Governance, risk & compliance instrument

AI Readiness Assessment

Go through your organization one control at a time, across ten areas. Mark each as Met, Partial, or Gap, and the results roll up into a maturity score for each area and an overall picture. This is the in-depth version of the quick self-checks, scored the way a real engagement would. New to this? The five-minute AI Maturity Assessment is the gentler place to start. Nothing you enter leaves your browser.

See the control crosswalk for how these domains map across NIST AI RMF, ISO 42001, SOC 2, ISO 27001, NIST 800-171 / CMMC, and privacy.

1. Governance, Accountability & Policy

What good looks like: Someone clearly owns AI, a group oversees it, there is an approved policy on what is allowed, and everyone knows who gets to decide what.

Not scored
1.1A named owner is accountable for AI across the organization.How to assess: A named individual or role accountable for AI, recorded in an org chart or governance charter.
1.2A chartered oversight body governs AI decisions.How to assess: A committee charter with membership, mandate, meeting cadence, and minutes.
1.3An approved AI policy and acceptable-use standard are in force and current.How to assess: An executive-approved, dated, version-controlled AI policy and acceptable-use standard.
1.4Decision rights for AI are defined and assigned.How to assess: A RACI (or equivalent) covering who approves, builds, operates, and reviews AI across its lifecycle.

2. AI System & Agent Inventory + Risk Classification

What good looks like: You keep a live list of every AI system and agent, including who owns it, what data and tools it can reach, and how risky it is.

Not scored
2.1A complete inventory of AI systems and agents is maintained and kept current.How to assess: A live register and a process that catches new or changed AI within a defined window.
2.2Each inventory entry records its business owner, data access, and tool access.How to assess: Populated owner, data-scope, and tool/API-scope fields for every entry.
2.3A defined risk-classification scheme is applied to every AI system and agent.How to assess: A documented risk-tiering method and a tier assigned to each entry.
2.4Shadow AI is detected and brought under governance.How to assess: Discovery scans across network, SaaS, and identity, with a remediation path for ungoverned AI.

3. Data Governance & Privacy

What good looks like: AI uses only the data it needs, that data is kept accurate and labeled by how sensitive it is, and personal information is handled with privacy built in.

Not scored
3.1Data used by AI is classified and handled by sensitivity.How to assess: Sensitivity labels or classification applied to the data AI trains on and reads at inference.
3.2Data minimization is enforced for AI systems.How to assess: A record showing each data element in scope is necessary; minimization applied to prompts, context, and training sets.
3.3Data quality is assessed before data feeds an AI system.How to assess: Accuracy, completeness, and currency checks on the datasets used by AI.
3.4Privacy-by-design governs PII handling and data sharing for AI.How to assess: DPIA or PIA records, PII-handling rules, and controlled data-sharing agreements for AI data flows.

4. Identity & Access Governance, incl. agentsFramework Domain 1

What good looks like: Every AI agent signs in as its own tracked identity, gets only the access it needs and only when it needs it, and its activity is logged separately from people.

Not scored
4.1Every agent runs under a unique non-human identity (no shared or human credentials).How to assess: An identity register mapping each agent to its own service principal / workload identity.
4.2Agent access is least-privilege and just-in-time scoped to the task.How to assess: Per-agent permission scopes; JIT elevation records; no standing broad grants.
4.3Privilege drift is prevented; agent entitlements are reviewed on a defined cadence.How to assess: Periodic access-review records for non-human identities; deprovisioning on agent retirement.
4.4Agent identity and access events are logged and auditable separately from human activity.How to assess: A distinct audit stream for agent auth and access events.

5. Human Oversight & Decision RightsFramework Domain 2

What good looks like: For each important decision you have decided in advance whether AI can act on its own, only advise, or must wait for a person, and you track how often people overrule it.

Not scored
5.1Material decision points are identified before deployment.How to assess: A workflow map that marks every point where the consequence of an error is material.
5.2Each material decision point has an assigned oversight tier.How to assess: A documented tier (autonomous, advisory, or required) for each decision point.
5.3Oversight interfaces give the reviewer what they need to make a real decision.How to assess: Review screens that surface decision context, options, and rationale, not just a log of what the agent did.
5.4Override rates are tracked and acted on.How to assess: Override metrics per decision point and a review when rates run high or sit at zero.

