Metric 0 Pre-Check
Before any Agentic Governance Charter review proceeds, three gates must pass. All three must be satisfied before proceeding.
Gate 1 — GOV-02 AI Use Policy Coverage: Confirm an AI Use Policy entry exists in GOV-02 for the agentic tool covering the specific use case proposed for autonomous operation. An entry covering only advisory use does not satisfy this gate for Tier 3–4 deployment. If failed: create or amend the GOV-02 entry before proceeding.
Gate 2 — DAT-06 AI BoM Registration at Tier 3 or 4: Confirm the tool is registered in the AI BoM (DAT-06) at Provisional or Active status with a draft agentic tier classification of Tier 3 or Tier 4. Unregistered tools must complete DAT-06 registration before governance review. If failed: initiate DAT-06 registration immediately — this Charter review is blocked.
Gate 3 — No Open Class 6 Shadow AI Incidents: Confirm via STR-07 that no open Class 6 incidents relate to this tool, its underlying model, or the proposed autonomous workflow. If failed: resolve or formally close the STR-07 incident before proceeding. If the tool was operating autonomously without authorisation, apply the Emergency Decommissioning Protocol (SUS-06) before any new authorisation.
---
1. Purpose
GOV-08 establishes the Agentic AI Governance Charter — the organisation’s binding policy framework for deploying, operating, overseeing, and decommissioning AI systems operating at Agentic Tier 3 (Supervised Autonomous) or Tier 4 (Executor) as classified under DAT-06.
This Charter governs what happens after registration: the formal approval process, the mandatory governance provisions that must be in place before autonomous operation begins, the human oversight requirements that persist throughout deployment, the escalation pathways when autonomous systems fail or exceed scope, and the renewal cycle that keeps governance current.
GOV-08 does not govern Tier 0 (Advisory), Tier 1 (Augmented), or Tier 2 (Automated) tools — those tiers are governed under standard DAT-06 registration and GOV-02 policy controls. This Charter is reserved for AI systems that act autonomously on client matters, execute multi-step legal workflows without human sign-off at each step, or take actions with direct legal, financial, or privilege consequences.
---
2. Strategic Context
Agentic AI represents a qualitatively different risk profile from advisory or automated tools. An AI system that drafts a document (Tier 1) creates risk only if the human fails to review. An AI system that files, sends, or commits on behalf of the organisation (Tier 4) creates risk through the act itself — before any human review is possible.
Legal environments amplify this risk in three ways:
- Privilege and confidentiality attach to specific communications and work product, and autonomous processing that does not respect these boundaries may waive privilege or breach professional obligations.
- Professional liability standards assume human professional judgment — automated decisions may not satisfy the duty of competence or the duty of supervision.
- Client expectations and contracts increasingly require disclosure of AI use and assurances of human oversight.
Regulatory frameworks are converging on the same requirement. The EU AI Act (Art. 14) mandates human oversight measures for high-risk AI systems, including meaningful intervention capability. ISO/IEC 42001 requires documented procedures for autonomous AI operations. The NIST AI RMF emphasises testing, monitoring, and incident response for agentic systems.
Without a dedicated Agentic Governance Charter, the organisation cannot demonstrate to regulators, clients, or courts that autonomous AI was deployed with appropriate control — a gap that transforms AI efficiency gains into professional liability exposure.
---
3. Operating Principles
Principle 1 — Tiered Governance: Governance intensity scales with agentic tier. Tier 3 and Tier 4 tools require Charter authorisation, mandatory provisions, and formal approval panels regardless of tool familiarity or perceived risk level.
Principle 2 — Mandatory Pre-Deployment Authorisation: No AI tool may operate at Tier 3 or Tier 4 without a current Agentic Deployment Authorisation Certificate. Provisional DAT-06 status does not confer operational authority. The Certificate is the authority.
Principle 3 — Kill-Switch First: The kill-switch mechanism must be tested and confirmed functional before any Tier 3 or Tier 4 tool goes live. A tool whose kill-switch has not been confirmed functional is not authorised to operate autonomously.
Principle 4 — Scope Immutability: The Autonomous Action Scope Boundary (DAT-06 Field 22) is immutable between formal reauthorisations. Scope drift — even with beneficial intent — triggers an out-of-cycle review and may trigger emergency decommissioning if confirmed.
