Scoring Your Agentic AI Readiness: The A7 Assessment

You have read the framework. This is the assessment.

From Framework to Score

Article 1 of this series established the problem: 40% of agentic AI projects will be canceled, most organizations are deploying at autonomy levels their readiness cannot support, and only about 130 of thousands of agentic AI vendors are genuine. The A7 Framework introduced seven dimensions, five autonomy levels, the Premature Autonomy concept, and the dimensional floor rule. That article gave you the map.

This article gives you the instrument. A structured scoring methodology that turns "are we ready for AI agents?" into a number between 7 and 35 — mapped directly to the autonomy level your organization can safely deploy. Full rubrics for all seven dimensions. The Agent Washing Detector. Industry benchmarks. And the Agentic Readiness Sprint — a phased plan for moving from your current autonomy level to the next.

By the end of this article, you will have everything you need to score your organization, identify your weakest dimensions, understand which autonomy level your readiness actually supports, and build a prioritized improvement plan. The A7 Readiness Worksheet packages this into a practical tool you can run with your leadership team in a single working session.

How the A7 Score Works

The A7 scoring system is built for clarity and utility. Seven dimensions, each scored 1-5, produce a total Readiness Score between 7 and 35. The score maps to an autonomy level — and the mapping includes a critical safety mechanism that prevents the most dangerous misinterpretation.

A7 Readiness Score = A1 + A2 + A3 + A4 + A5 + A6 + A7

The minimum score is 7 (all dimensions at Level 1 — the organization has virtually no agentic AI readiness). The maximum is 35 (all dimensions at Level 5 — the organization operates at the frontier of agent-native capability). Most organizations today score between 10 and 22. The score is not a grade. It is a diagnostic that tells you where you stand and what you can safely deploy.

1. 7-14 → L0 (Not Ready): Traditional AI only. No autonomous agents. Deploy predictive models and dashboards, but nothing that takes autonomous action.
2. 15-21 → L1 (Copilot Ready): AI assists, human decides. Copilots and suggestion-based tools. A human must approve every action.
3. 22-28 → L2 (Supervised Agent Ready): Agents execute within boundaries. Humans supervise. The agent acts; the human watches and can intervene.
4. 29-33 → L3 (Autonomous Ready): Agents operate independently with programmatic guardrails. Human review on exception. Requires mature governance, robust security, and reliable oversight.
5. 34-35 → L4 (Full Autonomy Ready): Self-directed agents. Rare and aspirational. Appropriate only for specific, well-bounded use cases in the most mature organizations.

The assessment process takes 2-3 hours with your leadership team. For each dimension, you review the scoring rubric, identify the description that most closely matches your organization's current state, assign a score with brief justification, and note the single highest-impact improvement action. Then you aggregate, map to autonomy level, and check the floor rule. The rest of this article provides the rubrics that make this possible.

Scoring the Foundation: A1-A3

The first three dimensions form the foundation of agentic readiness. Data Architecture (A1) determines whether agents can access the information they need. Technical Infrastructure (A2) determines whether agents can operate reliably. Governance (A3) determines whether agents operate within appropriate boundaries. A weakness in any one of these three compromises everything built on top.

A1 — Data Architecture (1-5)

MIT Technology Review confirms that most AI delays are data architecture gaps, not model shortcomings. More than two-thirds of technology executives identified data issues as the primary risk factor for failing to achieve AI goals. The A1 dimension measures whether your data infrastructure can serve agents — not just dashboards.

