Autonomy is earned: five levels for governing computer agents

Every operations leader considering computer agents asks a version of the same question: how much should we let the agent do on its own?

It is the right question, and the industry has mostly answered it badly. One camp says “keep a human in the loop” — which, applied literally, means a person approves every keystroke and nothing gets faster. The other camp says “full autonomy” — which no risk committee will sign, and rightly so.

The honest answer is that it depends on the workflow, and it should change over time. A new agent on an unfamiliar process deserves the same supervision as a new hire. An agent that has processed fifty thousand invoices without a single policy breach deserves more latitude. Trust should follow evidence.

We built Autofil around that principle.

The five levels

Level 1 — Observe. Autofil watches live work, drafts the Playbook and predicts what it would have done. It touches nothing. You get a measured accuracy score before an agent has write access to anything.

Level 2 — Suggest. Every case is gathered, keyed and validated, then staged for a person to review and submit. Nothing commits without a human, but the human’s job shrinks from ten minutes of typing to ten seconds of checking.

Level 3 — Supervised. Autofil runs the workflow end to end, but pauses for approval on anything irreversible — payments, submissions, deletions, external communications. This is where most workflows live in their first quarter.

Level 4 — Autonomous within policy. Straight-through processing inside thresholds you define. Anything novel, ambiguous, low-confidence or over limit is escalated with full context.

Level 5 — Orchestrated. Agents hand work to one another across workflows — vendor onboarding to invoice posting to payment — under a single control plane with continuous controls testing.

Promotion is a policy, not a vibe

The levels only matter if moving between them is disciplined. In Autofil, each workflow has a promotion policy written by its owner and approved by risk. A typical one reads:

Promote from L3 to L4 after 500 consecutive runs with ≥ 99.5% field-level accuracy, zero policy violations, and sign-off from the process owner and the controller.

Autofil tracks the evidence automatically and proposes promotion when the bar is met. A human still makes the call.

Just as important is automatic demotion. If Autofil detects drift — a screen that changed, a shift in the data, falling confidence — the workflow steps down a level and alerts its owner. Autonomy can be lost faster than it is gained, which is exactly how it works with people.

Why this matters to the board

Autonomy levels turn an abstract AI-risk debate into a concrete operating decision. Instead of asking “do we trust AI?”, leaders ask “which of our workflows have earned Level 4?” — and they get an answer backed by data.

It also gives internal audit something it has never had with automation: a documented, versioned record of why each agent is allowed to do what it does, who approved it, and the evidence they relied on.

How Retypd enforces it

• Policy-as-code guardrails in Retypd Governance, evaluated outside the model before every action
• Approval tasks routed by Retypd Orchestrate, including to Slack and Microsoft Teams
• Drift detection with automatic slow-down and demotion
• A versioned record in Retypd Governance of every promotion, the evidence behind it and who approved it

To see how the levels would apply to your workflows, talk to us — we map them with you in the first working session.