The VERITRAIL Framework is the backbone of everything VERITRAIL does. The self-assessment scores against it, the VERITRAIL Agent cites it, Shield policies reference it, and evidence packs are organized by it. It was designed around one test: can an organization show, with evidence, that a control is operating?
Why another framework?
NIST AI RMF, ISO/IEC 42001 and the EU AI Act each describe what good AI governance should achieve, at different levels of abstraction and with different legal force. Teams implementing them run into the same practical problems:
- The unit of work is unclear. A framework outcome such as "AI risks are identified and documented" does not tell an engineering lead what to build or an auditor what to look at.
- Ownership is diffuse. Without a named owner role per control, everything belongs to "governance" and nothing gets done.
- Evidence is an afterthought. Policies get written; the records that prove the policy is followed do not.
- Agents are not covered well. Tool access, autonomous actions and runtime containment need explicit controls, not just model-centric guidance.
The VERITRAIL Framework addresses these by defining each control with an objective, a scope (all, high_risk, genai, agentic or third_party), an owner role, a priority, the evidence that demonstrates it, the VERITRAIL module that implements it, and indicative framework mappings.
Structure
| Component | What it is | Where to go |
|---|---|---|
| Five pillars | The organizing structure: accountability, inventory, lifecycle, integrity/safety/security, and assurance | Five pillars |
| Control library | 64 controls (AO, EI, GL, IS, SE), versioned, with evidence requirements and mappings | Control library |
| Maturity model | Five levels and eight weighted domains used by the self-assessment | Maturity model |
| Assessment modules | Organization-level and use-case-level assessments, triggered by risk tier | Assessment modules |
| Implementation runbooks | Step-by-step procedures to stand up and operate key controls | Runbooks |
The five pillars at a glance
Accountability & Operating Model (AO-01 to AO-12). Policy, an executive forum with authority to stop use cases, RACI, risk appetite, AI literacy, escalation, obligations register, prohibited-use screening and resourcing. Read more.
Enterprise AI Inventory (EI-01 to EI-12). A single system of record for AI systems, models, prompts, data sources, agents and their tools, vendor AI and shadow AI, with risk tiers and named owners. Read more.
Governance & Lifecycle Control (GL-01 to GL-12). Intake triage, pre-deployment assessment, risk-tiered approvals, change triggers, validation, technical documentation, go-live checks, monitoring, retirement, exceptions, vendor due diligence and recertification. Read more.
Integrity, Safety & Security (IS-01 to IS-16). Threat modelling, data quality, privacy, redaction, fairness, safety, robustness, transparency, human oversight, supply chain, least-privilege tool access, runtime guardrails, human approval for high-impact actions, adversarial testing, kill switch and resource limits. Read more.
Stewardship, Evidence & Assurance (SE-01 to SE-12). Control testing, evidence retention, metrics, incident management, evidence packs, corrective action, management review, tamper-evident logging, post-market monitoring, internal audit, transparency reporting and drift monitoring. Read more.