1. Level 1 — Siloed: Data exists in departmental silos. Batch processing only. No API access to operational data. AI training data is manually extracted. No data catalog or lineage. Example: Inventory in SAP, customers in Salesforce, analytics in Google Analytics — an agent must wait for nightly ETL jobs.
2. Level 2 — Consolidated: Basic data warehouse or lake exists. Some API endpoints. Data quality is inconsistent. Integration requires manual engineering for each new use case. Freshest data is hours old. Example: Snowflake warehouse, but onboarding a new data source takes 6-8 weeks.
3. Level 3 — Integrated: Integrated lakehouse with broad API access. Data catalog operational. Core data quality metrics tracked. Near-real-time data for priority systems. Semantic layer emerging. Example: Databricks lakehouse with 80% catalog coverage, 15-minute refresh, semantic layer defining core entities.
4. Level 4 — Adaptive: Real-time streaming infrastructure (Kafka, Flink). Comprehensive semantic layer. Data mesh principles — domain teams own and publish data products. Agents can discover and access data through self-service interfaces. Example: Real-time data mesh where each domain publishes data products with SLAs.
5. Level 5 — Agent-Optimized: Agent-native data architecture: real-time streaming, comprehensive semantic layer, agent-specific interfaces, automated data quality enforcement, context-rich metadata, and feedback loops where agent actions generate data that improves future performance. Example: Every data product includes agent-readable schemas, confidence scores, and freshness guarantees.

A2 — Technical Infrastructure (1-5)

Traditional cloud infrastructure was designed for request-response patterns. Agents operate in continuous execution loops — observing, planning, acting, evaluating — often across multiple coordinated agents. Gartner's 40% cancellation prediction reflects, in significant part, infrastructure gaps that become apparent only after deployment.

1. Level 1 — Basic: Single cloud environment. Manual deployment. No container orchestration. No agent-specific infrastructure. Monitoring is application-level only. Example: AI workloads on a single AWS account, deployed manually via notebooks.
2. Level 2 — Standardized: Containerized workloads (Docker/Kubernetes). Basic CI/CD. Cloud-native services. Agents run as standalone processes. No undo capability. Example: ML on Kubernetes with Jenkins, but no orchestration for multi-step tasks.
3. Level 3 — Orchestrated: Container orchestration with auto-scaling. Agent framework deployed (LangGraph, CrewAI). Basic state management. Monitoring covers agent execution. Manual rollback procedures exist. Example: LangGraph agents with state persistence, monitoring of task completion and error rates.
4. Level 4 — Resilient: Multi-region or multi-cloud. Sophisticated agent orchestration with tool-use management. Reversible operations for critical actions. Automated failover. End-to-end execution traces. Performance SLAs defined and monitored. Example: Agents across AWS and Azure with transaction rollback and real-time SLA monitoring.
5. Level 5 — Agent-Native: Multi-cloud agent mesh with dynamic routing. Atomic operations with automatic rollback. Auto-scaling responsive to agent workload patterns. Agent lifecycle management (versioning, A/B testing, deprecation). Self-healing infrastructure. Example: Agent mesh with canary releases, automatic rollback on performance degradation, self-healing on failure.

A3 — Governance Framework (1-5)

According to industry research cited by the World Economic Forum, only 21% of leaders currently have a mature governance model for autonomous agents. And Gartner projects $5 billion in compliance spending by 2027. The A3 dimension maps directly to the Minimum Viable Governance framework: Level 3 on A3 = MVG implementation.

1. Level 1 — Absent: No AI governance structure. No policies covering AI decision-making. No inventory. No designated ownership. Agents deployed by whoever has access. Example: Engineers experiment with agents; CEO learns about deployments from customer complaints.
2. Level 2 — Informal: Basic AI policies exist but are not enforced. Governance is one person's part-time responsibility. No distinction between traditional AI and agent governance. Example: "AI Ethics Policy" drafted during an audit but addressing model bias, not agent autonomy.
3. Level 3 — Structured (MVG): Complete AI/agent system inventory, designated governance owners, risk-tiered classification, deployment gates, regular governance reviews. Policies address agent decision authority. Example: Every AI system inventoried with named owner, risk tier, and decision scope. Monthly governance reviews.
4. Level 4 — Comprehensive: Cross-functional governance council. Governance integrated into agent development lifecycle. Agent decision authority framework with explicit boundaries per use case. External audit capability. Example: Governance Board with decision authority matrices — which decisions agents make autonomously vs. with approval.
5. Level 5 — Embedded: Governance-as-code: agent policies enforced programmatically. Continuous monitoring of agent decisions against policy. Real-time dashboards. Board-level reporting. Governance enables faster deployment through automated guardrails. Example: Decision boundaries enforced by infrastructure, not manual review. Low-risk deployments clear governance automatically.

