Human in the loop: when to require approval before a message sends

Require approval for new campaigns, new audiences, any change to the offer or rules, and any reply that touches price or terms. Switch to automatic for routine conversations once you have read enough drafts to trust the pattern. In Sellinger a campaign can run in approval mode, where every message waits for a person, or send automatically, and a person can take over any conversation.

"Human in the loop" is often sold as a safety feature and treated as a speed tax. Both views miss the point. Approval is a way to teach the agent what you want, and then to stop spending time on it. This post sets out when each mode fits, and how to keep reviews quick.

What are the modes?

Two settings control what the agent does without asking.

Approval mode. Every message waits in a queue. A person reads it, edits if needed and approves, and then it sends.

Automatic. The agent sends by itself, within the daily limits and quiet hours of each account.

On the channel account, an Auto Approve Actions switch controls the second mode. Leave it off to review messages, or switch it on to let the agent act without approval. The steps are in approve messages. Separately, a person can take over any conversation at any time (take over a conversation).

When should you require approval?

• A new campaign. Read every first message until you trust it.
• A new audience. The same agent writes differently for engineers and for founders. Check the first twenty drafts.
• A new offer or new rules. After any change, the old drafts tell you nothing.
• Anything about price or terms. A discount, a contract change or a scope promise.
• A new language. Have a native speaker read the first drafts (write in the local language).
• High value accounts. A message to a strategic account is worth a minute of review.
• Regulated topics. Anything with legal, health or financial weight.

When can you go automatic?

When four things are true:

1. You have read at least a few dozen drafts for this campaign and made fewer and fewer corrections.
2. The agent's rules are written down and you know what they say.
3. A person reads flagged conversations daily.
4. You have limits in place: a daily cap, a quiet window and a stop on opt outs.

Even then, keep approval on for the pieces that carry the most risk. Many teams go automatic for replies to simple questions and keep approval for first messages, or the other way round.

What does the data say about running without people?

On our own LinkedIn outbound, 93% of replied conversations since May 2026 ran with no human message. In the Lamothermic study it was 94%. Neither number means nobody was involved. A person set the audience, the offer and the rules, read the flagged threads and picked up the meetings. The saving is in the routine conversation, which is most of the volume.

It also took a long time to get there. Both accounts started with a person reading what the agent wrote, correcting the rules and adding guardrails. Lamothermic's owner added lines about what the foundry does not cast, as the campaign ran. That is approval doing its job: finding the rules you did not know you needed.

What is the cost of approval?

Time, and speed. On our own account the median reply was 2.3 minutes when the agent ran on its own, and 48% of prospect messages arrived at night or on weekends. A draft that waits until Monday gives away that advantage. If you use approval, someone has to review often, ideally within the hour on workdays, and have a rule for what happens overnight. A common compromise is automatic for acknowledgements and simple answers, and approval for anything with a commitment.

The post on speed and the handoff covers why a flag that nobody sees is as bad as no flag.

How do you review quickly?

• Batch it. Review once or twice a day, not message by message.
• Review in order of risk. First messages of a new campaign, then replies with a price or a date, then the rest.
• Fix the rule, not the message. If you correct the same thing three times, change the instruction so you never see it again.
• Use shortcuts. Through the MCP connector you can work the approval queue from Claude or ChatGPT; see the approval queue triage recipe.
• Track corrections. Count how many drafts you changed per twenty. When it drops near zero, you are ready to loosen.

What should the agent stop on?

Independent of mode, some things should always stop the agent:

• A question it has no fact for. In Sellinger the agent stops on its own when it cannot answer, such as a price.
• An opt out. Replies such as STOP or unsubscribe put the contact on the suppression list and end the thread. A vague or unclear phrase never writes the permanent record: the agent stops and a person decides.
• A person taking over.
• Machine mail, such as automatic replies from helpdesks. These are marked, never counted as a human reply, and sent to a person with a reason.

What does a made up approval queue look like?

An illustration with invented drafts, to show how a morning review works. A team starts a new campaign for operations managers. Twenty first messages wait in the queue.

• 14 approved as they are. The observation is true and the ask is small.
• 4 edited. The agent wrote "you scale" where the profile says "you run". The editor fixes the wording and writes a rule: use the person's own verb.
• 2 rejected. Both mention a product the person does not use. The editor adds an exclusion for that segment.

Two things happened: the campaign got safer, and the rules got better. The next twenty drafts need fewer edits. When corrections fall close to zero for a full batch, the team moves first messages to automatic and keeps approval on for replies that touch price.

What are the failure modes of approval?

Rubber stamping. A reviewer who approves everything in one click is not reviewing. Count the edits per twenty drafts. If it is always zero in week one, the reviewer is not reading.

A queue nobody owns. If three people could approve and none does, messages age. Name one owner and one backup.

Approving the wrong thing. Reviewing the wording while the audience is wrong. Look at who the message goes to before you look at how it reads.

Never leaving approval. Teams that keep approval on for everything give away speed. On our own account the median reply was 2.3 minutes when the agent ran on its own.

Who should be the reviewer?

The person who knows the buyer and the offer, which is often the founder or the head of sales and rarely the most junior person. In an agency, the account manager for that client. For a language you do not speak, a native speaker (write in the local language). The reviewer needs authority to change the rules, not only to approve single messages.

What happens when a person takes over?

A person can take over any conversation. The agent stops writing in that thread, and the person sends the messages. Use it when a lead asks for a human, when a deal is large, or when the conversation turns to terms. The steps are in take over a conversation. When the person is done, the thread can go back to the agent or stay with the person.

How do you know the setting is right?

Three signals. Corrections per twenty drafts are near zero. Flagged conversations are read the same day. And no reply has gone out that you would not have signed. If any of the three fails, move one step back toward approval for that campaign only.

A simple rollout

1. Week 1: approval on for everything. Read every draft.
2. Week 2: approval on for first messages and anything with a price. Replies to simple questions go automatic.
3. Week 3 onward: automatic by default, with approval for new campaigns and changes.
4. Every week: read ten flagged conversations and adjust the rules.

The goal is not zero people. It is people where judgment matters, and nowhere else.