The AI Maturity Model for Managers

The Measurement Gap: Why Enterprise AI Maturity Models Miss Managers

Enterprise technology investments frequently encounter an unexpected operational bottleneck. According to the RAND Corporation (2024), "By some estimates, more than 80 percent of AI projects fail—twice the rate of failure for information technology projects that do not involve AI." When initiatives collapse, failure rarely stems from compute limits. Instead, it traces to the human-agent supervisory boundary: ambiguous delegation, uninspected drafts flooding queues, and managers lacking frameworks to govern non-deterministic work.

This friction exposes a critical measurement gap. Established capability frameworks evaluate transformation from a macro perspective. The Microsoft Learn (2026) Agentic AI maturity model adapts the Capability Maturity Model (CMM) across five enterprise levels, from experimentation to organization-wide operations. Mainstream consulting and cloud-vendor frameworks similarly prioritize data lake governance, model clusters, and corporate compliance.

These frameworks guide procurement committees, leaving daily supervisory practices unaddressed. In automated search testing, a direct probe of Google's autocomplete API returns no completion for 'AI maturity model for managers' ([]). While industry benchmarks audit cloud infrastructure, supervisory practices remain unmeasured.

When supervisory capability is unmonitored, tooling distribution is mistaken for transformation. In the vendor-sponsored Cisco AI Readiness Index 2024, while 98% of organizations report increased urgency, only 13% characterize operations as fully prepared. Bridging this divide requires focusing on the supervisory layer macro models overlook. Read what AI readiness means for modern leaders.

What the M0–M3 Model Measures (and What It Does Not)

The M0–M3 Manager AI Maturity Model benchmarks supervisory practices across four operational levels:

• M0: Traditional (Reactive bans, unmonitored shadow adoption, or uncritical acceptance of synthetic drafts)
• M1: AI-Enabled (Personal copilot acceleration and static prompt wikis without workflow restructuring)
• M2: AI-Ready (Human-machine workflow redesign, versioned context assets, and structured quality gates)
• M3: AI-Native (Multi-agent pipeline governance, capacity decoupling, and dynamic value contracts)

The model evaluates observable behaviors across the five capability dimensions of organizational leadership: Task Delegation & Boundary (D1), Decision Quality & Anti-Hallucination (D2), Review & Quality Assurance (D3), Team Enablement & Knowledge Asset (D4), and Workforce Capacity & New KPIs (D5).

In ACM CHI research (arXiv:2603.18895), Min Hun Lee demonstrated that evaluation must shift from model accuracy benchmarks to team collaboration readiness. Failures stem primarily from miscalibrated reliance—both overuse when models err and underuse when they provide valid assistance. The M0–M3 model benchmarks how managers calibrate authority envelopes and construct recovery gates.

The model is neither a compliance checklist nor a promotion ladder guaranteeing commercial ROI. Applying Andrew Grove's production principles from High Output Management, it functions as a diagnostic radar to locate a team's "limiting step"—the constraint dictating pipeline throughput.

The Four Levels and Twelve Stages: Observable Behaviors, Blind Spots, and Transition Triggers

The matrix below details the twelve developmental stages of the manager AI maturity model. Each stage specifies an observable supervisory behavior, its associated blind spot, and the concrete transition trigger required to advance.

Level M0: Traditional Supervisory Habits

At Level M0, supervisors respond to non-deterministic technology with administrative defensiveness or inattention. Research on Bring Your Own AI adoption (arXiv:2609.05236) demonstrates that prohibition-based approaches yield residual risk profiles comparable to unmanaged baselines by driving usage into covert shadow tools. Concurrently, uncritical adoption overlooks findings documented by Ji et al. (2023, arXiv:2202.03629), where large models generate hallucinations that corrupt production systems when unverified.

Level M1: The Individual Tooling Plateau

Level M1 captures personal productivity gains that fail to improve organizational throughput. While managers save individual hours, distributing static prompt repositories cannot solve dynamic operational problems without embedded domain context. Without redesigned verification architecture, increasing individual drafting speed merely overwhelms the supervisor during manual sampling reviews, cementing management as the pipeline's limiting step.

Level M2: Workflow Integration and Verification Infrastructure

Advancing to Level M2 transforms supervisory focus from individual oversight to systemic process governance. Workflows are decomposed into specialized human-machine swimlanes anchored by version-controlled RAG repositories. This architecture operationalizes the core principle of the NIST Artificial Intelligence Risk Management Framework (AI 100-1): "Functions organize AI risk management activities at their highest level to govern, map, measure, and manage AI risks. Governance is designed to be a cross-cutting function to inform and be infused throughout the other three functions."