Scoring the Human Layer: A4-A7

The next four dimensions form the human layer — the organizational capabilities that determine whether technology investments produce results or produce expensive failures. An organization can score highly on A1-A3 (strong data, robust infrastructure, mature governance) and still fail at agentic AI if human oversight is absent, the culture resists, security is traditional, or autonomy is miscalibrated.

A4 — Human Oversight Protocols (1-5)

Fortune's reporting captures the central question: "What happens when an agent goes rogue?" The answer for most organizations is troubling. AI agents do not go rogue from malicious intent — companies give them too much freedom. Permission fatigue sets in and agents gradually accumulate broad privileges.

1. Level 1 — None: No human oversight mechanisms. No approval workflows for automated decisions. No escalation paths. No kill switches. No monitoring of agent actions. Example: Customer service agent operates autonomously issuing refunds and modifying accounts with no human review.
2. Level 2 — Minimal: Basic logging of agent actions exists but is not reviewed regularly. Manual intervention requires engineering effort. Escalation is informal. Example: Agent actions logged to a database, but no one reviews unless a customer complains. Stopping requires an engineer to modify config.
3. Level 3 — Defined: Approval gates for critical actions. Action logging with regular review cadence. Defined escalation paths for anomalies. Pause capability exists. Example: Procurement agent approves orders under $5,000 autonomously; above that requires human approval. Weekly log review. Dashboard pause button.
4. Level 4 — Graduated: Graduated oversight model: autonomy per agent matched to risk tier. Real-time monitoring dashboards. One-click kill switches with graceful degradation. Automated anomaly detection triggers human review. Example: Three oversight tiers — low-risk autonomous, medium-risk with review, high-risk with real-time monitoring and automatic halt.
5. Level 5 — Adaptive: Oversight intensity adjusts dynamically based on agent confidence, task novelty, and environmental signals. Kill switches with instant graceful degradation. Oversight telemetry feeds governance reporting. Regular simulation exercises. Example: Monitoring intensity scales with risk signals — routine actions logged, unusual actions trigger review, high-risk anomalies trigger automatic pause with human takeover.

A5 — Organizational Readiness (1-5)

Google Cloud's 2025 review states it directly: "Success depends on preparing people to trust and use the technology, not just the power of the code." The World Economic Forum identifies three persistent blockers — infrastructure, trust, and data — but behind all three lies organizational readiness.

1. Level 1 — Resistant: No AI culture. Active or passive resistance to AI adoption. No AI skills development. No executive sponsorship. AI perceived as a threat. Example: Partners view AI as a cost threat to billable hours. IT experiments with copilots in secret.
2. Level 2 — Aware: Executive awareness exists but has not translated to strategy or investment. Limited AI skills. Change management is ad hoc. Example: CEO mentions AI in all-hands but allocated no budget. Two data scientists report to IT with no business integration.
3. Level 3 — Engaged: AI-aware leadership with budget. Skills development programs operational. Dedicated AI team. Change management playbook. Executive sponsor identified. Cross-functional collaboration. Example: VP of AI with team of 12, skills program reaching 200 employees, change management playbook for each deployment.
4. Level 4 — Committed: Executive team aligned on agentic AI strategy. Dedicated agent operations teams. Organization-wide AI literacy. Incentive structures support AI adoption. Culture embraces human-agent collaboration. Example: C-suite completed agent literacy training. "Agent Operations" team manages deployed agents. Performance reviews include AI adoption metrics.
5. Level 5 — Agent-Native: Organization designs processes assuming agents are participants. Every function has agent integration expertise. Board-level agent strategy. Continuous learning embedded. Contributes to industry knowledge. Example: Every new process includes "agent role" specification alongside human roles. Department heads have agent deployment authority within guardrails.

