Autonomous vs approval mode: how to decide for each AI agent
Five pieces of evidence that should move an agent, the one measurement people use that is not evidence, why splitting by job beats choosing a level, and what going back does not mean.
Sohail Safdar · · 5 min read
Every AI agent should start supervised. The interesting question is not whether, it is what evidence should let it stop being.
What each mode actually trades
Approval mode trades speed for certainty. The agent drafts and a person sends, so nothing reaches a customer that nobody read. The cost is real: during opening hours that might be minutes, and overnight it is until somebody arrives.
Autonomous mode trades certainty for speed. Replies go in seconds at any hour, and a mistake reaches the customer before anybody sees it, which is why the guardrails underneath matter more rather than less.
What does not change between them: escalation keywords, sensitive-topic refusal and prompt-injection handling all run in both. Autonomy removes the human review of ordinary replies. It should not remove anything else, and if a vendor's autonomous mode skips a check that approval mode ran, that is a different product rather than a different setting.
The five pieces of evidence
All five, not three of five.
1. A near-zero edit rate, over a run. Almost every draft approved as written, across twenty consecutive drafts rather than twenty you remember.
2. Few remaining knowledge gaps. The questions it could not answer are rare, and the frequent ones now have approved answers. A gaps list that is still growing means the business has not finished teaching it.
3. Escalation you have tested yourself. Not the vendor's demo. Send your own trade's emergency words, a complaint phrased the way an annoyed customer phrases one, and a request for a person worded as "is there anyone there". Then confirm somebody actually saw each one, which is a separate check from whether it handed over.
4. A readiness test passed, including its emergency gate. What an AI readiness test should cover sets out the five scenarios and why the emergency check must not be part of the average.
5. A named person who owns handovers, including after hours. An agent that escalates correctly into a queue nobody owns has escalated to nothing.
The measurement people use that is not evidence
Elapsed time. "It has been two weeks, so we are fine."
Two weeks of three conversations a day is eighteen data points about a quiet practice. Three days of thirty conversations is ninety. The calendar version of this rule is how a seasonal or small business ends up trusting an agent on a handful of exchanges, all of which happened to be about opening hours.
Count drafts, not days. And count a run of consecutive drafts rather than a percentage, because a 95% approval rate still means one in twenty went out wrong, and you want to know whether the failures are clustered in one topic.
Split by job instead of choosing a level
This is the move that gets missed, and it is usually better than either mode applied to everything.
Some conversations are safe and some are not, and the difference is topical rather than gradual. So use separate agents with narrow jobs:
- One autonomous agent for hours, directions, parking, and the booking link. A wrong answer here is inconvenient. The volume is high and the risk is low.
- One approval-mode agent for pricing, new-customer enquiries and anything a salesperson would handle. The volume is lower and a mistake is expensive.
Narrow agents are easier to trust, and not because they are better: because what they can get wrong is bounded by what they are allowed to discuss. That is a structural limit rather than a behavioural hope, and structural limits are the kind you can rely on at 9pm on a Saturday.
When to stay supervised, or go back
Keep approval mode on when:
- The agent handles high-stakes conversations: pricing for large jobs, clinical or legal intake.
- Your knowledge changes often and has not settled.
- You have just added a service, a location or a channel.
- A mistake happened recently and the cause is not fixed yet.
And go back when any of those becomes true again. A price change, a new policy, a new member of staff whose services differ, or simply a run of edits.
Going back is not a failure, it is the control working. The failure mode to avoid is a business that treats the move to autonomous as a milestone it has passed, and therefore leaves an agent unsupervised through exactly the weeks when its knowledge is least stable. Treat it as a dial you turn both ways.
One honest note about the comparison
A practice reading this will notice that approval mode is slower, and speed to lead matters. Both are true.
The resolution is not to rush autonomy. It is to be honest about what approval mode costs overnight and to say so to the customer: an after-hours holding message that says when somebody will reply is a promise you can keep, and it beats both silence and an autonomous reply you have not earned the right to trust yet.
Plenty of practices run permanently in approval mode on their high-stakes agent and are entirely right to. That is a legitimate place to stop, not a stage you failed to leave.
In Frontiva
Each AI agent has an action mode: autonomous, approval, or off, set per agent rather than per business, which is what makes the split-by-job approach above available rather than theoretical.
Approval is the default posture, and a queued draft is deliberately not counted as a conversation the AI handled, because counting it would make a supervised agent look busier than it is and flatter exactly the number you are using to decide.
Escalation keywords and sensitive topics hand off in either mode. The sensitive-topic and prompt-injection checks are deterministic pattern checks that run before the answering path, so they are not a thing autonomy can skip past.
Unsupervised answering is locked until the readiness suite passes, and its two emergency checks are required rather than averaged, so an agent that cannot get an urgent message to a person cannot be switched to autonomous however well it scored on prices. Approval mode, why start supervised covers the first week, and your AI receptionist's first week is the day-by-day version.