Voxvencer

Sales and outbound

How to Write a Call Script an AI Agent Can Follow

AI phone agents need instructions, not word-for-word scripts. How to write goals, facts, rules and handoffs so an AI agent stays on plan.

Person writing notes on paper at a desk
Photo: Startup Stock Photos, StockSnap (CC0)

The short version

  • Don't give an AI agent a word-for-word script. Give it a goal, the steps, the facts it may use and the rules it must follow.
  • Write the handoff and opt-out rules as carefully as the pitch.
  • Test the script against awkward callers before launch, then refine from real transcripts.

A traditional call script is written for a person. It gives them the words for the opening, a few key questions and rebuttals, and trusts their judgment for the rest. An AI agent needs something a bit different. It has no judgment about your business beyond what you write down, but it's very good at adapting phrasing to the caller as long as it understands the goal.

Here's how to write instructions that keep an AI phone agent on track.

Script vs instructions

If you give an AI agent a word-for-word script, two things happen. It sounds stiff, because people never answer the way a script assumes. And it breaks the moment a caller goes off-script, which happens on almost every call.

Instead, write the script as instructions: what the call is for, what steps to take, what facts to use, what rules never to break, and when to hand off. The agent supplies the wording.

You'll still include some exact phrases where wording matters: legally required disclosures, the opening line, and anything your compliance team needs said a certain way.

The structure that works

1. Role and goal

Start with who the agent is and what a successful call looks like.

You are the automated assistant for Brightwave Solar, calling people who requested a solar quote on our website in the last hour. Your goal is to confirm their interest, ask four qualifying questions and book a 30-minute home assessment with a solar advisor. If they're not interested, end the call politely.

One clear goal per agent. "Qualify leads and also handle billing questions and also upsell batteries" produces muddled calls.

2. The opening, word for word

Hi, this is the automated assistant from Brightwave Solar. You asked about a solar quote on our website a little while ago. Is now a good time for a couple of quick questions?

This is where disclosures go: business name, that it's automated, and recording notice if you record.

3. Steps, in order, as goals

  1. Confirm they own the home. If they rent, explain we can only install for homeowners and end the call politely.
  2. Ask roughly what their monthly electric bill is in summer.
  3. Ask whether the roof is less than 15 years old, or if they know of any planned roof work.
  4. Ask when they'd ideally like the system installed.
  5. If they own the home and their bill is over $100, offer an assessment slot using the calendar tool.

Writing each step as a goal ("confirm they own the home") rather than a fixed question lets the agent ask naturally and skip steps the caller has already answered.

4. Facts the agent may use

List the information it's allowed to share. Be specific.

  • Assessments are free and take about 30 minutes.
  • We install in Lewis and Clark, Cascade and Gallatin counties.
  • Financing is available. Rates depend on credit, so never quote a rate.
  • A typical install takes one to two days once permits are approved.

Anything not on this list, the agent should not state as fact. That's how you prevent confident wrong answers.

5. Rules

Rules are the "never" and "always" list. Write them plainly.

  • Never quote a price or savings estimate. Say the advisor will cover that at the assessment.
  • Never claim to be a person. If asked, say you're an automated assistant.
  • If the person asks not to be called again, confirm you'll remove them, use the opt-out tool, and end the call.
  • If the person sounds confused about why we're calling, explain the website request once. If they still don't recall it, apologize and end the call.
  • Keep each reply to two sentences or fewer unless the caller asks for detail.

The opt-out rule deserves extra care. People say it many ways: "take me off your list," "stop calling," "I'm not interested, don't call back." List several examples so the agent recognizes them. Our article on TCPA and AI voices explains why.

6. Handoffs

When should the call go to a person, and how?

Transfer to a live advisor if the caller asks for a person, has a question about an existing installation, or wants to talk pricing now. Before transferring, say: "Let me connect you with one of our advisors. One moment." Pass a summary of their answers so far.

We cover handoff design in more depth in designing the human handoff.

7. Objections

You can include brief guidance for common objections, but keep it honest and short.

  • "I'm busy": offer to call back at a time they choose.
  • "Is this a sales call?": "It's a follow-up to the quote request you made on our website. I just need a couple of details to set up a free assessment."
  • "How much does it cost?": explain that pricing depends on the home and the advisor covers it at the assessment.

Don't write pushy rebuttals. An agent that argues with people who said no creates complaints, and in outbound calling it creates legal risk.

8. Ending the call

Before ending, confirm any booking with day, date and time, and tell them they'll receive a text confirmation. Thank them and end the call.

