• Explainer
  • Concepts
  • 5 min read

The Autonomy Ladder

A four-level model for granting an AI employee more freedom in a way you can defend — and step back down when circumstances change.

Author
ftes.ai Editorial
ftes.ai Team

The short answer

The autonomy ladder is a four-level model for granting an AI employee freedom gradually: propose only, low-risk actions, actions inside caps, then running the week. Promotion is earned per class of action against approval history, and every level can be stepped back down.

Put a Digital FTE on thisEvery role proposes with evidence and reports to you.

Key takeaways

  • Autonomy is granted per class of action, never as a single global setting.
  • Level 0 means propose everything; it is where every role should start.
  • Promotion evidence is approval history: repeated approvals without amendment.
  • Every level must be reversible, or it is not a control.
  • Caps and escalation rules stay in force at every level above zero.

Why should autonomy be a ladder rather than a switch?

Autonomy should be a ladder because trust in an AI employee is specific rather than general. A system can be reliably excellent at drafting replies in a known style and simultaneously unready to decide a refund outside policy. A single on-off setting forces you to choose the lowest common denominator for everything.

A ladder lets the grant follow the evidence. You promote the classes of action where the record supports it and keep gates where it does not, which is how a manager already treats a new colleague: more rope on the work they have proved, none on the work they have not touched.

What are the four levels?

The ladder has four levels, each defined by what the AI employee may do without a human decision. On ftes.ai every Digital FTE begins at Level 0 regardless of role.

Level 0 is propose only: every publish, send, and spend waits in the approvals queue. Level 1 allows low-risk actions the manager has approved many times, such as routine refreshes or reply drafts. Level 2 allows the system to act inside explicit caps, such as pausing an underperforming ad set or publishing already-scheduled work. Level 3 lets it run the week, with the manager reviewing the brief, the exceptions, and the numbers rather than each item.

  • Level 0 — Proposes only: nothing leaves the company without a decision
  • Level 1 — Low-risk actions: routine patterns already approved repeatedly
  • Level 2 — Acts inside caps: bounded actions within written budget and rule limits
  • Level 3 — Runs the week: scheduled production, with review of brief and exceptions
LEVEL 0Proposes only
every action waits for a named approver
LEVEL 1Tight bounds
one class auto-approved within small limits
LEVEL 2Broader bounds
same class, wider limits, still logged
LEVEL 3Delegated class
runs unattended; every action in the trail
PROMOTED PER CLASS · FROM A CLEAN APPROVAL RECORDSTEPPED BACK IN ONE LOGGED CHANGE
Figure 1The four levels, climbed one class of action at a time from the approval record, and stepped back in one logged change.

What evidence justifies a promotion?

The evidence for promoting an AI employee is its approval history for that specific class of action. If the last several dozen instances of a class were approved without amendment, the system has demonstrably learned the standard for that class.

Two warning signs should block a promotion regardless of volume. First, approvals that came with edits, because an edited approval is a partial rejection. Second, an approval rate of essentially one hundred percent with no send-backs anywhere, which usually means the reviews were not real. Promotion should follow genuine scrutiny, not the absence of it.

What stays gated at every level?

Some classes of action should keep their gate permanently regardless of how high the ladder goes, because their cost is asymmetric. Anything affecting an individual's outcome, anything that is legally or reputationally irreversible, and anything novel belongs in this category.

A rejection, an offer, a refund outside policy, a public statement, a first payment to a new recipient, a new market or platform: these are not gated because the system is untrusted but because the downside is not symmetric with the upside. Keeping them gated is also what makes the promoted classes defensible.

How do you step autonomy back down?

Stepping autonomy down should be a single deliberate action, available at any time, and it should be used without drama. Circumstances change: a new product launch, a regulatory shift, a period of unusual sensitivity, or simply a run of send-backs that suggests the standard has moved.

Treat a step-down as a normal management action rather than an admission of failure. The willingness to reverse is what makes the original grant reasonable; a level you cannot revoke is not a delegation, it is an abdication.

How does the ladder change the manager's job?

The autonomy ladder changes a manager's work from reviewing items to curating classes. At Level 0 the job is deciding on individual proposals. By Level 2 or 3 the job is deciding which classes deserve promotion, where the caps should sit, and what the exceptions are telling you.

That shift is the actual productivity gain, and it is measurable in a calendar. Time spent approving individual items falls; time spent setting targets and adjusting limits rises; total management time falls while the function's throughput increases.

What does the ladder look like for a specific role?

For a Finance Controller Digital FTE the ladder is applied per class of action in its Setup tab. Exact-match reconciliations, where the bank line and the invoice correspond precisely, are usually the first class promoted, because the pattern is unambiguous and the approval record is long within weeks. Partial matches, write-offs, adjustments, and anything above a threshold the manager sets stay gated, because judgement is involved. Posting entries never leaves the manager's approval unless a narrow class is deliberately promoted.

For a Customer Support Digital FTE the same ladder is applied per ticket category: shipping-status and password-reset replies from a well-maintained knowledge base earn auto-send first, while refunds, account access, complaints, and distress sit on the always-escalate list and never climb. The ladder is the same mechanism in both roles, applied to different classes, which is why a manager who has climbed it once with one role can climb it with the next from the first morning.

Frequently asked questions

What is the autonomy ladder?

It is a four-level model for granting an AI employee freedom gradually: propose only, low-risk actions, actions inside caps, then running the week under review.

Where should an AI employee start?

At Level 0, proposing everything. Every role should start there regardless of vendor claims, because trust is specific to a class of action and has to be demonstrated.

What evidence justifies more autonomy?

Repeated approvals of that specific class without amendment. Edited approvals are partial rejections, and a perfect approval rate with no send-backs suggests reviews are not real.

What should stay gated permanently?

Anything affecting an individual's outcome, anything irreversible, and anything novel: rejections, offers, public statements, first payments, and new platforms.

Can autonomy be reduced again?

Yes, and it should be a single deliberate action available at any time. A level you cannot revoke is not a delegation.

  • Autonomy
  • AI governance
  • Approval gates
  • Management
  • Digital FTE
Further reading

Keep going.

Related guides on hiring and managing Digital FTEs.

Ready to hire your first Digital FTE?

Pick a role, connect the tools it needs, and review its first proposals this week.