Level M3: Agentic Orchestration and Decoupled Capacity

Level M3 governs networks of autonomous agents executing multi-step planning, tool interaction, and programmatic reflection. True maturity at this tier demands robust runtime controls—including execution sandboxes, state rollbacks, and token circuit breakers—to prevent cascading systemic failures. By replacing input hours with verified throughput metrics, organizations decouple operational capacity from headcount while using continuous synthetic simulations to refine strategy.

False Maturity: Four Patterns That Look Advanced but Are Not

Organizations frequently mistake superficial activity for genuine operational capability. Four prevalent patterns illustrate this false maturity:

1. The Prompt Collector's Illusion (M1.2 Masquerading as M2)

A team lead compiles prompt templates on a wiki and claims AI maturity. Yet static prompts lack dynamic business context, schema constraints, and automated acceptance tests. Applying them yields verbose drafts that omit critical constraints, increasing review overhead. True maturity requires decomposing delivery pipelines and enforcing machine-readable input contracts.

2. The Firefighting Hero Trap (M1.3 Masquerading as M2.3)

A manager personally proofreads every AI-assisted draft, taking pride in catching subtle errors late into the night. However, this creates a severe pacing bottleneck. On Hacker News concerning 'The AI Productivity Gap' (Item 49152222, 2026), practitioners observed that while AI accelerates individual drafting 5x, downstream reviews, integration, and testing remain serial. Accelerating non-bottlenecks expands inventory rather than throughput. For more on this dynamic, read why AI adoption stalls at middle managers. Advancing requires erecting automated linting and source verification upstream.

3. The Multi-Agent Toy Demo (Prototype Masquerading as M3.1)

An engineering lead demos autonomous agents debating specs in a sandbox, proclaiming the team AI-native. Yet without token circuit breakers, database rollback mechanisms, and sandboxed execution, deploying such frameworks triggers infinite loops and unauthorized data changes. Industrial M3.1 implementation requires benchmarking agents against live production constraints with enforced human escalation boundaries.

4. Ungoverned Laissez-Faire Mistaken for Autonomous Agility (M0.2 Masquerading as M3.2)

A manager grants contributors unrestricted freedom to use arbitrary AI tools, labeling it agile autonomy. This abdication invites severe operational degradation. In the HBS/BCG field experiment (Dell'Acqua et al., 2026; first circulated as a 2023 working paper), knowledge workers using AI on tasks outside its capability frontier were "19% less likely to produce correct solutions compared with those without AI." Unbounded usage without verification degrades decisions. Advancing requires explicit data fences and throughput-based evaluation criteria.

Objections and Limits: Maturity Models, Labels, and Delayering

A rigorous supervisory framework must address four foundational criticisms:

Methodological Arbitrariness and the CMMI Legacy

Maturity models derived from CMMI traditions face skepticism regarding empirical validity. In Business & Information Systems Engineering, Becker, Knackstedt, and Pöppelbuß (2009) observed: "The evergrowing number of maturity models indicates a certain degree of arbitrariness concerning their development processes. Especially, this is highlighted by incomplete documentation of methodologies applied for maturity model development." A systematic review analyzing Cybersecurity Capability Maturity Models (CCMMs) within the CMMI lineage (arXiv:2408.16140) similarly found that compliance-centric models often emphasize process conformance over operational agility. We concede this critique: M0–M3 rejects linear determinism and compliance certifications. Drawing on Eliyahu Goldratt's Theory of Constraints, it functions solely as an operational diagnostic to expose workflow bottlenecks.

Feedback Psychology and the Risk of Defensive Labeling

Classifying managers into developmental stages risks triggering psychological defensiveness. In Kluger and DeNisi's 1996 meta-analysis of 131 empirical studies, as reported in a 2021 Frontiers in Psychology review, one third of the studies reviewed showed negative effects of feedback on learning, particularly when perceived as a threat to personal competence rather than a task diagnostic. Consequently, M0–M3 evaluates workflows and inspection mechanisms, never personal competence. Feedback adheres to task-focused principles, diagnosing specific interventions like automated schema validation or verifiable citation checks.