A6 — Security & Safety (1-5)

Microsoft's Security Blog is explicit: "Autonomous agents aren't a minor extension of existing identity or application governance — they're a new workload." Their 2026 analysis identifies specific threat vectors: prompt injection, tool-use boundary violations, credential exposure, task drift, and Cross Prompt Injection Attacks.

1. Level 1 — Standard AppSec: Traditional application security only. No agent-specific measures. No awareness of agent attack vectors. Security team has not assessed agent risk. Example: Mature AppSec program — SAST, DAST, pen testing — but agent prototype accesses production DB with shared service account.
2. Level 2 — Aware: Security team aware of agent risks. Basic input validation. Agent credentials managed (not hardcoded). No systematic adversarial testing. Agent security on roadmap. Example: Agent credentials in secrets manager. Input validation filters obvious prompt injection. No adversarial testing conducted.
3. Level 3 — Defensive: Prompt injection defenses deployed. Agent tool-use boundaries defined and enforced. Credential isolation per agent. Output validation for customer-facing actions. Regular security review. Example: Customer service agent can query balances but cannot initiate transfers. Isolated credentials with least-privilege. Quarterly security reviews.
4. Level 4 — Proactive: Agent threat model maintained. Regular adversarial testing (red-teaming). Sandboxed execution environments. Behavioral monitoring for anomalous actions. Agent security integrated into development lifecycle. Example: Dedicated red team conducts monthly adversarial testing. New capabilities tested in sandbox before production. Behavioral monitoring flags anomalous patterns.
5. Level 5 — Comprehensive: Full agent security program: scope control, adversarial testing, sandboxing, credential isolation, behavioral monitoring, supply-chain security for tools and plugins, and security automation. Security enables deployment by providing verified guardrails. Example: Every agent in sandbox with strict scope control. Ephemeral credentials per task. Continuous adversarial testing pipeline. ML-based anomaly detection.

A7 — Autonomy Calibration (1-5)

The meta-dimension. Gartner estimates only about 130 of thousands of agentic AI vendors are genuine. An organization that cannot distinguish real agents from rebranded products will systematically miscalibrate its autonomy level — believing it deploys L2 agents when it has L1 copilots with better marketing.

1. Level 1 — Uncalibrated: Cannot distinguish between AI assistants, copilots, and autonomous agents. No taxonomy for autonomy levels. Vendor claims accepted at face value. Example: Procurement selects "AI agent" vendor based on marketing — deploys what are actually chatbots with better UX.
2. Level 2 — Aware: Basic understanding that autonomy levels exist. Can distinguish chatbots from more sophisticated AI, but taxonomy is informal and inconsistent. Some vendor claims questioned. Example: CTO pushes back on some vendor claims, but no formal taxonomy. Different departments use "agent" to mean different things.
3. Level 3 — Defined: Clear autonomy level taxonomy adopted organization-wide. Vendor evaluation includes technical assessment of actual capabilities. Organization can accurately classify current deployments. Example: Four-level taxonomy adopted. Every AI system classified. Vendor evaluations include technical deep-dives: "Show us the planning loop, tool-use, state management."
4. Level 4 — Calibrated: Autonomy level matched to readiness per use case. Formal readiness assessment conducted before each deployment. Regular reassessment. Agent washing detector operational. Example: A7 assessments before each deployment. Customer service approved for L2; trading deferred because A6 insufficient for L3. Three vendors rejected as agent-washed.
5. Level 5 — Optimized: Right autonomy level per use case, continuously reassessed, with clear upgrade paths. Organization can articulate exactly what improvements would unlock the next level. Contributes to industry standards. Example: Each deployment has documented autonomy level, readiness justification, and upgrade path. Quarterly reassessments. Participates in industry working groups.

