When a human should take over: speed is table stakes, the handoff is the bottleneck

A human should take over when the agent cannot answer with a fact, when the prospect asks for a price, a contract or an exception, and when the person is ready to talk. Speed is table stakes: on our own LinkedIn outbound, cutting the agent's reply time from about 19 to about 2 minutes did not raise meetings per conversation. What loses meetings is a handoff that nobody sees.

The AI SDR market competes on speed: "replies in seconds". Speed does matter, and we will show where. But once a reply arrives in a couple of minutes, the next bottleneck is not the agent at all. It is the moment a conversation needs a person, and whether that person knows.

Why is speed only table stakes?

Because people write when it suits them, not you. On our own outbound, 48% of prospect messages arrived at night or on weekends. A human team answers those on Monday. Our agents answered 74% of them before the next morning, with a median reply of 2.3 minutes since May 2026 and 90% inside an hour. That is a real advantage over an inbox that sleeps, and it keeps conversations alive.

But the reply time fell from about 19 minutes to about 2 over the year, and meetings per conversation did not rise with it. Once nobody waits overnight, shaving more minutes buys nothing. The list and the offer were the bigger levers, as the case study shows.

Where do meetings get lost?

In the gap after the agent has done its part. Four patterns come up.

1. The unseen flag. The agent correctly decides it cannot answer, and the conversation waits in a list nobody checks.
2. The cold handoff. A person picks up a thread with no idea why the agent stopped, and asks the question the prospect already answered.
3. The slow draft. In approval mode, messages wait for a person. If a draft sits for a day, the speed advantage is gone.
4. The missing person. The prospect books a call and nobody shows up on time. In our own data, one prospect told us that a late host on a 15 minute call is a reason to walk away.

Notice that none of these is the agent's fault. They are workflow faults around the agent.

When should a person take over?

Draw the line in advance. A workable list:

• A question with no fact behind it. If the agent does not know, it must say so and hand over. An agent that invents an answer is worse than one that waits. In Sellinger the agent stops on its own when it cannot answer, for example a price it has no answer for.
• Price, terms and exceptions. A discount, a contract change, a custom scope. These are decisions, not facts.
• A signal of a big deal. In the Shelly programme the agent read every installer answer for project size and handed the large ones to Shelly's team; 66 converted threads went to a person.
• A complaint or a refusal that needs care. A real "please stop" ends the thread and goes on the suppression list. A vague or unclear phrase should stop the agent and ask a person to decide.
• A regulated or sensitive topic. Anything with legal, medical or financial weight.
• A request for a person. If they ask for a human, give them one.
• Readiness. When someone says "send me the link", the agent can book it. When they say "let's talk about rollout", a person should.

Some of these are rules, some are judgment. The rules should be written down. The judgment should be reviewed weekly, using the threads the agent flagged.

What controls should you look for?

• Approval mode. Every message waits for a person to approve it before it sends. Use it for new campaigns until the messages read the way you would write them. Sellinger campaigns can run in approval mode or send automatically.
• Takeover. A person can take over any conversation at any time, and the agent stops in that thread. See take over a conversation.
• A named reason. When a lead is sent to a person, the reason is recorded. Automatic replies from helpdesks, no reply senders and survey emails are marked as machine mail, never counted as a human reply, and the lead goes to a person with a named reason.
• Suppression checks. Before every send the platform checks the team's suppression list and refuses the send if the contact is on it.
• Notifications that reach someone. A flag nobody sees is the same as no flag.

The rules for approval are covered in human in the loop and the steps in approve messages.

What does a handoff gap look like in another business?

Hotel Pirin, a hotel owned by a Sellinger co-founder, gives a clear picture. Its reservations agent answered guests with a median first reply of 41 seconds over 54 guest inquiries. Where the agent answered first (53 threads), the median first reply was 0.7 minutes. Where staff answered first (10 threads), it was 159 minutes. That is not a like for like race, because staff answered first only where the agent had not, but it shows how wide the gap is when a conversation waits for a person.

The handoff problem showed up in the quotes. Before a rule told the agent to quote as soon as it had the dates and the party, families with two or more children were handed to "a colleague who will send an offer", and those offers took from 3 hours to 7 days. After the rule, the share of the agent's replies that carried a price went from 50% to 92%. The fix was not a faster agent. It was a better line between what the agent may do and what a person must do. Groups, agencies and Booking.com messages still went to staff by design.

A made up example: a cold and a warm handoff

The two notes below are invented to show the difference.

Cold handoff: "Lead needs attention." The person opens the thread, reads twelve messages and writes: "Thanks for your interest, what is your budget?" The prospect had already said the budget in message four.

Warm handoff: "Stopped because the prospect asked for a custom scope for 40 users. They said they need it live by the first of the month. Budget stated in message four. Suggested next step: reply with a call time and confirm the date." The person answers in one message and books the call.

The agent does the same work in both. The second note is what turns a flag into a meeting.

What service levels should you set?

Set targets before launch and review them weekly. These are suggested targets for you to adapt, not Sellinger settings.

The agent already stops on its own at these points, so the targets are about people. Pick a named person for each row, and a backup for when they are away.

How do you run a weekly review?

1. Open the threads flagged for a person in the last week.
2. For each, note the time from flag to first human action.
3. Mark the ones where the person had to ask something the prospect had already said.
4. Fix the cause: a missing fact in the knowledge, a rule that stopped the agent too early, or a notification that reached nobody.
5. Count meetings held that came from handed over threads.

What should the person receive?

A good handoff has four things:

1. The whole conversation, across channels, in one place.
2. Why the agent stopped, in one sentence.
3. What the prospect asked for, in their words.
4. A suggested next step the person can accept or change.

Without the second item people repeat questions and prospects feel processed. With it, a person can answer in a minute.

How do you measure the handoff?

Three numbers, reviewed weekly:

• Time from flag to first human action. The one that decides whether the speed advantage survives.
• Share of flagged threads that end with a human message. If it is low, flags are being ignored.
• Meetings held, not booked. On our account 351 meetings were booked through the agents and 33 were confirmed held. A booking is a promise. The handoff includes the call itself. See count meetings honestly.

What is the shape to aim for?

On our own LinkedIn outbound, 93% of replied conversations since May 2026 ran with no human message. That is the agent carrying the routine. The other 7% is where judgment lives, and where the quality of the handoff decides the result. Design for both: let the agent run the routine at any hour, and make the exceptions impossible to miss.