Voice agent for lead qualification — pre-launch checklist

Keep this list open while you build and launch an agent that calls every new lead. The sections follow the order of the work: preparation, call script, data to log, testing, handover to a human, and what to watch after launch. Every item is an action with a result you can check.

0. Before you build

• The form states clearly that an AI agent will call after submission. Anyone who isn't expecting a call shouldn't get one.
• The form has a field or checkbox where the lead agrees to be contacted by phone. Without it, you don't call.
• You've decided which hours the agent may call. A submission sent at night can wait until morning.
• The form collects at least: name, phone, email, company name, role, description of the need, company size.
• The phone field enforces one format, or the flow has a step that cleans the number.
• If leads come from several countries, the form has a separate country code field. A hard-coded country code only works for one country.
• You've chosen the three variables passed to the agent on every call: name, company, description of the need.
• The flow is sketched on paper before you open the tool: form, number cleanup, call, wait, log.

1. Call script and qualifying questions

Starting settings

• The agent waits for the person to speak first.
• The system prompt has sections: identity, style, response guidelines, lead details, call flow, error handling, behaviour rules.
• Lead details are written as variables: {{lead_name}}, {{lead_company_name}}, {{lead_request}}.
• The agent's end-call function is switched on.

Opening (template)

"Hello, this is [agent name], an AI agent calling on behalf of [company]. Am I speaking with [name] from [lead's company]?"

• Wrong number: the agent apologises and ends the call.
• Bad time: the agent apologises and asks when to call back.
• The person has time: the agent mentions the form and the described need.

Questions (template to adapt)

1. Interest: "Am I right that this is about [description of need]?"
2. Motivation: "What made you look into this now?"
3. Timeline: "When would you like to get started?"
4. Experience: "Have you worked with automation or AI before? How did it go?"
5. Budget: "Do you have a budget set aside for this project?"
6. Working model: a question confirming the person accepts how your next stage works, for example a paid scoping phase. Replace it with your own.

Closing (template)

"Thank you for your time. Someone from our team will be in touch about next steps."

• The agent makes no promises about dates, prices, or scope that aren't in its prompt.

2. Data to extract and log

You define each field in the voice platform with a name, a type, and a description of what the agent should extract. Each field is linked to the right agent.

• The sheet has a column for each field plus the form columns.
• You know where the structured outputs sit in the platform's response (usually deep, under the call's artifacts).

3. Test plan

• Sample data is pinned on the input step, so you don't fill in the form for every test.
• A number with one digit missing: the record lands in the sheet as an invalid number, and no call starts.
• A full conversation: you answer, go through every question, and check that each field in the sheet matches what you said.
• Wrong number and bad time: the agent behaves as the prompt says.
• Voicemail: the row gets the "call back" status.
• Polling loop: the first wait is about a minute, later checks come every few seconds, and the loop only exits on the "ended" status.
• The call-details result is limited to the first item before the status check.
• Call limits: you know whether a number bought on the platform has a daily outbound cap, and you have a plan for your own number from an internet telephony provider.

4. Handover to a human

• Every row with a "completed" status has someone responsible for the follow-up.
• The person on your team reads motivation, timeline, and budget before the call, not during it.
• "Call back" rows have a set time for the next attempt.
• Invalid-number rows go to email follow-up.
• If the person asks for a human or is clearly getting upset, the agent either ends politely or transfers the call, depending on what you configured.

5. After launch

• You listen to recordings and read transcripts regularly, not just in the first week.
• You compare the fields in the sheet with what was actually said on the call.
• You note the questions people stumble over and rewrite them in the prompt.
• You check which fields the team actually uses before calling, and remove the ones nobody reads.
• You track how many calls end in an error and how many go to voicemail.

One last line

The agent asks the questions; a person runs the relationship. Launch with three or four questions, and add more only once the team starts using the answers.