Principle 5 — Intervention Logging is Non-Negotiable: All human interventions in autonomous AI workflows must be logged. Disabling, bypassing, or deleting intervention logs without AI Governance Lead authorisation is a Class 6 event.
Principle 6 — Bias Monitoring Persists: Continuous bias and drift monitoring for Tier 3–4 tools does not lapse between reviews. A monitoring gap of more than 72 hours without documented justification triggers a mandatory review.
Principle 7 — Incidents Reset the Clock: Any Tier 3 or Tier 4 incident triggers an immediate review regardless of where the tool sits in its reauthorisation cycle. A tool may not return to autonomous operation until the incident is resolved and reauthorisation confirmed.
Principle 8 — Transparency to Clients: When a Tier 3 or Tier 4 tool takes autonomous action on a client matter, disclosure obligations apply. Non-disclosure is not an option for confirmed autonomous actions on client work.
Principle 9 — Evidence-First Mindset: Every governance action under this Charter — approval, monitoring, intervention, incident response, renewal — is documented at the time it occurs. Retrospective reconstruction of governance evidence is not acceptable.
Principle 10 — No Autonomous Authorisation: AI systems may not authorise other AI systems to operate at Tier 3 or Tier 4. All agentic authorisation requires human principal approval through the Agentic Governance Panel.
---
4. Scope and Tier Applicability
This Charter applies to all AI tools operating at Agentic Tier 3 (Supervised Autonomous) or Tier 4 (Executor) as defined in DAT-06.
Tier 3 — Supervised Autonomous: Executes multi-step workflows autonomously with escalation triggers that pause for human review at defined checkpoints. Examples include automated contract review and flagging workflows, and AI-orchestrated due diligence with human sign-off at key milestones.
Tier 4 — Executor: Executes autonomous actions within a defined boundary without mandatory human review at each step. Examples include autonomous filing, scheduling, or notification systems acting on behalf of the organisation; AI agents orchestrating other AI tools as sub-agents; and systems making financial or legal commitments within a predefined scope.
This Charter does not apply to Tier 0 (Advisory), Tier 1 (Augmented), or Tier 2 (Automated) tools — those tiers are governed by standard DAT-06 registration and GOV-02 policy controls.
Where a tool’s tier classification is disputed or uncertain, it is treated as the higher tier until formal reclassification is completed.
---
5. Agentic Tier Governance Framework
The following governance requirements apply in addition to standard DAT-06 registration requirements.
Tier 3 — Supervised Autonomous
- All five Mandatory Governance Provisions must be operational before deployment.
- Agentic Governance Panel approval is required.
- Human oversight: minimum 20% random sampling of autonomous decisions per operating week, with escalation review for all sampled exceptions.
- Kill-switch test: confirmed functional before go-live and monthly thereafter.
- Bias monitoring: minimum quarterly reports; immediate escalation on detecting statistically significant drift.
- Reauthorisation: annual, plus event-triggered.
Tier 4 — Executor
- All five Mandatory Governance Provisions must be operational before deployment.
- Agentic Governance Panel approval is required.
- Executive Sponsor sign-off additionally required (Partner, CISO, or equivalent).
- Human oversight: continuous parallel monitoring of all autonomous actions; human review required within 24 hours of any autonomous commitment.
- Kill-switch test: confirmed functional before go-live; tested weekly thereafter; test result logged.
- Bias monitoring: continuous monitoring with automated alerts; monthly reports reviewed by AI Governance Lead.
- All three Agentic-Specific Fields (DAT-06 Fields 20–22) must be fully populated.
- Reauthorisation: every 6 months, plus event-triggered.
---
6. Mandatory Governance Provisions — Tiers 3 and 4
All five provisions are mandatory. No Tier 3 or Tier 4 tool may operate without all five confirmed in place. The AI Governance Lead must sign off that each provision is operational before issuing the Agentic Deployment Authorisation Certificate.
Provision 1 — Kill-Switch Mechanism
A confirmed, tested ability to immediately and completely halt all autonomous actions of the tool within a defined time window.
- The kill-switch must be accessible to the AI Governance Lead and at least one IT Security designee at all times, including outside business hours.
- Kill-switch activation must immediately halt all queued and in-progress autonomous actions without data corruption or partial commits.
- Activation time must be confirmed at or below the agreed threshold: Tier 3, within 15 minutes of decision; Tier 4, within 5 minutes.
