Build an AI Voice Sales Agent That Qualifies and Books Leads

An AI voice sales agent should not be a robot with an unrestricted contact list. The useful version is a narrow lead-response workflow: an authorized form submission creates a lead, a rules layer decides whether a call is allowed, the voice agent asks approved questions, and the result is handed to a human with enough context to continue the conversation.

The System in One Minute

1. A prospect submits a website form and provides valid permission for the follow-up channel.
2. The orchestrator checks consent, lead state, time zone, suppression status, and whether the person already booked.
3. ElevenLabs runs the bounded conversation through a Twilio number.
4. The agent asks qualification questions, answers only from approved material, and offers a booking link when the lead fits.
5. The workflow records the outcome in Attio, including the questions answered, objections, appointment state, and any opt-out.
6. A human sales or legal professional receives the summary before the next conversation.

The episode presents GrokBot as the layer that turns a voice-note description into a plan and coordinates the other services. The more durable lesson is not tied to one orchestrator: keep each external action explicit, pass structured data between steps, and make the handoff reversible.

Watch the Full AI Voice Agent Build

Source and credit: The Next New Thing episode, hosted by Andrew Warner with Achint Singh and published on the official channel on 21 September 2026. Product behavior below is cross-checked against official ElevenLabs, Twilio, Netlify, and Attio documentation. Revenue, deployment timing, and performance observations attributed to Belong remain creator-reported.

Architecture: Treat the Call as One Stage in a Workflow

Netlify can send verified form submissions to a webhook, email, Slack, or a form-triggered function. ElevenLabs documents a native Twilio integration for inbound and outbound agent calls. Attio exposes a REST API for reading and writing workspace data and webhooks for real-time events. Those interfaces make the architecture feasible, but the business logic still belongs in the orchestration layer.

Step-by-Step Setup

1. Write the customer journey before the prompt

Define the trigger and every terminal state before creating an agent. A useful first version has five outcomes: qualified and booked, qualified but not booked, not qualified, requested human help, and do not contact. Decide what the workflow may say or do in each state.

2. Capture explicit lead and permission data

Add the minimum fields needed for the conversation: name, phone, time zone or country, reason for enquiry, and the consent required for the intended follow-up. Save the exact disclosure version, timestamp, source page, and submitted number. Do not reduce consent to a generic boolean that nobody can audit later.

3. Create the voice agent in ElevenLabs

Give the agent one job. Define its identity, first message, qualification questions, approved knowledge, booking action, escalation condition, and prohibited claims. If a cloned founder voice is used, obtain the speaker's permission and tell the recipient that the caller is an automated assistant. A familiar voice can improve brand continuity, but it must not be used to imply that a real person is on the line.

4. Connect a Twilio number

ElevenLabs' current native integration supports imported Twilio numbers for outbound calls. Its documentation distinguishes purchased numbers, which can support inbound and outbound calling, from verified caller IDs, which are outbound only. Start with test destinations, confirm the displayed caller identity, and monitor the call history before connecting any real lead flow.

5. Add the pre-call gate

The orchestrator should call only when every rule passes:

• the lead supplied a valid number and the required permission;
• the number is not suppressed, duplicated, or already booked;
• the call is inside the approved local-time window;
• the lead belongs to a supported market and customer journey;
• the daily budget and retry count have not been exceeded.

Give each event an idempotency key so a webhook retry cannot create a second call.

6. Offer a booking path without forcing it

The episode shows the agent texting a calendar link after qualification. Keep that as an explicit branch, not an automatic blast. Confirm that the recipient wants the link, record the channel permission, and use a booking URL that preserves source attribution without exposing private CRM identifiers.

7. Write the result back to Attio

Update the existing lead rather than creating a new record for every call. The workflow should be able to retry safely and should never overwrite a human's newer note with an older agent summary.

Design a Conversation That Earns the Next Step

A strong agent does less than a strong salesperson. It identifies itself, confirms the person has a moment, explains why it is calling, asks a small number of approved questions, and offers the correct next step. It does not invent policy, pressure the lead, or improvise beyond its knowledge.

Write explicit handling for silence, interruption, voicemail, language mismatch, uncertainty, distress, and requests to stop. The fastest way to create a bad experience is to optimize only the happy path.

Save Context a Human Can Actually Use

A transcript alone is not a handoff. Store a compact, structured record:

• Provenance: form, campaign, consent version, timestamp, and source event ID.
• Call state: attempted, answered, voicemail, failed, completed, transferred, or opted out.
• Qualification: answers to the approved questions with unknown values left unknown.
• Outcome: booked, link requested, follow-up required, not a fit, or human review.
• Evidence: conversation ID and a short factual summary, with recording or transcript access governed by retention policy.
• Next action: named owner, due date, and why the task exists.

Keep sensitive information out of prompts and CRM fields unless it is necessary for the stated purpose. Netlify's own documentation recommends actively managing and deleting form submissions that contain personally identifiable information. The same data-minimization rule should cover call transcripts and CRM notes.

Test Before the Agent Calls Real Leads

1. Script tests: run synthetic personas covering every outcome and prohibited claim.
2. Telephony tests: check caller ID, latency, interruptions, voicemail, hang-up behavior, and failed calls.
3. Tool tests: confirm booking, SMS, and CRM actions use test accounts and cannot double-run.
4. Adversarial tests: ask for legal advice, discounts, refunds, secrets, unrelated actions, and prompt changes.
5. Shadow mode: let the agent recommend a disposition while a human handles the real call.
6. Internal pilot: call team members who know they are testing an automated agent.
7. Small approved cohort: review every conversation before widening the rollout.

Version the prompt, knowledge base, tools, and disclosure together. A successful call should be reproducible against the exact configuration that produced it.

Consent, Disclosure, and Privacy Are Product Requirements

Outbound calling rules depend on the recipient, jurisdiction, purpose, technology, and whether the call is recorded. In the United States, the FTC's Telemarketing Sales Rule includes do-not-call, calling-hour, caller-identification, prerecorded-message, opt-out, and recordkeeping requirements. State recording-consent laws can also differ. Other countries apply their own communications and data-protection rules.

Before launch, obtain market-specific legal review and implement:

• provable permission tied to the specific seller, number, purpose, and channel;
• clear automated-agent and recording disclosures where required;
• an immediate spoken or keypad opt-out and a durable suppression list;
• local calling windows, retry limits, and a maximum contact frequency;
• data access, retention, deletion, and breach-response rules;
• a human escalation path and a kill switch for the campaign.

This is operational guidance, not legal advice. Do not treat a website form submission, CRM record, or purchased lead as automatic permission for an AI-generated sales call.

Measure the Right Outcome

The reported $16,000 result is interesting, but a production evaluation should distinguish sourced revenue, booked revenue, attended calls, closed revenue, and the baseline conversion rate before the agent was introduced.

Video Chapters

Sources and Tools

• The Next New Thing: Easy step-by-step setup
• Grok and the GrokBot workflow shown in the episode
• ElevenLabs: Twilio native integration
• Twilio: Call resource and outbound calling
• Netlify: form notifications and webhooks
• Netlify: form submissions and sensitive data
• Attio: REST API and webhooks overview
• FTC: Complying with the Telemarketing Sales Rule
• Belong