The Autonomy Ladder

From Shadow to Guardrailed Autonomy — a safe ladder for AI in production

The same person wouldn't put a brand-new employee in charge of wire transfers on day one. Neither should AI. HUMAIPAIR's 5-mode ladder graduates trust one certified capability at a time, with approval gates, telemetry, and instant regression.

See the ladder Governance model

The five modes at a glance

Each mode is defined by what the AI may do autonomously and what still needs a human signature.

Horizontal 5-step autonomy ladder from Shadow to GuardrailedAutonomy 1 Shadow Observes, never acts AI does Mirror decisions, log reasoning Approval Every action (none permitted) 2 Advisor Suggests, human decides AI does Propose options, rank and explain Approval Human selects and executes 3 Copilot Drafts, human reviews AI does Draft outputs, prefill forms Approval Human signs off before submit 4 Controlled Action Acts with per-action approval AI does Execute in-scope actions live Approval Each action gated by human or policy 5 Guardrailed Autonomy Runs in policy envelope AI does Executes without per-call approval Approval Exceptions only: out-of-envelope
Figure 1. The ladder: intensity of color tracks autonomy. Each step is a distinct certified mode.

Each mode, spelled out

Entry criteria, division of labor, and the KPI gate that promotes to the next rung.

1

Shadow

Observation only
Entry criteria
Role blueprint approved; AI counterpart provisioned with read-only access.
What the AI does
Mirrors every decision the human makes; records its own preferred action and reasoning.
What the human does
Works normally; periodically reviews the AI's shadow log and flags disagreements.
Promotion gate → Advisor
Agreement rate ≥ 85% over 100 tasks & zero high-severity disagreements.
2

Advisor

AI suggests, human decides
Entry criteria
Shadow promotion passed; Advisor rubric certified by supervisor.
What the AI does
Surfaces ranked suggestions inline with the human's workflow; explains top pick.
What the human does
Accepts, edits, or rejects suggestions. Rejections become training feedback.
Promotion gate → Copilot
Acceptance rate ≥ 70% + quality rubric pass for 200 tasks.
3

Copilot

AI drafts, human reviews
Entry criteria
Advisor gate passed; knowledge pack versioned and signed.
What the AI does
Produces full drafts end-to-end. Completes forms, writes responses, proposes dispositions.
What the human does
Reviews and submits. Override rate is measured per task class.
Promotion gate → ControlledAction
Override rate ≤ 10% over 500 tasks; task class scoped to the policy envelope.
4

ControlledAction

AI acts with per-action approval
Entry criteria
Copilot gate passed; ai_ops has signed off on action-class scope.
What the AI does
Executes in-scope actions (API calls, writes, notifications). Each call logs a pre-action record.
What the human does
Approves or blocks each action via the pair console or policy auto-approval.
Promotion gate → Autonomy
Approval rate ≥ 95% + incident rate < 0.5% over 1,000 actions.
5

GuardrailedAutonomy

Policy envelope enforces
Entry criteria
ControlledAction gate passed; org_admin countersignature on envelope.
What the AI does
Executes without per-call approval inside the envelope; cost, rate, and risk caps enforce.
What the human does
Reviews exception queue only; handles out-of-envelope or high-risk escalations.
Promotion gate → next envelope
Quarterly recertification; expanded envelope requires a full ladder pass for new actions.

The transition state machine

Promotions always require supervisor + ai_ops approval. Regressions are always allowed and can be triggered by the pair, supervisor, or an automated policy.

State machine showing forward promotions requiring approval and regressions always allowed Shadow Mode 1 Advisor Mode 2 Copilot Mode 3 Controlled Action Mode 4 Guardrailed Autonomy Mode 5 promote (KPI + approval) promote (KPI + approval) promote (KPI + approval) promote (KPI + approval) regress (anytime) regress (anytime) regress (anytime) regress (anytime) promotion (gated) regression (always allowed)
Figure 2. Green arrows need approval; red arrows are always allowed and trigger re-certification.

Ready to map this to your roles?

See how we translate the ladder into blueprints for customer support, claims, compliance, and sales ops.

See role blueprints