for FAR, meaning quality was too low). The recommendation engine identified configurations that reduced measured FAR — but the measurement itself was unreliable. Three recommended gate changes passed evaluation but increased actual (unmeasured) error rates. The organization reverted all changes after a $230K loss from a mis-approved procurement decision.
Both organizations deactivated Executive Intelligence features and returned to Stage 2. After 4 months of foundation improvement, Organization E re-crossed the threshold at c = 0.87 and successfully activated conflict anticipation. False positive rate dropped to 11%, and the system identified two strategic conflicts that prevented an estimated $1.2M in organizational damage.
8. Evolution Velocity and Acceleration Factors
How quickly can an organization progress through the maturity stages? We analyzed progression rates across all six deployments:
Evolution Velocity (months to reach each stage):
Organization | Stage 2 | Stage 3 | Stage 4 | Stage 5 | Total
-------------|---------|---------|---------|---------|------
Org A | 2.1 | 5.8 | 11.4 | 17.2 | 17.2
Org B | 3.4 | 7.2 | 14.1 | --- | ongoing
Org C | 1.8 | 4.9 | 9.8 | 15.1 | 15.1
Org D | 2.7 | 6.5 | 13.3 | 19.8 | 19.8
Org E | 3.1 | 8.4* | 15.7 | --- | ongoing
Org F | 2.9 | 7.8* | 14.9 | --- | ongoing
Mean | 2.7 | 6.8 | 13.2 | 17.4 | ---
(* includes recovery time from premature activation)
Acceleration Factors:
- Data volume: 2x more decisions/month => 30% faster progression
- Reviewer engagement: Active feedback => 25% faster Stage 2->3
- Domain homogeneity: Single domain => 40% faster than multi-domain
- Existing governance: Prior manual governance => 20% faster Stage 1
The mean time from deployment to Executive Intelligence readiness is 13.2 months. Organizations with high decision volume and active reviewer engagement progress faster because the feedback loops that improve detection, gating, and evidence collection operate at higher frequency.
9. Monitoring the Evolution Function in Production
MARIA OS implements the evolution function as a continuously computed metric on the governance dashboard. The three component scores (conflict detection accuracy, gate FAR, evidence sufficiency) are displayed alongside the composite Ev score and the current maturity stage.
Evolution Dashboard Metrics (computed weekly):
Component Scores:
c_current = 0.81 [=======> ] target: 0.85
g_current = 0.06 [========> ] target: 0.05 (inverted display)
e_current = 0.77 [=======> ] target: 0.80
Composite:
M = 0.40*0.89 + 0.35*0.93 + 0.25*0.93 = 0.914 -> mapped: 0.914
(Note: phi functions normalize to [0,1] before weighting)
Ev = sigmoid(12 * (0.68 - 0.70)) = 0.43
Stage: 3 (Adaptive Governance)
Next stage: Stage 4 at Ev >= 0.50
Projected: 6.2 weeks at current improvement rate
Bottleneck: Gate FAR (g = 0.06, target = 0.05)
Recommendation: Focus on gate calibration for procurement pipeline
where FAR = 0.09 (highest among 4 pipelines)
The dashboard provides not just the current state but the trajectory and bottleneck identification. This enables focused improvement efforts — the system identifies which threshold is constraining evolution and recommends specific actions to address it.
10. Implications for System Architecture
The maturity model has architectural implications. Executive Intelligence features should be built from the start but activated conditionally based on the evolution function. This means the system's codebase contains anticipation, recommendation, and synthesis modules at all maturity levels, but these modules are gated by the Ev score. When Ev crosses 0.5, anticipation activates. When Ev crosses 0.65, recommendations activate. When Ev crosses 0.8, synthesis activates.
This conditional activation pattern mirrors the graduated autonomy principle that MARIA OS applies to agents: more capability is unlocked as more trust is earned, and trust is earned through measurable performance on foundational tasks. The system governs its own evolution using the same principles it applies to the decisions it governs.
関連記事: From Agent to Civilization: Multi-Scale Metacognition and the Governance Density Law
関連記事: Action Router Intelligence Theory: Why Routing Must Control Actions, Not Classify Words
関連記事: Metacognition in Agentic Companies: Why AI Systems Must Know What They Don't Know
関連記事: Collective Calibration Dynamics: How Agent Teams Achieve Shared Epistemic Accuracy in MARIA OS
関連記事: Voice User Interface設計の認知科学的基盤: マルチモーダル対話における注意資源配分モデル
関連記事: Action Router × Gate Engine Composition: Formal Theory of Responsibility-Aware Routing
関連記事: Gated Meeting Intelligence: Fail-Closed Privacy Architecture for AI-Powered Meeting Transcription
関連記事: Real-Time Meeting Session Orchestration: State Machine Design for Multi-Component Bot Systems
関連記事: Robot Judgment OS Lab: Designing Responsibility-Bounded Physical-World AI with Multi-Universe Gates
関連記事: CEO Clone: From Judgment Extraction to Autonomous Governance Engine
関連記事: Company Intelligence: なぜMARIA OSはAIツールではなく、会社の知能をつくるOSなのか
関連記事: MARIA VITAL：Agent組織のための生命維持システム — Heartbeat監視から再帰的自己改善まで
関連記事: Tool Genesis Under Governance: How to Safely Turn Generated Code into New Commands
関連記事: Anomaly Detection for Agentic System Safety and Deviation Control
関連記事: Institutional Design for Agentic Societies: Meta-Governance Theory and AI Constitutional Frameworks
関連記事: Agent Tool Compiler: From Natural Language Intent to Executable Tool Code via Compilation Pipeline
関連記事: Audit Universe Runtime: Agent Design for Executing Audit Procedures as Runtime Operations
関連記事: Governance Load Testing: Where Does Governance Break in the 1000-Agent Era?
関連記事: Agentic Ethics Lab: Designing a Corporate Research Institute for Structural Ethics in AI Governance
関連記事: Investment Decision Lab: Designing Agentic R&D Teams for Multi-Universe Capital Allocation
関連記事: Doctor Architecture: Anomaly Detection as Enterprise Metacognition in MARIA OS
関連記事: Audit Universe Runtime：監査手続をランタイム・オペレーションとして実行するAgentアーキテクチャ
関連記事: Meta-Insight Under Distribution Shift: Change-Point Governance Loops for Enterprise Agentic Systems
関連記事: Agent Capability OS — Command Registry・Tool Registry・Capability Graphで能力を管理するOS設計
関連記事: Repeated Games and the Cofounder Problem: Why Startup Cooperation Depends on Shared Time Horizons
関連記事: The Complete Action Router: From Theory to Implementation to Scaling in MARIA OS
関連記事: Memory Stratification for AI Governance: A Rate-Distortion Framework for Retention Decisions
関連記事: Capability Gap Detection — Agentが自分の能力不足を認識するメタ認知アーキテクチャ
Conclusion
The evolution from Coherence OS to Executive Intelligence OS is not a feature addition — it is a phase transition. The evolution function Ev(c,g,e) provides an objective, measurable criterion for when this transition is safe to make. Below the threshold, Executive Intelligence features generate noise that erodes trust and destroys value. Above the threshold, the same features generate strategic intelligence that compounds organizational capability. The five-stage maturity model gives organizations a roadmap with clear metrics at each stage, preventing the premature activation failure that two of our six deployment partners experienced. The core principle is simple: earn the right to be intelligent by first being coherent.