Twelve months ago, AI voice agents were impressive demos that weren't quite ready for real business use. The latency was noticeable, the voices were subtly robotic, and the context handling fell apart in anything but linear conversations. The tools available today — particularly Vapi, Retell AI, and ElevenLabs' conversational AI product — have cleared those hurdles decisively. We are actively building production voice automation for agency clients right now, and the results are worth examining in detail.
What an AI Intake Call Actually Looks Like
Here is a concrete example. A prospective client lands on an agency's website at 11pm on a Wednesday. They fill out a contact form. Previously, they'd get an auto-responder email saying the team would be in touch within one business day. With a voice agent integrated into the intake process, they receive an automated call within two minutes. The call lasts an average of four minutes, during which the voice agent:
- Introduces itself as the agency's intake assistant (not as a human)
- Asks about the nature of the project, timeline, and budget range using a structured conversation flow
- Handles follow-up questions and clarifications conversationally
- Captures all responses as structured data and writes them to the CRM
- Books a discovery call with the account manager for the following day based on calendar availability
The prospective client has spoken to "someone" within minutes of expressing interest, has a confirmed meeting on the calendar, and the account manager arrives at that meeting already briefed on the prospect's core requirements. The conversion rate from lead to discovery call increases. The quality of discovery calls improves because the intake data allows for preparation.
The Infrastructure Stack
Vapi is our current platform of choice for voice agent infrastructure. It handles the real-time audio processing, integrates with multiple TTS (text-to-speech) providers including ElevenLabs for high-quality voices, and provides a robust function calling system that allows the voice agent to interact with external systems — CRMs, calendars, databases — during the call.
The conversation logic is built on a combination of a large language model (we use Claude 3.7 for the reasoning layer) and a deterministic state machine that manages the conversation flow. Pure LLM-driven conversations are too unpredictable for intake scenarios where you need specific data captured reliably. The state machine ensures the conversation covers required topics while the LLM provides the natural language generation that makes it feel conversational rather than like a phone tree.
The Limitations to Know About
Voice agents handle linear conversations well and branch conversations reasonably well. They handle non-linear, context-heavy conversations poorly. If a prospect wants to have a nuanced strategic discussion about whether they need web development or automation — the kind of consultative conversation that requires real expertise — a voice agent is not the right tool. The technology excels at qualification and intake, not consultation.
There are also real considerations around disclosure. Regulatory environments vary, but our strong recommendation is always to be transparent that the caller is speaking with an AI agent. Beyond the ethical considerations, non-disclosure creates trust risk if discovered, and most prospects respond positively to the transparency — it demonstrates that the agency is forward-thinking about how they use technology.
The Deployment Timeline
A production voice intake agent — integrated with your CRM, calendar system, and lead routing workflow — takes between three and six weeks to build, test, and deploy. The majority of that time is in the conversation design and testing phase: running the agent against edge cases, tuning the language model prompts for your specific qualification criteria, and ensuring the CRM integration handles all the data states correctly. The infrastructure build itself is relatively fast.
The agencies that deploy this earliest will have a meaningful competitive advantage in lead response time — one of the highest-leverage metrics in the sales process. The window for that advantage is narrower than it was a year ago.
