• Explainer
  • Governance
  • 5 min read

AI Governance for Autonomous Agents

Five controls, each with a mechanism behind it. Governance you can point at rather than describe.

Author
ftes.ai Editorial
ftes.ai Team

The short answer

AI governance for autonomous agents rests on five controls: a written remit bounding what the agent owns and escalates; an approval gate requiring human approval per action until earned otherwise; an autonomy ladder granting autonomy per class from the approval record; an audit trail of every action; and revocation from one place at any time.

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

Key takeaways

  • Remit: what the agent owns, escalates, and never does, written down.
  • Approval gate: human approval per action until earned otherwise.
  • Autonomy ladder: per-class promotion from the approval record, reversible.
  • Audit trail: every action, who, what, when, with evidence.
  • Revocation: stop, step back, or disconnect from one place.

What does governance need to answer?

Governance for an autonomous agent needs to answer four questions: what is it allowed to do, who approves what it does, how do we know what it did, and how do we stop it. A framework that answers all four with a mechanism, rather than a policy document, is one that holds under pressure.

The five controls below are the ftes.ai answer. Each is a feature of the platform, visible in the workspace, rather than a principle to be applied by hand.

What is the remit?

The remit is the written scope of a Digital FTE: what it owns, what it escalates, and what it never does. It is set at hiring in the first-run wizard and refined from the approval record. A Customer Support FTE's remit names the categories it resolves, the fixed list it always escalates, and the tone it uses.

The remit is the first control because it bounds everything else. An agent with no written remit cannot be governed, because there is nothing to hold it to.

How do the approval gate and autonomy ladder work together?

The approval gate requires a named person to approve each action before it happens. Every Digital FTE starts at Level 0, where every action is gated. The autonomy ladder lets a manager promote a specific class of action, such as exact-match reconciliations or first-touch outreach to cold prospects, to auto-approve, based on a long run of unchanged approvals for that class.

Promotion is per class, never blanket, and reversible with one change. Some classes never qualify, such as anything affecting a person's outcome. Together the two controls mean autonomy is earned from evidence and can be withdrawn at any time.

  • Level 0: everything gated, the starting point for every FTE
  • Promotion: per class, from the approval record, by a named person
  • Step back: one change, logged, effective immediately
  • Never: a fixed list that cannot be promoted
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 gate and the ladder together: everything starts at Level 0, promotion is per class from evidence, and any class can be stepped back in one change.

What does the audit trail record?

The audit trail records every proposal, approval, send-back with reason, publish, send, spend, autonomy change, and credential change, with who did it and when, and with the evidence the proposal carried. Per-run audit trails show what a Digital FTE did in each piece of work.

The trail is what makes the other controls demonstrable. A remit that is not enforced, an approval that is not recorded, and an autonomy grant that nobody can trace are governance in name only; the trail turns them into governance in fact.

How does revocation work?

Revocation is the ability to stop. A manager can pause a Digital FTE, step any class back to gated, disconnect any integration from the workspace, and revoke credentials from one place, at any time, with each action logged. The wallet's hard stop is a form of revocation too: an empty wallet pauses everything.

The test of a governance framework is what happens on a bad day. On ftes.ai the answer is: step back, disconnect, read the trail, and fix the remit, all from the workspace, all recorded. The vendor evaluation guide lists the questions to ask any vendor about the equivalent.

How do you audit the framework once a quarter?

Audit the framework quarterly by sampling and tracing. Pick a handful of actions per role from the quarter, including at least one auto-approved action from a promoted class and one escalation, and trace each end to end in the audit trail: the proposal, the evidence it carried, the decision and reason or the grant it was auto-approved under, the timing, and the effect. Confirm that every action on the always-escalate and never-automated lists in the sample reached a person.

Then read the management record: which classes were promoted and stepped back, by whom, on what record; which connections were added or removed; which budgets changed. Compare the remit as written with the send-back reasons of the quarter, because repeated reasons show where the remit and the work have drifted apart. Record the audit and any changes. The quarterly confirmation of the Brand Brain and the review of connections fit naturally into the same session, and an hour per role is usually enough, because the trail was built as the work happened and nothing has to be reconstructed.

Frequently asked questions

What does AI governance for autonomous agents involve?

Five controls: a written remit, an approval gate per action, an autonomy ladder that promotes per class from evidence, an audit trail of everything, and revocation from one place.

What is a remit?

The written scope of what an agent owns, what it escalates, and what it never does. Without one, an agent cannot be governed.

How is autonomy granted?

Per class of action, by a named person, from a long run of unchanged approvals for that class. It is reversible with one change and some classes never qualify.

What does the audit trail contain?

Every proposal, approval, send-back, publish, spend, autonomy change, and credential change, with who, when, and the evidence.

How do I stop an agent?

Pause it, step any class back, disconnect integrations, or revoke credentials from the workspace at any time. An empty wallet also pauses everything.

  • Governance
  • AI governance
  • Autonomous agents
  • Framework
  • Controls
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.