ascendraAI READINESS & ADVISORYBook Consultation
← Back to AI ClarityAscendra research · Proposed methodology v0.3

AI Clarity.
From evidence to action.

A risk-based framework for understanding AI, assessing applicable controls and planning how to enforce them. For AI you build, AI you consume and the systems connecting both.

The Ascendra AI Clarity Framework

Five dimensions.
A clearer view of AI.

Understand the capabilities behind responsible AI adoption. Explore each dimension to see the question it answers and the evidence behind it.

AI CLARITYFIVE DIMENSIONS
Five independent perspectives.
Conceptual visual — no assessment scores shown.
Visibility

Do we know what AI exists and what it does?

Establish the AI footprint: applications, embedded AI, models, agents and connections. Reconcile what is approved with what is observed, and make remaining blind spots explicit.

EVIDENCE TO EXAMINE

Inventory records, named owners, discovery coverage, data flows and agent or tool activity.

Explore the assessment offer
The right controls for the actual risk

Every AI use is different.
Its controls should reflect that.

AI you consume

Purchased tools and embedded features. Assess intended use, information shared, integrations, reliance on outputs and the split between supplier and customer responsibilities.

AI you build

Applications, retrieval systems and agents. Extend the assessment to data pipelines, prompts, models, tools, deployment and ongoing operation.

One connected assessment

A purchased model connected to internal data and tools can involve both. Assess what the system does and what your organization controls.

Illustrative use case

Drafting public marketing content

A person reviews the output before use. No sensitive information or connected action is involved.

Routine approval under delegated authority, with a clear route to report problems.

Proportionate control expectations

Establish the baseline.

  • Approved service
  • Named owner
  • Acceptable-use rules
  • Human review
A lifecycle for AI built and consumed

From understanding AI.
To knowing its controls work.

The three pillars establish the baseline, evidence and roadmap. This six-step lifecycle connects them to implementation and ongoing assurance.

  1. 01

    Discover and register

    Identify AI uses, owners, data and connections. Record visibility gaps.

  2. 02

    Assess and classify risk

    Evaluate consequences and authority. Assign a justified use-case risk tier.

  3. 03

    Select proportionate controls

    Define required safeguards, shared responsibilities and justified exclusions.

  4. 04

    Implement and enforce

    Apply agreed controls across identities, gateways, pipelines and operations.

  5. 05

    Validate and decide

    Test effectiveness. Remediate, restrict, accept residual risk or stop.

  6. 06

    Monitor and improve

    Track activity, performance and changes. Reassess and verify remediation.

From policy to enforcement

Your AI
control plane.

01Ownership and inventory

Know which agents, models and connectors are approved, what they do and who owns them.

How we verify it: Reconcile observed activity to approved records and resolve unowned assets.

02Identity and permissions

Identify each agent and constrain its authority, including read versus write access and delegation on behalf of a user.

How we verify it: Attempt prohibited actions and verify that authorization and revocation work.

03Data and tool protection

Apply sensitive-data restrictions, approved connections, DLP and checks at tool execution boundaries.

How we verify it: Test attempted data disclosure and unauthorized tool invocation.

04Testing and release

Define quality and security acceptance criteria. Control changes to models, prompts, sources and tools.

How we verify it: Review release evidence and revalidate representative scenarios after material change.

05Human oversight

Make required approvals enforceable. Name escalation owners and people authorized to intervene.

How we verify it: Test approval bypass, stale approvals and actions changed after approval.

06Monitoring and production support

Connect relevant inputs, versions, sources, tool calls, outputs and approvals so incidents and questioned decisions can be investigated.

How we verify it: Exercise alerts, suspension, rollback and recovery. Protect traces with appropriate redaction, access and retention.

What the framework examines

25 criteria.
A traceable assessment.

Five proposed criteria in each dimension. Define applicability, evidence and test expectations before scoring. A detailed reference crosswalk and criterion-specific anchors require pilot development.

Visibility

  • V1 · AI inventory
  • V2 · Discovery coverage
  • V3 · Attribution
  • V4 · Data and action visibility
  • V5 · Change reconciliation

Governance

  • G1 · Accountability
  • G2 · Risk decisions
  • G3 · Human oversight
  • G4 · Supplier and lifecycle oversight
  • G5 · Management assurance

Security

  • S1 · Identity and authorization
  • S2 · Data and model protection
  • S3 · Secure lifecycle
  • S4 · Runtime safeguards
  • S5 · Security response

Data Stewardship

  • D1 · Ownership and provenance
  • D2 · Quality and suitability
  • D3 · Permitted use and privacy
  • D4 · Retention and deletion
  • D5 · Knowledge maintenance

Operational Assurance

  • O1 · Task performance
  • O2 · Safety and impact testing
  • O3 · Behavior monitoring
  • O4 · Intervention and fallback
  • O5 · Operational recovery

One outcome, one primary criterion

Shared evidence can support distinct outcomes. Avoid duplicate credit for the same fact. Governance defines oversight; operational assurance tests whether intervention works.

Agent and MCP assessments feed the same dimensions when those architectures are present.

Evidence-based scoring

Demonstrated capability.
With its limits in view.

0Absent
1Ad hoc
2Defined
3Implemented
4Tested effective
5Sustained

Rate each dimension independently

Award the highest cumulative level supported by all required anchors. A criterion with several required outcomes takes the lowest demonstrated level. The dimension is the equal-weight mean of applicable criteria, shown to one decimal.

Equal weighting is a proposed pilot convention. Do not calculate an overall average across dimensions or interpret maturity as the probability of loss.

Unknown is not zero

Insufficient evidence means Unrated. Publish a dimension score only when every applicable criterion in it is rateable. Document and independently review N/A exclusions.

Report scope, assessment period, method version, evidence completeness, operational coverage and significant findings alongside any rating.

Keep risk decisions separate

Assess inherent risk before controls, then residual risk after considering demonstrated effectiveness. Severe consequences trigger minimum requirements; they must not disappear in an average.

Critical exposures remain visible as Priority action. Serious unknowns require further evidence. Risk acceptance does not raise a capability rating.

Retain the record

Keep artifacts, sources, dates, populations, samples, test results and assessor rationale. Distinguish provider controls from customer duties. Review material judgments and retain disagreements.

A policy supports design. Operating records support implementation. Representative tests support effectiveness.

Read the foundational white paper

A method built
for scrutiny.

Explore the five dimensions, scoring rules and the path to independent validation.

Read the white paper

PDF · v0.2 · Consultation draft. This site’s v0.3 proposal adds explicit three-pillar delivery, use-case risk tiering and control implementation. Pilot validation remains ahead.

Reference foundations and development status

Built on established guidance.
Open to independent challenge.

MITRE SAFE-AI supplies the security assessment foundation through AI-specific consideration of controls. NIST AI RMF informs broader governance and risk outcomes. Ascendra’s five dimensions, scoring, risk-tier examples and presentation are proposed design choices, not official mappings or endorsements.

  1. MITRE SAFE-AI — A Framework for Securing AI-Enabled Systems ↗
  2. NIST AI RMF — Core outcomes ↗
  3. NIST — AI risks and trustworthiness ↗

Version 0.3 · 13 October 2026 · Consultation proposal. The next stage is to complete assessment anchors and crosswalks, run calibration pilots and test agreement between independent assessors. No certification, industry percentile or validated benchmark is claimed.

Contribute feedback
Make the next decision with clarity

Bring your AI ambition.
Build an informed plan.

Start with your use cases, priorities and existing controls. We’ll help define a practical assessment scope.

Book Consultation