FollowAI Cold AI Calls: Build a Voice Manager That Qualifies and Books

A real AI cold-calling manager is a phone system, a conversation agent, and a sales workflow connected together. It selects an eligible lead, opens a call through a business number, speaks in real time, asks the approved qualification questions, books a meeting when the rules allow it, and writes the outcome into the CRM. FollowAI can build this manager with LiveKit Agents, a SIP carrier, a voice stack, Sales OS, and a controlled tool layer.

The practical use is easy to picture: a company has more qualified prospects than its team can call. The manager calls the approved queue, handles the first conversation, detects voicemail or an unavailable line, records a structured outcome, and transfers a serious opportunity to a person. This is very different from an article about cold messages. The deliverable is an actual outbound voice operation.

The voice architecture

What the company that provides the phone line does

LiveKit does not magically create a caller identity for every country. The implementation needs a telephony provider or a number purchased through a supported phone-number service. That provider supplies a DID, caller ID, SIP trunk or connector, country coverage, billing, and carrier-side call records. Twilio, Telnyx, Sinch, and similar providers are examples; the right choice depends on the target countries, number type, routing, and compliance requirements.

The normal setup is:

1. purchase or port the business number;
2. configure the provider’s SIP trunk and authentication;
3. create a reusable LiveKit outbound trunk with the provider endpoint;
4. dispatch the voice agent to a room for the selected lead;
5. create a SIP participant that dials the destination number into that room;
6. listen for call status, voicemail or IVR detection, connection failure, and hang-up events;
7. write the call result and next action to Sales OS.

The same architecture can accept inbound calls. Dispatch rules can send the number to the correct agent or a human queue, while the CRM record gives the agent the caller’s context.

How the manager is trained

“Train the model” does not have to mean fine-tuning a new model before the first call. A production voice manager is taught through an operating package:

• system instructions: role, tone, language, opening, disclosure, and stop conditions;
• offer knowledge: services, supported integrations, boundaries, examples, and approved answers;
• qualification schema: the exact fields to collect, such as need, current process, timeline, authority, and meeting goal;
• tools: read the CRM, create an activity, check calendar availability, book a meeting, send a confirmation, and request a human;
• conversation states: greeting, permission, discovery, qualification, objection, booking, transfer, voicemail, and close;
• examples: approved conversations showing good questions, bad assumptions, and correct handoffs;
• evaluation set: scripted test calls for silence, interruptions, accents, objections, wrong numbers, sensitive questions, and tool failures.

After test calls, we label where the agent interrupted, missed a field, made an unsupported promise, failed to stop, or wrote the wrong CRM state. Only approved corrections enter the next version. That is safer and cheaper than blindly feeding every recording back into a model.

What the agent is allowed to do

The agent can greet, explain the reason for the call, ask approved questions, summarize the need, offer available meeting times, send a confirmation, and create a follow-up. It must not invent a discount, promise a legal result, collect sensitive payment data over an unapproved channel, or keep talking after an opt-out. A request for a human, an angry customer, a regulated topic, or a low-confidence answer should trigger a transfer or task.

The workflow also needs operational fallbacks. If the calendar is down, the agent collects a preferred time without pretending it booked. If the CRM is unavailable, the call event is stored for replay. If the number reaches voicemail, the configured voicemail path runs instead of a full conversation.

What a call costs

The price is a stack of separate components, not one mysterious “AI call” fee:

As a transparent example, a current US local outbound rate published by Twilio is $0.014 per minute, with a $1.15 monthly local number. A thousand outbound minutes would therefore be about $15.15 for the carrier portion before LiveKit, speech, model, recording, and application costs. Rates change by country, number type, volume, and provider, so the quote must be checked before launch.

What FollowAI can build

FollowAI can build the complete cold-calling operation: lead queue, business number and SIP connection, LiveKit agent, voice instructions, offer knowledge, qualification tools, calendar booking, CRM updates, recording policy, human transfer, Telegram alerts, evaluation calls, dashboards, and maintenance. The goal is not to make the most calls. The goal is to create more qualified conversations while keeping every action visible and controlled.