Approval mode: why an AI receptionist should start supervised

Approval mode means the AI drafts every reply and a person approves, edits or discards it before it sends. Running the first week or two this way is how you find the facts that are missing, the phrasing that sounds wrong, and the questions that should be handoffs, before any customer sees a mistake. Then you release categories one at a time, not all at once.

What approval mode actually is

Every inbound message still gets a draft within seconds. The difference is that the draft waits in the inbox for a person. Approve sends it as written. Edit sends your version and records the change. Discard drops it and you reply yourself. The customer sees a reply a few minutes later instead of a few seconds later, which in week one is a fine trade.

What you learn in the first week

Missing facts. Every draft that says "I am not sure" about something you know is a knowledge entry to add. Most businesses find a dozen in the first two days.

Wrong facts. A price that changed in March and nobody updated. Approval mode catches it before a customer books at the old price.

Tone. Read twenty drafts and you know whether the tone paragraph works. Too many exclamation marks, too formal, too eager to apologise. Fix the paragraph, not the drafts.

Handoffs you had not thought of. A question that a person should always take, which nobody put on the list. Add the trigger.

The edits themselves. Every edit is a small signal. If you keep changing "we would be happy to" into "we can", that is a tone rule.

Releasing categories, not the whole thing

After a week of clean drafts, release the categories where the edit rate is near zero: hours, parking, booking confirmations, reminders. Keep the rest supervised: pricing questions, anything from a new customer, anything with a complaint in it. Each category is a decision you can make on evidence rather than on nerve.

The categories to keep supervised longer

New customers, because first impressions. Pricing, because a wrong number is a promise. Anything the AI flagged as uncertain. Anything from a customer with an open complaint. Some businesses keep pricing supervised permanently and lose nothing by it, since the draft is still instant and approving takes two seconds.

What to watch after release

The gaps log for questions it could not answer. The share of conversations handed off. Any edit a team member makes to a sent reply. A monthly fifteen-minute review is enough once the first month is behind you.

Frequently asked questions

Does approval mode slow replies down too much?

In week one, a few minutes instead of a few seconds. Customers who were getting voicemail before do not notice. From week two, released categories are instant again.

Who should be approving?

Whoever answers the phone today. They know the right answers and they will spot the wrong tone fastest. It should not be the owner alone unless the owner is the front desk.

Can I go back to supervised later?

Yes, per category. A price change, a new service or a new location is a good reason to supervise the affected category for a few days.

What Frontiva does here

In Frontiva, approval is the default posture: the agent writes each reply as an ordinary outbound message and it waits in the unified inbox for a teammate to approve or edit, and a queued draft is not counted as a conversation the AI handled. Autonomy is set per agent, so you widen it when you have watched it work, and every question the agent could not answer lands, counted, in the knowledge gaps log.

Approval mode is one of the settings on the AI agents page. For what the thing is before you decide how to supervise it, see what is an AI receptionist.