The controls that matter most on day one
Of the 64 controls, 17 are marked critical priority. If you can evidence nothing else, these are the ones that close the largest exposures — no inventory, no pre-deployment assessment, unprotected sensitive data, untested agents and no way to stop a misbehaving system:
VERITRAIL controls cited
- AO-01 AI governance policyNIST AI RMF: GOVERN 1.1, GOVERN 1.2ISO/IEC 42001: 5.2, A.2.2, A.2.4EU AI Act: Art. 17, Art. 26
- AO-02 Executive accountability and AI governance forumNIST AI RMF: GOVERN 2.1, GOVERN 2.3ISO/IEC 42001: 5.1, 5.3, A.3.2EU AI Act: Art. 17
- AO-10 Prohibited and restricted AI use policyNIST AI RMF: GOVERN 1.1, MAP 1.1ISO/IEC 42001: A.9.4, A.2.2EU AI Act: Art. 5
- EI-01 Enterprise AI inventory (system of record)NIST AI RMF: GOVERN 1.6ISO/IEC 42001: A.4.2, 7.5EU AI Act: Art. 26, Art. 49
- EI-02 AI risk classification and tieringNIST AI RMF: MAP 1.1, MAP 1.5, GOVERN 1.3ISO/IEC 42001: 6.1.2, A.5.2EU AI Act: Art. 6
- EI-03 Accountable owner per AI systemNIST AI RMF: GOVERN 2.1ISO/IEC 42001: A.3.2, A.10.2EU AI Act: Art. 26
- EI-07 Agent and tool registryNIST AI RMF: GOVERN 1.6, MAP 4.1ISO/IEC 42001: A.4.4, A.6.2.3EU AI Act: Art. 11
- GL-02 Pre-deployment risk and impact assessmentNIST AI RMF: MAP 1.1, MAP 3.1, MAP 5.1, MEASURE 1.1ISO/IEC 42001: 6.1.4, 8.4, A.5.2, A.5.3, A.5.4EU AI Act: Art. 9, Art. 27
- GL-03 Risk-tiered approval gatesNIST AI RMF: MANAGE 1.1, GOVERN 1.4ISO/IEC 42001: 8.1, A.6.2.5EU AI Act: Art. 9, Art. 43
- IS-03 Privacy assessment and data minimisationNIST AI RMF: MEASURE 2.10, MAP 1.1ISO/IEC 42001: A.7.4, A.5.4EU AI Act: Art. 10
- IS-04 Sensitive data redaction and maskingNIST AI RMF: MEASURE 2.10ISO/IEC 42001: A.7.4, A.6.2.6EU AI Act: Art. 10
- IS-09 Human oversight designNIST AI RMF: GOVERN 3.2, MAP 3.5ISO/IEC 42001: A.9.2, A.6.2.6EU AI Act: Art. 14, Art. 26
- IS-11 Least-privilege tool accessNIST AI RMF: MEASURE 2.7, MANAGE 2.2ISO/IEC 42001: A.6.2.6, A.4.4EU AI Act: Art. 15, Art. 14
- IS-13 Human approval for high-impact actionsNIST AI RMF: GOVERN 3.2, MANAGE 2.2ISO/IEC 42001: A.9.2, A.6.2.6EU AI Act: Art. 14
- IS-14 Prompt-injection and data-leakage testingNIST AI RMF: MEASURE 2.7ISO/IEC 42001: A.6.2.4EU AI Act: Art. 15
- IS-15 Kill switch and rollbackNIST AI RMF: MANAGE 2.4ISO/IEC 42001: A.6.2.6EU AI Act: Art. 14
- SE-08 Immutable action loggingNIST AI RMF: MEASURE 2.8, MANAGE 4.1ISO/IEC 42001: A.6.2.8EU AI Act: Art. 12, Art. 19, Art. 26
Framework mappings are indicative VERITRAIL interpretations (Control Library v1.0.0); they are not legal advice and do not establish conformity or compliance.
Evidence-first by design
Every VERITRAIL finding carries one of five evidence labels:
| Label | Meaning |
|---|---|
| Verified | Evidence exists and has been reviewed by a named reviewer |
| User-asserted | Someone states the control is implemented, but evidence has not been provided |
| Inferred | Inferred as a likely gap or fact from responses or metadata |
| Missing evidence | Required evidence has not been provided |
| Critical gap | Material privacy, security, safety, legal or operational exposure |
Self-assessment answers are always User-asserted. Only evidence reviewed by a named human becomes Verified. This distinction is what lets a board pack, an audit response or a customer questionnaire be honest about what is known. The full method is described in Gap analysis.
Relationship to external frameworks
Each control carries indicative mappings to NIST AI RMF 1.0 subcategories, ISO/IEC 42001:2023 clauses and Annex A controls, and EU AI Act (Regulation (EU) 2024/1689) articles. See NIST AI RMF, ISO/IEC 42001 and EU AI Act readiness for how to use them.
Get the full framework
The VERITRAIL Enterprise Blueprint packages the pillars, the control library, the maturity model and a phased rollout plan in a single document for your governance forum.