6. AI / Agent Security & Tool-Use, MCPFramework Domain 3

What good looks like: Each agent can use only the tools you have approved and nothing else, every tool call is logged as a security event, its outputs are checked, and you test it against attempts to trick it (prompt injection).

Not scored
6.1Each agent runs against a deny-by-default tool and MCP allowlist.How to assess: A per-agent allowlist of approved MCP servers and tools, with default-deny enforced.
6.2Tool invocations are logged as security events with full parameters.How to assess: Tool-call logs, including parameter values, routed to the security log pipeline rather than operational logs alone.
6.3Tool outputs are validated before they re-enter the agent context.How to assess: Schema validation, sanitization, and anomaly flagging applied to tool results.
6.4The tool interface is red-teamed for prompt injection before deployment.How to assess: Prompt-injection test cases run against tool responses, with findings remediated.

7. Audit Trail, Logging & ExplainabilityFramework Domain 4

What good looks like: For decisions that matter you can see the AI's full reasoning, not just what went in and out; the record cannot be altered; and a non-technical person can follow it.

Not scored
7.1Consequential AI decisions capture a full reasoning trace, not just input and output.How to assess: Stored traces that include the decision path, tools called, and data referenced.
7.2Audit logs are immutable and tamper-evident.How to assess: Write-once / append-only storage; integrity verification.
7.3Logs are retained for the period the applicable regime requires.How to assess: A retention policy mapped to the regime in scope.
7.4A non-technical compliance officer can reconstruct and explain a given decision from the trace.How to assess: A walk-through reconstructing one real decision end to end.

8. Monitoring: Bias, Fairness & PerformanceFramework Domain 6

What good looks like: You decide up front what fair looks like, then watch for unfair patterns and slipping accuracy over time, including bias picked up from the tools the AI relies on.

Not scored
8.1Fairness metrics are defined before deployment.How to assess: Documented fairness metrics for the decision domain and a statement of what an equitable outcome looks like.
8.2Disparate impact and performance drift are monitored on a defined cadence.How to assess: Scheduled monitoring with results retained over time.
8.3Bias inherited from called tools is included in the analysis.How to assess: A bias assessment that covers the scoring, search, or external APIs the agent relies on.
8.4Defined thresholds trigger fairness review and intervention.How to assess: Review thresholds and a record of interventions taken when they are crossed.

9. Incident Response & Resilience, autonomous failuresFramework Domain 5

What good looks like: Every agent has a clear off switch (who flips it, when, and what happens to work already in progress), your responders are trained on how the AI is built, you rehearse failures, and you review which control broke.

Not scored
9.1Every agent has a documented kill switch before go-live.How to assess: A per-agent kill switch defining who can activate it, under what conditions, and how in-flight tasks are handled.
9.2Incident response personnel are trained on agent architecture.How to assess: IR training that covers tool-call chains and MCP connections, with named responders.
9.3Agentic failure scenarios are rehearsed before deployment.How to assess: Tabletop exercise records that simulate autonomous failure.
9.4Post-incident reviews examine the governance control that failed.How to assess: Post-mortems that name the governance failure, not only the technical symptom.

10. Regulatory Alignment & Third-Party / Supply ChainFramework Domain 7

What good looks like: Each AI use is matched to the rules that apply to it (such as the EU AI Act, HIPAA, or federal contracting rules), your AI vendors are risk-checked, and clearing those rules is required before you go live.

Not scored
10.1Each deployment is mapped to the regulatory regimes that apply.How to assess: A per-deployment mapping to EU AI Act tier, HIPAA, CMMC/CUI, and federal AI transparency where applicable.
10.2In-scope regimes drive their specific obligations into the design.How to assess: For each applicable regime the required controls are in place; for high-risk EU AI Act, that means human oversight, transparency documentation, and conformity assessment.
10.3Model and tool vendors are risk-reviewed.How to assess: Vendor and AI supply-chain risk reviews on file for each model and tool in use.
10.4Compliance is a documented prerequisite for deployment.How to assess: A release gate that records regulatory alignment before go-live, not after.

v1 draft: controls are at the family / criteria level and validated per engagement. The Met / Partial / Gap point values and the maturity thresholds are a sensible default, not a published standard. Editions referenced: NIST AI RMF 1.0 (2023), ISO/IEC 42001:2023, ISO/IEC 27001:2022, NIST SP 800-171 (r2/r3) and CMMC 2.0, ISO/IEC 27701, NIST Privacy Framework 1.0.