How to Write an AI Call Agent Script: Step-by-Step Guide

A 2025 study in Scientific Reports asked people to tell AI-generated voices apart from real human ones. They were not very good at it, failing to flag the AI voice more than a quarter of the time (the researchers concluded people are "poorly equipped" to spot a voice clone). That is the reality your AI call agent works in. The prospect on the other end usually can't hear the difference, which is exactly why telling them the caller is AI matters more, not less.

Here is the trap most teams walk into. They take a script written for a human SDR, drop it into an AI agent and expect it to perform. It doesn't. A human script is a suggestion a rep improvises around. An AI script is a decision tree the agent has to follow. Same goal, completely different build.

This guide covers the six components every AI call agent script needs, plus a step-by-step process for writing your first one. The goal is a script that uses what AI does well, sounds like a person wrote it and knows the moment to hand the conversation to a human.

Key takeaways

• Disclose that the caller is AI, up front. It is the ethical default, and when prospects can't hear the difference, it is also what keeps trust intact.
• The first 10 to 15 seconds decide the call. The opener has to acknowledge the interruption and earn permission to keep going.
• Write branches, not a monologue. An AI agent needs a defined response for every way a prospect can react, not one suggested path.
• Plan objection trees, not single rebuttals. Map two or three responses for each common objection.
• Build compliance into the script, not on top of it. TCPA disclosures and human-handoff triggers are script elements, not afterthoughts.

Why AI call scripts work differently

Most cold-calling advice assumes a human is holding the phone. The script runs in a straight line, opener to pitch to objection to close, and the rep fills the gaps with intuition, reading tone and improvising when the conversation wanders. That works because people are good at wandering back.

An AI agent is good at other things. It never forgets a follow-up, it sounds the same on call 200 as it did on call 1, and it can work a warm call list while your reps sit in back-to-back meetings. What it can't do is improvise out of a branch you never wrote. Say "I'm not interested" and the agent needs an instruction. Say "Send me an email instead" and that's a different instruction. "Who is this?" is a third. A linear script breaks the moment a real person goes off-script, which is to say immediately. The ultimate guide to AI cold calling covers the broader strategy; this piece is about the script itself.

The modern tech does buy you one thing, though. The dialogue no longer has to sound robotic. The conversational AI behind a good call agent, like the ElevenLabs technology that powers AvairAI's AI Call Agent, hears tone and intent, filters background noise and responds in well under a second. It delivers whatever you write in a natural voice. So your job is to write lines a person would actually say, then wire the branches underneath them. (And prospects are far more receptive to this than the "nobody talks to a robot" assumption suggests, as the real hang-up data on AI calls shows.)

The six components every AI call script needs

Miss one of these and the call either underperforms or creates compliance risk.

1. An opener that earns the next 30 seconds

Gong analyzed more than 300 million cold calls and found the first few words make or break the call: its strongest openers advanced conversations several times more often than the weakest. Your opener has three jobs and about 10 to 15 seconds to do them. Acknowledge that you're interrupting, say why you're calling in terms of their world and ask permission to continue. It should also name that the caller is AI, early, so nobody feels misled later. Done well, that kind of transparency builds trust rather than eroding it.

What the opener is not is an introduction to your company. Nobody has agreed to hear about you yet.

A template that does all of this:

"Hi Caleb, this is Olivia, an AI agent calling on behalf of [Company]. I know I'm catching you out of the blue. I'm reaching out because we've helped [similar teams] with [specific problem]. Do you have 30 seconds for me to tell you why?"

2. A value proposition you can say in one breath

Once you have permission, you have maybe 20 seconds. Name the problem in language specific to their role, name the outcome they'd get and, if you have it, name who else has gotten it. Then stop.

"We help [role] at [company type] [achieve outcome]. Teams like [a customer you can name] have [specific result]. I wanted to see whether that's worth a closer look for [company]."

Keep it short. You're not closing on this call. The job is to earn enough genuine interest that the prospect wants the next conversation with one of your reps.

3. Discovery questions that surface genuine interest

After the value prop, ask questions that get the prospect talking and tell you whether there's a real fit. Open-ended, not a yes-or-no survey.

Instead of "Are you interested in improving your lead generation?", ask "What's the biggest challenge your team is hitting with [pain point] right now?"

One distinction worth being precise about: these questions surface interest and relevant pain, the fit-plus-engagement that makes someone an interested lead (a marketing qualified lead, or MQL). They are not sales-qualification. Whether this becomes a real opportunity is something your rep works out in the next conversation, not something the agent decides on the call. Position the agent as finding the people worth your rep's time, never as qualifying the lead itself.

Then branch on the answer:

• A relevant pain surfaces, go to the matching benefit.
• No challenge surfaces, offer an industry angle or a peer comparison.
• The prospect sounds rushed, move straight to the next-step ask.

4. Objection branches

This is where AI scripts diverge hardest from human ones. A rep handles "I'm not interested" on instinct. An agent needs the instinct written down, two or three responses per objection.

"I'm not interested."

• Acknowledge and probe: "Completely fair. Is that because you've already got something in place, or just not a priority right now?"
• Remind of value: "Understood. Before I let you go, would it change anything to know [specific result] is what teams like yours have seen?"
• Exit gracefully: "No problem at all. Want me to send a short email with the details in case timing changes?"