Organizational Delayering and the Contraction of Middle Management

A third critique argues that AI will flatten hierarchies, eliminating middle managers. Economic research documents an association between AI adoption and flatter corporate structures. In NBER Working Paper 31325 (Babina et al., 2023), researchers documented that "AI investments are associated with a flattening of the firms' hierarchical structure, with significant increases in the share of workers at the junior level and decreases in shares of workers in middle-management and senior roles." While administrative coordinators routing status updates face elimination, in our interpretation, surviving supervisors face expanded spans of control across human-agent teams. Orchestrating hybrid units requires higher supervisory maturity, not less.

Capital Expenditure Does Not Dictate Return on Investment

The RAND Corporation (2024) finding—that by some estimates more than 80 percent of AI projects fail—confirms software licensing does not ensure returns. The M0–M3 model does not promise that climbing stages guarantees profitability. Rather, it shows capital expenditure collapses when organizations omit supervisory architecture, allowing unvetted drafts to overwhelm downstream capacity.

A Self-Assessment: Eight Questions to Locate Your Team's Real Level

Evaluate your team's operational practices against these eight diagnostic criteria:

The Diagnostic Checklist

1. [Tool Whitelisting & Boundary] Does your team maintain an explicit whitelist of authorized AI tools with clear operational boundaries rather than blanket bans or unmonitored use? (Yes / No)
2. [Data De-Identification Protocols] Do contributors follow mandatory de-identification rules before submitting operational data to external models? (Yes / No)
3. [Source-Provenance Gate] Does your team enforce a rule prohibiting deliverables unless accompanied by traceable primary sources or tests? (Yes / No)
4. [Shared Institutional Assets] Have you progressed beyond private tool use to publish validated prompt templates in shared repositories? (Yes / No)
5. [Active Hallucination Auditing] When reviewing synthetic deliverables, do you systematically test for factual contradictions rather than approving polished formatting? (Yes / No)
6. [Human-Machine Swimlanes] Have you restructured core pipelines into explicit human-machine swimlanes rather than proofreading all outputs manually? (Yes / No)
7. [Curated Knowledge Pipelines] Does your team maintain a version-controlled RAG or context repository with documented updating procedures so contributors query verified organizational knowledge rather than unvetted public models? (Yes / No)
8. [Throughput-Based Metrics] Have you retired input metrics like hours logged in favor of net verified throughput and first-pass yield? (Yes / No)

Mitigating Social Desirability Bias

While checklists offer a reflective mirror, self-reported scores suffer from response bias. In an extensive meta-analysis on response distortion in behavioral inventories published in Frontiers in Psychology (Martínez & Salgado, 2021, PMC8511514), researchers documented that respondents systematically exhibit faking and social desirability distortion, inflating adherence to ideal behaviors. Managers working late on manual triage often over-rate their workflow governance.

Objective evaluation requires simulation-based behavioral testing. Leaders can consult our guide on how to assess manager AI skills. To benchmark your supervisory judgment against realistic trade-offs across all five dimensions, complete the free 28-question Situational Judgment Test.

What to Do Next Week: Three Moves Across the Limiting Step

Advancing from the tooling plateaus of M0 and M1 toward the integrated governance of M2 requires focused operational actions. Supervisors can execute three moves starting next week:

Move 1: Map Your First Human-Machine Workflow Swimlane

Select one recurring deliverable causing review bottlenecks (such as technical design briefs). Map it across four distinct swimlanes: AI drafting, contributor triage, automated verification checks, and senior boundary approval, preventing unvetted drafts from flooding managerial queues.

Move 2: Erect a Minimal Verification Gate with Source-Provenance Chains

Enforce an immediate standard: no proposal or commit may be submitted without an auditable provenance chain linking claims to primary sources or test cases. Requiring explicit evidence guards against unverified outputs, addressing vulnerabilities highlighted by Dell'Acqua et al. (2026; first circulated as a 2023 working paper), where consultants relying on AI outside its capability frontier were 19% less likely to produce correct solutions than consultants working without AI.

Move 3: Refactor One Non-Hours-Based Performance Metric

Eliminate activity metrics that reward generation volume, such as lines of code or draft counts. Replace them with operational throughput indicators: first-pass inspection yield (percentage passing quality gates without remediation) or net verified cycle velocity.

Upgrading managerial AI maturity is not about prompt tricks or added software licenses. It is the operational discipline of setting delegation boundaries, erecting verification gates, and protecting throughput. To see where this maturity model sits relative to external governance, literacy, and assessment frameworks, read our comparison of nine AI competency frameworks. Explore the wider research series in the AI-Ready Leader Insights hub, or to benchmark your team's readiness and locate your limiting step, take the free 28-question Situational Judgment Test.