Autonomous vs approval mode: how to decide for each AI agent

In approval mode, an AI agent drafts each reply and a person approves or edits it before it sends. In autonomous mode, the agent sends on its own and hands off when an escalation keyword or sensitive topic appears. Start every agent in approval mode, and move one to autonomous only when a week or two of drafts needed almost no edits, the questions it could not answer are few, and a person owns its handoffs.

What each mode trades

Approval mode trades speed for certainty. Customers wait for a person to approve, which during office hours might be minutes and overnight might be hours. Nothing goes out that a person did not see.

Autonomous mode trades certainty for speed. Replies go out in seconds, at any hour. Mistakes reach customers before anyone sees them, which is why the guardrails and escalation rules matter more.

The evidence to move an agent to autonomous

1. Edit rate near zero. Over the last week or two, almost every draft was approved as written.
2. Few knowledge gaps. The questions it could not answer are rare, and the common ones now have approved answers.
3. Escalation that works. Test messages with your escalation keywords, and a request for a person, all hand off.
4. A readiness check passed. The scenarios you care about (an unknown question, a prompt-injection attempt, a sensitive question, a real price question) behave correctly.
5. An owner for handoffs. Someone watches what it hands over, including after hours.

When to keep approval mode

• The agent handles high-stakes conversations: pricing for large jobs, legal or medical intake.
• Your knowledge changes often and is not yet stable.
• You are adding a new service, location or channel.
• A mistake has happened recently and the cause is not fixed.

Splitting by job

If some conversations are safe and others are not, use separate agents with narrow goals: one autonomous agent for hours, directions and booking links, one approval-mode agent for pricing and new customer enquiries. Narrow agents are easier to trust.

Going back

Moving an agent back to approval mode is not a failure. A price change, a new policy, or a run of edits is a good reason to supervise for a few days.

Frequently asked questions

How long should approval mode last?

Usually one to two weeks, measured by edit rate rather than by the calendar.

What if nobody approves drafts at night?

Then customers wait until morning. Decide whether that is acceptable for that agent's job, and say in your after-hours message when a reply will come.

Does autonomous mode skip the guardrails?

It should not. Guardrails and escalation run in both modes; autonomy only removes the human review of ordinary replies.

What Frontiva does here

Each of Frontiva's AI agents has an action mode (autonomous, approval, or off), approval is the default posture, and a queued draft is not counted as a conversation the AI handled. Escalation keywords and sensitive topics hand off in either mode, and the readiness suite scores an agent before and after changes. See approval mode, why start supervised.