"Just send me an email."

• "Happy to. So I make it relevant, what's the one thing you'd want it to answer?"
• "Of course. What's the best address for you?"

"How did you get my number?"

• "Your contact details are in our database of [industry] professionals. I'm reaching out because [value]. Worth a quick conversation?"

Plan for at least five common objections this way, each with its own fork.

5. A close that hands off a warm prospect

Every call should end on a specific next step, and for an AI agent that next step is almost always a human. The agent's job is to confirm the prospect is genuinely interested and connect them to a rep, not to lock a deal or run a full demo.

So make the ask concrete and give the prospect an easy path:

"It sounds like this is worth a proper conversation. I can have one of our specialists reach out to walk you through it. Would later this week work, or would you rather I send a link so you can grab a time that suits you?"

If they hesitate, offer the lighter option, a calendar link or a follow-up email with relevant results, and leave the door open. The meeting itself gets set by your rep or by the prospect picking a time. The agent's win is the warm, interested handoff. That division of labor is the whole point: AvairAI surfaces the interested lead, and your reps book and close.

6. Compliance and handoff triggers

For AI calling, TCPA compliance has to live inside the script, not get bolted on after. Build in:

• AI disclosure and identification: who is calling, that it's an AI agent and any required recording notice.
• Immediate opt-out: "take me off your list" gets honored on the spot, every time.
• Human handoff triggers: the conditions where the agent should stop talking and transfer.

Hand off to a person when the prospect asks for one, when a high-value buying signal shows up, when a question runs past the agent's knowledge or when sentiment turns negative. AvairAI's intelligent handoff framework runs these transitions automatically, but your script still has to define when they fire.

One framing worth getting right. Because US TCPA law restricts automated and AI calls to people who have given prior consent, AI call agents do their best work on warm and opted-in lists, inbound leads, existing contacts, event registrants and renewals, always AI-disclosed. Whether AI cold calling is legal at all turns on that consent, which is why the heavy cold relationship-building stays with your reps.

How to write your first script

The components above are the parts. Here is the order I'd assemble them in.

Step 1: Define the one outcome

Start at the end. What single result do you want from this call? For most outbound, it's a clear, genuine yes to a next step, an interested lead your rep can pick up. Not a closed deal. Not a full pitch. Just real interest and agreement to talk again. Write that outcome at the top of the document, and make every line serve it.

Step 2: Map the conversation tree

Before you write a word of dialogue, sketch the branches:

• Prospect engages, go to value prop, then discovery, then the next-step ask.
• "Not interested," go to the objection branch, then re-engage or exit.
• "Send email," capture the address, confirm and end.
• Prospect hangs up, end the call.

That map is the skeleton. The dialogue hangs off it.

Step 3: Write the words

Now write the lines, out loud. A few rules keep an agent sounding human. Talk like a person ("I'd love to," not "I would be delighted to"). Keep sentences short, because long ones land awkwardly when spoken. Leave room for the prospect to actually respond. And use specifics, real names, numbers and company references, since vague is what sounds robotic.

Step 4: Build the objection branches

Take your five most common objections and write two or three paths for each. Then say them aloud. If you wouldn't say it to a real person, the agent shouldn't either.

Step 5: Add the compliance layer

Drop the required disclosures in at the right moments, make the opt-out instant and wire the human-handoff triggers for anything the agent shouldn't handle alone.

Step 6: Test on yourself, then iterate

Before a single real prospect hears it, run the script on yourself. AvairAI gives you two ways to do that: Quick Test (one email plus one AI call to your own number) and Full Test (all 12 touches at compressed intervals). Listen for phrasing the agent stumbles on, branches that feel stiff, response paths you forgot and any compliance gap. Good scripts go through several rounds before launch, and once yours is live, you can A/B test your openers and objection branches so the data, not your gut, picks the winner.

Common mistakes to avoid

A handful of patterns sink AI call scripts more than any others.

Cramming in too much. The agent isn't running a webinar. One idea per section.

Treating it like a human script. A linear flow fails the moment the prospect goes off-script, and they always do. Branches are not optional.

Stiff, formal language. Write the way people talk. Contractions are fine, and so is starting a sentence with "And" or "But."

Bolting on compliance. Under the TCPA, a single violating call carries statutory damages of $500, and up to $1,500 if a court finds it willful. Build compliance in from the first draft.

No handoff plan. Some moments need a human. Define them, or the agent keeps talking when it should pass the baton.

Putting it together

Writing for an AI call agent is a different craft than writing for a rep. You're building a decision tree, not a monologue: branching logic, objection paths that fork, compliance baked into the structure and a clear point where a human takes over.

The advantage you have is that you can hear it before anyone else does. Run the script on yourself, listen to exactly what your prospects will hear and revise until it sounds like a person who happens to be an AI agent. Get that right and the agent does the patient, repeatable work of surfacing genuine interest, while your reps spend their hours where people win, in the conversations that close. That is Pair Selling: the AI handles the prospecting grind, you handle the relationships.

Start your first AI calling campaign and hear how it sounds in your own voice.