Your Score, Your Autonomy Level

Sum the seven dimension scores to produce your A7 Readiness Score (7-35). Map the score to the autonomy level. Then apply the dimensional floor rule — the most important safety mechanism in the scoring system.

The floor rule is simple and non-negotiable: a single low-scoring dimension can block an autonomy level regardless of the total score.

1. L1 (Copilot Ready): No dimension below 1. Any score qualifies.
2. L2 (Supervised Agent Ready): No dimension below 2. A single Level-1 dimension blocks L2.
3. L3 (Autonomous Ready): No dimension below 3. Requires at least Defined maturity across all dimensions.
4. L4 (Full Autonomy Ready): No dimension below 4. Requires Advanced maturity across all seven dimensions.

Why the floor rule? Because aggregate scores mask critical weaknesses. An organization scoring 28 overall — technically in the L2 range — but with A4 (Human Oversight) at Level 1 and A6 (Security) at Level 1 is not safely L2 Ready. It has no mechanism to catch an agent that drifts from its intended purpose and no agent-specific security to prevent exploitation. The aggregate number lies. The floor rule tells the truth.

If your floor-adjusted autonomy level is lower than your aggregate-indicated level, the gap between the two numbers is your improvement priority list. A score of 30 with A3 at Level 2 means you are L2 until governance matures — and governance is the single investment that would unlock L3 deployment across the organization.

The Agent Washing Detector

Gartner estimates that roughly 97% of agentic AI vendors are not genuine — they are chatbots, RPA bots, and AI assistants rebranded as agents without the underlying capabilities. The A7 Framework includes a five-question diagnostic for detecting agent washing in vendor products and internal deployments. A product that fails three or more of these questions is likely agent-washed.

1. 1. Planning: Does the system decompose goals into sub-tasks, or does it follow a fixed script? A fixed script is not an agent — it is a workflow automation tool with a marketing upgrade.
2. 2. Tool Use: Does the system dynamically select and invoke external tools, or does it call a predetermined API? A predetermined API call is not agentic — it is an integration.
3. 3. Memory: Does the system maintain state across interactions and learn from previous actions, or does each interaction start fresh? Stateless interaction is a chatbot, regardless of the label.
4. 4. Autonomy: Does the system take actions without human approval for each step, or does every action require a human click? If every action requires approval, it is a copilot — a useful copilot, but not an agent.
5. 5. Recovery: When the system encounters an unexpected situation, does it adapt its plan, or does it fail and escalate? A system that always escalates on the unexpected is a rule-based system, not an autonomous agent.

Run this diagnostic on every vendor product you evaluate and every internal system labeled "agent." The results will likely surprise you. Most organizations discover that what they call "agents" are actually copilots or chatbots — which is not a failure. It is a recalibration. Knowing you are at L1 with L1 tools is safer than believing you are at L3 with L1 tools. The Agent Washing Detector brings honesty to a market that has very little.

Industry Benchmarks

Where does your score sit relative to your industry? These benchmarks are directional — based on available evidence as of early 2026 — not definitive. Every organization's score will be unique. The purpose is orientation: knowing whether you are ahead of, behind, or in line with your peers prevents the two most dangerous assumptions — "we must be behind" (leading to panic-driven deployment) and "we must be fine" (leading to complacent inaction).

