Improve · autonomy program
Autonomy is earned one exit criterion at a time
Nobody jumps to eighty percent agentic. Each phase has published exit criteria, and an agent only moves up a band when the evidence says it has earned the next one.
Estate autonomy today vs. committed target
73% → 83%The autonomy ladder
What each band actually permits.
Assisted
Human does the work; the agent only drafts or retrieves.
Recommended
Agent proposes, a human reviews and commits every time.
Supervised
Agent executes inside limits; a human samples and can reverse.
Delegated
Agent executes and closes; humans see exceptions only.
Autonomous
Agent owns the outcome and the policy loop; humans audit after.
Phase 1 — Sensing and evidence
A0-A1
Agents observe, retrieve and assemble. They produce briefs, exception lists and evidence packs, but they do not act on systems of record. The objective is to establish provenance discipline and a trustworthy baseline before any authority is delegated.
Exit criteria
- Provenance attached to every material output
- Baseline measured for each in-scope process
- Agent registry populated with named human owners
- Evidence vault operating with tamper-evident logs
Phase 2 — Supervised execution
A1-A2
Agents draft and prepare actions that a human confirms before execution. Override reasons are captured as structured data and become the primary training signal for threshold calibration.
Exit criteria
- Override rate below 15% on target processes
- Structured override reasons captured on every rejection
- Rollback tested for every reversible action type
- Human review capacity measured and staffed
Phase 3 — Functional orchestration
A2-A3
Each function runs a Chief of Staff agent that decomposes work, assigns specialists and escalates by published threshold. Autonomous execution is permitted for high-confidence, reversible, in-policy work.
Exit criteria
- Chief of Staff live in every in-scope function
- Touchless rate above 60% on standardized work
- Every gate mapped to a named accountable human
- Continuous control testing replacing quarterly sampling
Phase 4 — Governed autonomy
A3-A4
Selected workflows operate at A4 within tight envelopes, with continuous evaluation, automatic suspension on drift and quarterly recertification. Autonomy increases require business-owner and risk approval, never an engineering decision.
Exit criteria
- Quarterly recertification passing for every A4 agent
- Automatic suspension proven on injected drift
- Value ledger reconciled to finance-grade baseline
- Cross-functional flows running without dropped handoffs
Phase 5 — Enterprise coordination
A3-A4 with enterprise coordination
The Super Orchestrator holds enterprise commitments across all fourteen functions, allocates scarce human review capacity and rehearses scenarios before they occur. Human effort concentrates on strategy, judgment and exception.
Exit criteria
- Enterprise outcome map maintained continuously
- Scenario rehearsal run before each planning cycle
- Human capacity allocated by measured demand
- Released capacity redeployed against a published plan