Write for the ear

Phone replies need to be short. Tell the agent to:

  • Use one or two sentences per turn
  • Ask one question at a time
  • Read back important details (times, numbers, addresses)
  • Avoid lists of more than three options

Test before you launch

Run test calls where you deliberately make it hard:

  • Answer two questions at once
  • Go off-topic, then come back
  • Ask something not in the facts list
  • Ask for a person
  • Say "stop calling me" in an unusual way
  • Mumble, or call from a noisy place

Read every transcript. Each failure usually points to a missing fact, an unclear rule or a step written as a question instead of a goal.

Improve from real calls

After launch, review transcripts weekly. Look for questions callers ask that aren't covered, places where the agent's replies are too long, and calls that should have been handed off sooner. Small edits to instructions compound quickly.

If you want to test how a language model follows instructions like these, paste a draft into our free AI chat tool, play the caller, and see where it drifts. The call script writer can also produce a first draft you can turn into instructions.

A complete example: appointment confirmation agent

Here is a full set of instructions for a different use case, so you can see the structure applied end to end.

Role and goal You are the automated assistant for Lakeside Physical Therapy. You call patients the day before their appointment to confirm, reschedule or cancel. A successful call ends with the appointment confirmed, moved or cancelled in the scheduling system.

Opening (say exactly) "Hi, this is the automated assistant from Lakeside Physical Therapy, calling for [first name]. Is this [first name]?" If the person isn't the patient, say: "No problem. Could you ask [first name] to call us at 555-0160? Thanks." Then end the call. Don't share appointment details.

Steps

  1. Confirm you're speaking with the patient.
  2. State the appointment: day, date, time and therapist.
  3. Ask if they can still make it.
  4. If yes, confirm and remind them to wear comfortable clothing.
  5. If no, offer to reschedule using the scheduling tool. Offer up to two times at once.
  6. If they want to cancel without rescheduling, cancel and tell them they can call to book later.

Facts

  • Address: 48 Lake Road, Suite 2. Free parking in front.
  • Cancellations with less than 24 hours' notice may incur a fee. Mention this only if they cancel.
  • Office phone: 555-0160, open Monday to Friday, 7 to 6.

Rules

  • Never discuss diagnosis, treatment or billing. Offer a callback from the office.
  • If asked whether you're a person, say you're an automated assistant.
  • If they ask not to receive these calls, confirm and record the opt-out.
  • Keep replies short.

Handoffs If they have a question you can't answer, offer to have the office call them back and record their question.

Short, specific and testable. Every line answers a question the agent would otherwise have to guess.

Man sketching a process flow on a whiteboard
Photo: Startup Stock Photos, StockSnap (CC0)

Turning a human script into AI instructions

If you already have a script for human callers, here's a quick conversion process:

  1. Highlight the compliance lines that must be said exactly. These go into the opening or rules as quoted text.
  2. Turn each question into a goal. "Do you own your home?" becomes "Confirm whether they own the home."
  3. Pull out facts buried in rebuttals and talking points, and list them separately.
  4. Write down the unwritten rules your best reps follow. "Don't push if they mention a recent death in the family." "Never promise installation dates." These are often missing from human scripts because people learn them on the job.
  5. Define success for the call in one sentence.
  6. Add handoff rules for situations your human reps would escalate.

Versioning and testing

Treat instructions like code. Keep each version with a date and a note about what changed. Before publishing a change, run the same set of test conversations against old and new versions and compare. If a change makes one problem better and another worse, you'll see it before customers do.

A simple test set might include fifteen to thirty scenarios: the common path, each objection, each handoff trigger, an opt-out, a wrong person answering, an angry caller and a few odd questions. Add a new scenario every time you find a real call that went wrong.

Frequently asked questions

How long should AI agent instructions be?

As long as needed to cover the goal, steps, facts, rules and handoffs, and no longer. One to two pages is typical for a focused agent. Very long instructions with many edge cases can make behavior less predictable.

Can I reuse my existing human script?

Yes, as a starting point. Pull out the goal, the key questions, the facts and the compliance language, and restructure them using the sections above.

What if the agent says something wrong?

Find the transcript, work out which instruction was missing or unclear, fix it and re-test. Most mistakes come from information the agent was never given.

Written by the Voxvencer editorial team. We build and run AI voice agents for call centers and small businesses, and we write about what we see on real phone lines. Questions or corrections: info@voxvencer.com.

Keep reading