1. Large Tech Companies (Score 22-28, L2): Strong A1 and A2. Weak spots are A3 (Governance) and A4 (Oversight) — technical dimensions outpace organizational ones. Move fast, govern later.
2. Financial Services (Score 18-24, L1-L2): Strong A6 (Security) from regulatory culture. Weaker A1 (Data Architecture) due to legacy systems. The gap between security maturity and data maturity is the defining challenge.
3. Healthcare (Score 14-20, L0-L1): Strong A3 (Governance) from regulatory culture. Weak A2 (Infrastructure) due to legacy clinical systems. Governance leads technology — the inverse of tech companies.
4. Manufacturing (Score 12-18, L0-L1): Weak across most dimensions. A5 (Organizational Readiness) is the primary blocker — the culture and skills gap is wider than the technology gap.
5. Mid-Market Enterprises (Score 10-16, L0-L1): Resource constraints limit all dimensions. Should focus on L1 readiness — reliable copilot deployment — before attempting agent deployment.
6. Startups / Digital Natives (Score 16-24, L1-L2): Strong A1, A2 (modern infrastructure). Weak A3, A4, A5 — move fast, governance and oversight lag. The opposite pattern from regulated industries.

Note the pattern: no industry segment is consistently above L2. L3 is rare. L4 is aspirational. If your competitors claim to be at L3, apply the Agent Washing Detector to their claims. The evidence says they are probably at L1-L2, and what they call "autonomous agents" are probably supervised copilots.

The Agentic Readiness Sprint

A score without a plan is a diagnosis without treatment. The Agentic Readiness Sprint converts your A7 assessment into a phased improvement roadmap — designed to close your most critical gaps first, using the same logic that personal finance advisors use when helping clients escape compounding debt: pay down the highest-interest balance first.

Phase 1: Assess (Weeks 1-2)

Run the full A7 assessment with your cross-functional leadership team. Score all seven dimensions. Calculate the total score and map to autonomy level. Apply the floor rule. Identify every dimension that falls below the minimum for your target autonomy level. Rank gaps by impact — which single improvement would unlock the most value?

Deliverables: A7 scorecard with dimension-level scores, floor-adjusted autonomy level, gap analysis with prioritized improvement targets, and ownership assignment for each gap.

Phase 2: Remediate the Floor (Weeks 3-12)

Focus exclusively on dimensions that violate the floor rule for your target autonomy level. If you score 24 overall (L2 range) but A4 is at Level 1, all improvement energy goes to A4 until it reaches Level 2. Do not spread investment across all dimensions — concentrate on the constraint. The floor violation is the bottleneck; everything else is optimization.

Common Phase 2 targets and quick wins: A3 at Level 1 → Level 2: Draft basic AI policies, assign governance ownership, begin an AI system inventory. A4 at Level 1 → Level 2: Implement basic logging of agent actions, establish an informal escalation path, add manual intervention capability. A6 at Level 1 → Level 2: Brief security team on agent risks, move agent credentials to a secrets manager, add basic input validation.

Phase 3: Level Up (Months 4-12)

With floor violations resolved, focus on raising the weakest dimensions toward your target autonomy level. The goal is balanced improvement — increasing the lowest scores first because they represent the highest-leverage opportunities. A dimension at Level 2 improved to Level 3 has more impact than a dimension at Level 4 improved to Level 5. The Agentic Readiness Sprint follows the same compounding logic as the Liability Ledger: small improvements to the weakest dimensions compound faster than large improvements to the strongest ones.

From Assessment to Action

This article has given you the complete scoring methodology — rubrics for all seven dimensions, the floor rule, the Agent Washing Detector, industry benchmarks, and a phased sprint for improvement. The A7 Readiness Worksheet packages all of this into a single working document designed for a leadership team session: scoring sheets for each dimension, the autonomy level mapping, the floor rule calculator, and the sprint planner.

The A7 Framework does not exist in isolation. It connects to the broader AskAjay ecosystem. Minimum Viable Governance provides the governance foundation — an organization that implements MVG achieves Level 3 on A3. The Trust Premium quantifies the business value that higher A7 scores create. The Liability Ledger captures the compounding cost when organizations deploy without adequate readiness. And the PRIME Framework governs responsible development of the agents themselves.

Start with the question this framework was built to answer: What autonomy level can your organization actually support? If you cannot answer that question with a number, you have found your starting point. Download the A7 Worksheet, assemble your leadership team, and run the assessment. The organizations that will scale agents successfully are the ones that know their number — and deploy accordingly.