How an AI follow-up call works, and when it hands off to you

An AI follow-up call is a loop that runs on every turn of the conversation: hear the person, understand what they meant, decide what to say, say it, and listen again. The whole point of running that loop is to get a lead to the moment where they want to talk to a real agent, and then hand the call over with context.

This post explains each part of that loop in plain terms, where it tends to break, and why the conversation that actually matters should belong to you.

What is actually happening on an AI call?

There are four stages, and they repeat for every turn in the conversation.

1. Speech in. The caller’s audio streams in and a speech recognition model turns it into text as they talk, not after they finish.
2. Understanding. A language model reads that text alongside what it already knows: the lead’s name, the listing they asked about, the script your team approved, and what has been said so far on this call.
3. Deciding what to say. The model picks the next move: answer the question, ask one, confirm a time, or decide this is the moment to bring in a person.
4. Speech out. A voice model turns the reply into audio and starts playing it, ideally before the full sentence has even been generated.

None of this is a recording. A recording plays the same thing whether the lead says “yes,” “who is this?” or “I already bought.” A conversational call has to respond to what was actually said.

Why does latency matter so much?

Because people are extremely sensitive to timing in conversation. A study of turn-taking across 10 languages, published in PNAS in 2009 by Tanya Stivers and colleagues, found that all of them showed “a general avoidance of overlapping talk and a minimization of silence between conversational turns,” with differences between languages in the average gap staying within a range of 250 ms of the cross-language mean.

In other words, the timing differences that make one conversational style feel different from another are measured in fractions of a second. An AI call that pauses noticeably before every answer feels wrong, even if the words are right. The lead starts saying “hello?” into the silence, and now both sides are talking at once.

So the engineering work is mostly about overlap. Transcription runs while the person is still speaking. The model starts forming a reply before the transcript is final. The voice starts speaking the first words while the rest is still being generated. Every stage that can stream, streams.

What happens when the lead interrupts?

The call has to stop talking. People interrupt constantly on the phone: to correct a detail, to say “sorry, I’m driving,” to jump ahead. A system that keeps reading its script over the top of a person is the fastest way to make them hang up.

Handling interruptions well means the system is listening while it speaks. When the caller starts talking, playback stops, the new words are transcribed, and the model decides whether the interruption changes the plan. “Wait, is this about the house on Maple?” needs an answer. A cough does not.

This is harder than it sounds, because background noise, a TV or a second voice in the car can all look like an interruption. Tuning that sensitivity is a real tradeoff: too eager and the AI stops mid-sentence for nothing, too relaxed and it talks over people.

How does it know it reached voicemail?

In the first seconds after the call connects, the system classifies what it hears: a person saying “hello,” or a greeting followed by a beep. The two sound different in rhythm and length, and a live person usually responds to silence while a recorded greeting just keeps going.

If it is voicemail, the right behavior is to wait for the beep and leave a short message that names who is calling and why. If it is a person, the AI needs to start the conversation immediately, because that is where the latency problem above starts to bite.

When should the AI stop and transfer the call?

This is the most important design decision, and the answer is: earlier than you might think. The job of an AI follow-up call is not to replace the agent. It is to reach the lead quickly, keep the conversation warm, answer the simple questions, and find out whether they are ready.

Good reasons to hand off include:

• The lead wants to see a home or talk about a specific offer.
• A seller wants to talk about pricing or timing in real terms.
• The lead asks for a person, directly or indirectly.
• The AI is not confident it understood, or the question is outside what it was given.
• The conversation turns emotional: a divorce, a death in the family, a job loss.

A live transfer means the AI keeps the lead on the line while it rings you. When you pick up, you get a short brief first (who this is, what they asked about, what they want) and then you are connected. The lead should never have to repeat the first two minutes of the call to you.

If you do not pick up, the AI should tell the lead honestly that you will call back, and make sure that callback actually gets scheduled.

What are AI calls bad at?

Being honest about this is the only way to deploy it well.

• Judgment. A model can follow a script and answer factual questions. It is not good at knowing when a “maybe” means “no,” or when a lead needs to be pushed and when they need space.
• Local knowledge it wasn’t given. Which side of the street floods, which school boundary just changed, what the listing agent is really like to work with. If it is not in the data, the AI does not know it, and it should say so rather than guess.
• Emotion and trust. People buying or selling a home are often making one of the largest financial decisions of their lives, sometimes under stress. That conversation deserves a person.
• Negotiation. Anything that involves price, terms or strategy belongs to you.
• Edge cases in audio. Bad connections, heavy background noise and people talking over each other still cause misunderstandings.

The right mental model is a very fast, very consistent assistant who handles the first contact and the repetitive follow-up, and knows when to get you.

What rules apply to AI calls?

The same ones that apply to other artificial voice calls. In February 2024, the FCC unanimously adopted a Declaratory Ruling recognizing that calls made with AI-generated voices are “artificial” under the Telephone Consumer Protection Act. The FCC’s release notes that its rules require telemarketers to obtain prior express written consent before robocalling, and that the ruling holds AI-generated voices to those same standards.

The general federal rules also still apply, including no telephone solicitations to residential subscribers before 8 a.m. or after 9 p.m. local time at the called party’s location, and a requirement that artificial or prerecorded voice messages state clearly at the beginning who is responsible for the call.

How Forward Flow’s calls work

Forward Flow’s calls are real conversations, not recordings. The AI listens, answers questions, adapts, and introduces itself as part of your team. It can qualify a lead and book an appointment on your team’s calendar. Voicemail is detected and a voicemail can be left instead.

When a lead is ready to talk, the call transfers to you live with a short brief: who they are and what they want. Scripts, cadence, hours and hand-off rules are your team’s to approve and change, and calls respect quiet hours. For what the whole first minute looks like around the call, see The first 60 seconds after a lead fills out a form.

The honest way to evaluate an AI call is to take one. Go to /experience, play a buyer or a seller, and let it call you. Then decide whether it sounds like something you would want representing your team.

This is general information, not legal advice. Talk to a lawyer about your own situation.