Agentic AI Deployment Readiness Assessment
Require a structured pre-deployment readiness assessment for tool-enabled AI agents, verifying that key governance controls are in place and that the agent's impact on connected systems has been evaluated before go-live.
Objective
Prevent premature deployment of agentic AI systems by establishing a governance gate that verifies control maturity, documents deployment impact, and obtains cross-functional sign-off before an agent is permitted to operate in production.
Maturity Levels
Initial
Agentic AI systems are deployed without a structured readiness process. Go/no-go decisions are made informally by the engineering or product team.
Developing
Some pre-deployment checklist exists but it is not specific to agentic systems. Coverage of tool access, permission scope, and impact on connected systems is absent or inconsistent.
Defined
A formal agentic AI deployment readiness assessment is required before any tool-enabled agent reaches production. The assessment covers control maturity (permissions, kill switch, audit logging), impact on connected systems, and escalation procedures. Sign-off from the AI governance function is required.
Managed
Assessment results are recorded and retained. A deployment register tracks all agentic systems in production with their readiness assessment date and outstanding remediation items. Re-assessment is triggered when an agent's tool access or autonomy scope changes materially.
Optimizing
Readiness assessments are automated in part: control checks that can be verified programmatically (e.g., kill switch wired, audit logging active) are checked automatically. Human review focuses on judgment-dependent items. Assessment results feed into the enterprise AI risk register.
Evidence Requirements
What an auditor or assessor would expect to see for this control.
- —Completed agentic AI deployment readiness assessment for each tool-enabled agent in production, including permission scope documentation, control completeness checklist, and impact assessment.
- —Sign-off records from technical owner, security/data function, and AI governance function.
- —Agentic AI deployment register showing all production agents, assessment dates, and any outstanding remediation items.
Implementation Notes
Distinction from model deployment gates
This control is distinct from model-performance-based deployment gates (which evaluate accuracy, drift, and bias metrics). Agentic deployment readiness focuses on governance and operational controls: does the agent have a kill switch? Are its permissions appropriately scoped? Has the blast radius of a malfunction been documented?
Key assessment domains
Permission and scope verification
- Agent permissions are scoped to the minimum required for the defined task. No open-ended tool access.
- Permission scope is documented and approved by a named owner.
- Any elevated permissions (write access, cross-system access, identity assumption) have an explicit justification.
Control completeness
- Kill switch or emergency halt is wired and tested.
- Audit logging is active and outputs are being collected.
- Human approval gates are defined for irreversible or high-consequence actions.
- Agent identity is registered in the NHI management system.
Impact assessment
- Systems the agent can read from, write to, or call are documented.
- Maximum blast radius of a malfunction or misuse is estimated: what data could be corrupted or exfiltrated? What services could be disrupted?
- Rollback or recovery procedure is documented.
Stakeholder readiness
- Operations team responsible for monitoring the agent post-deployment is identified and briefed.
- Escalation procedure for agent-related incidents is defined.
- Users or counterparties affected by agent actions are aware the agent exists (where disclosure is appropriate).
Gating and sign-off
The assessment should require sign-off from: (1) the technical owner verifying control completeness, (2) the data or security function verifying impact assessment, and (3) the AI governance function verifying overall readiness. For high-risk agents, board-committee review under AGT-023 may also be required.
What makes an agent 'high-risk' for this control
Triggers for elevated review: agents with write access to production databases, agents acting on behalf of users without per-action confirmation, agents with cross-system tool access, agents handling regulated data, and agents operating in environments with limited reversibility.
Example Implementation
Agentic AI Deployment Readiness Assessment (template excerpt)
Agent: Customer Refund Processing Agent | Version: 1.2 | Assessment date: 2026-06-01
1. Permission and scope verification
| Check | Status | Notes |
|---|---|---|
| Tool access list documented | Pass | CRM read, Payments API write (refund endpoint only), Audit log write |
| Permissions scoped to minimum required | Pass | Payment write scope limited to refunds ≤$500; amounts above require human approval |
| Elevated permissions justified | Pass | Payment write approved by Head of Payments and CISO on 2026-05-28 |
| No open-ended or wildcard tool access | Pass | — |
2. Control completeness
| Check | Status | Notes |
|---|---|---|
| Kill switch wired and tested | Pass | Tested 2026-05-30; halt confirmed within 8 seconds |
| Audit logging active | Pass | All tool calls logged to agent-audit stream |
| Human approval gate for irreversible actions | Pass | Refunds >$500 route to human queue; cancellations always require confirmation |
| NHI identity registered | Pass | NHI ID: SVC-REFUND-AGENT-001 |
3. Impact assessment
- Systems affected: CRM (read), Payments API (write — refund endpoint), Audit log (write)
- Maximum blast radius: Erroneous refunds up to $500 per transaction. Agent rate-limited to 50 transactions/hour. Maximum exposure per hour: $25,000. Rate limit alert at 40 transactions/hour.
- Rollback procedure: Payments team can reverse agent-initiated refunds within 24 hours via Payments API reversal endpoint. Procedure documented in runbook P-027.
4. Sign-off
| Role | Name | Date | Status |
|---|---|---|---|
| Technical owner | J. Reyes | 2026-06-01 | Approved |
| Security function | T. Okafor | 2026-06-01 | Approved |
| AI governance | C. Müller | 2026-06-02 | Approved — cleared for production |