Operate · Approvals
The approval workflow.
Nothing modifies your infrastructure without explicit approval. Risk classification determines who must approve. Approvals are immutable, audit-logged, and can be revoked before apply.
The default
01
Why approvals exist
Autonomous agents that bypass human review are unsafe at enterprise scale. Three reasons approvals are mandatory:
- Accountability — every change is attributed to a human approver in the audit trail
- Risk containment — multi-party approval prevents single-actor mistakes on broad changes
- Trust calibration — auto-apply autonomy escalates per action class only after a measured history of successful outcomes
02 · Risk tiers
What requires approval
Low-risk (contained blast radius)
Examples: deleting an unused EBS volume, terminating an idle Lambda, modifying a non-production tag.
Default: single approver. On the Trust Ladder enterprise tier, low-risk operations can be set to auto-apply after a configured number of successful outcomes.
Medium-risk (moderate blast radius)
Examples: rightsizing an EC2 instance behind an ALB, modifying an RDS parameter group, rotating an IAM access key.
Default: single approver with mandatory rollback path verification. Never auto-applied.
High-risk (broad blast radius)
Examples: modifying IAM policy on a privileged role, changing a security group attached to many resources, dropping a database resource.
Default: multi-party approval (2+ approvers). Can never be auto-applied regardless of trust ladder state.
03
Who can approve
By default, any organization member can approve. Enterprise tier supports:
- Role-based approval gates (only members with the "approver" role)
- Resource-scoped approvers (only specific resource owners)
- Environment-specific approvers (production approvers vs. staging)
- External approval via ServiceNow Change Request or PagerDuty escalation
04 · Audit
Audit trail
Every approval generates an immutable AxiomApprovalRequest + AxiomAuditEvent pair with:
- Approver identity
- Timestamp
- Plan items approved and explicitly rejected
- Optional note from approver
- Pre-apply state and post-apply state
- Rollback strategy at time of approval
Approvals are revocable until apply begins. After apply starts, rollback is the only path.
05 · Safe autonomy boundaries
The Trust Ladder
Trust Ladder is the model that lets agent autonomy escalate over time — only after measured success.
- Per action class, the agent tracks success/failure ratio
- After N consecutive successful outcomes (configurable, default 25), the action class can be promoted to auto-apply
- Promotion always requires explicit org-admin approval — the agent cannot self-escalate
- Any failure in an auto-apply class immediately demotes that class back to manual approval
- High-risk classes can never be auto-applied regardless of history
Trust questions
What is happening at approval?
A human is reviewing a proposed change to your infrastructure before Axiom applies it.
Why approvals?
Accountability + risk containment + safe autonomy. The default for any autonomous system at enterprise scale.
Is it safe to approve?
Yes — each item includes blast radius, rollback path, RTO, and Terraform diff. You see exactly what will change.
What happens after I approve?
Pre-flight snapshot → apply → verification. Rollback fires automatically on health failure.
Can I revoke an approval?
Yes, until apply begins. After apply, rollback is the recovery path.
What if approval expires?
After 24 hours unapplied, the plan expires. Re-scan to generate a fresh plan against current state.
Need a human?
Most flows are documented — but we'll help if anything is unclear.