- Each kill-switch activation is documented with timestamp, reason, authorising officer, and post-activation state of all in-progress tasks.
- Kill-switch testing logs are retained for 7 years.
Provision 2 — Intervention Log
A system of record capturing all human interventions in the autonomous AI workflow.
- Every human decision to override, modify, escalate, or approve an autonomous action is logged with timestamp, decision-maker identity, and the specific action affected.
- Logs are immutable — no deletion or modification after entry without AI Governance Lead written authorisation.
- Logs are exported quarterly and retained for 7 years from the date of the last entry.
- On decommissioning, a final export is performed before shutdown and retained for 7 years.
Provision 3 — Scope Limitation Verification
Formal, documented confirmation that the tool’s autonomous actions are bounded to the Approved Autonomous Action Scope Boundary (DAT-06 Field 22).
- The scope boundary is a written, granular statement of all permitted autonomous action types, data categories, counterparties, and output formats.
- Technical controls (API restrictions, data access controls, network isolation) are implemented to enforce the boundary — the tool must be technically incapable of actions outside the scope.
- Any automated action outside the technical boundary triggers an immediate alert and automatic escalation to the AI Governance Lead.
- Scope changes require a full reauthorisation cycle before the new scope is operational.
Provision 4 — Escalation Protocol
A defined, documented protocol specifying conditions under which the autonomous AI workflow pauses and routes to a named human decision-maker.
- Escalation triggers are defined in writing: specific decision types, confidence thresholds, data categories, or flag conditions requiring human review.
- Each trigger maps to a named responsible person and a backup.
- Response time commitments are defined for each escalation class.
- All escalation recipients are notified when a new agentic deployment goes live and when the protocol changes.
- Escalation records are retained for 5 years.
Provision 5 — Continuous Bias and Drift Monitoring
Ongoing monitoring of the tool’s outputs for statistically significant bias, quality drift, or factual degradation.
- A monitoring baseline is established during Technical Review before authorisation.
- Monitoring frequency: continuous alerting for Tier 4; weekly batch reviews at minimum for Tier 3.
- Monitoring scope covers output bias indicators, hallucination rate proxies, accuracy delta from baseline, and any demographic or jurisdictional disparities.
- Monitoring data is retained for 5 years.
- A statistically significant adverse trend triggers escalation to the AI Governance Lead within 24 hours and a mandatory scope review within 5 business days.
- At decommissioning, final monitoring data is exported and retained for 5 years.
---
7. Approval Authority and Governance Sign-Off
Agentic Governance Panel
All Tier 3 and Tier 4 deployments require approval from an Agentic Governance Panel comprising three members:
- AI Governance Lead (Chair),
- Legal representative (Partner or Senior Associate with relevant practice area expertise), and
- IT Security representative.
For Tier 4 deployments, an Executive Sponsor (Partner-level or CISO) provides additional sign-off confirming organisational risk acceptance.
Approval requires unanimous agreement of all three Panel members. A single Panel member may block deployment by raising a written objection. The objection is logged, addressed, and re-reviewed before any further vote. Conditional approvals are permissible for Tier 3 only — all conditions must be satisfied before any Tier 4 Certificate is issued.
The AI Governance Lead may not delegate their Panel role. Legal and IT Security representatives may designate qualified alternates in writing, reviewed annually.
---
8. Human Oversight Requirements by Tier
Tier 3 — Supervised Autonomous
- Weekly: review a 20% random sample of autonomous decisions from the prior week; document review completion and any exceptions identified.
- Monthly: review escalation protocol logs; confirm escalation recipients remain current and response times are being met.
- Quarterly: review bias monitoring report; confirm scope limitation remains technically enforced; update AI BoM entry if any material change has occurred.
- Per incident: immediate human review of all autonomous actions taken within the 24-hour window preceding the incident trigger.
Tier 4 — Executor
- Continuous: parallel monitoring of autonomous actions by a designated human monitor or automated monitoring system with human review of alerts within 4 hours.
- Daily: review of all autonomous commitments made in the prior 24-hour period; flag any actions approaching scope boundaries.
- Weekly: kill-switch test; bias monitoring alert review; escalation protocol status check.
- Monthly: AI Governance Lead reviews full monitoring logs; certifies ongoing compliance with the Agentic Deployment Authorisation Certificate conditions.
- Per incident: immediate halt of autonomous operations pending root cause determination.
---