An AI voice agent is the part of a lead generation system people want to talk about. It is not the part that decides whether the system works.
A voice agent that makes excellent calls to a badly assembled list produces excellent calls to the wrong people. LeadChirp was built around that observation: the calling is one stage of six, and the five around it are what determine whether the calling is worth doing.
The workflow, in order
Discovery. The system finds and sorts prospects matching the business, which removes the manual prospecting that usually consumes the front of a sales process.
ICP building. This is the stage that compounds. The Ideal Customer Profile is not set once at configuration; it is refined continuously against which leads actually converted. The definition of a good prospect therefore moves with the business rather than reflecting whatever was true when someone filled in a form.
Verification. Every lead goes through both automated and manual checking before it enters the pipeline. This is the least interesting stage and among the most valuable — an agent calling a dead number burns a credit and produces nothing, and at volume the difference between a verified list and an unverified one is most of the cost.
Email campaigns. Personalised sequences go out first, with follow-ups adjusted according to how each prospect engaged. Email is cheaper than calling, so it goes first and does the initial sorting.
Voice. Leads that warrant direct contact get a call. The agent holds a natural conversation, handles objections, and qualifies through dialogue rather than by reading a script at someone.
Scheduling. Where a prospect shows genuine interest, a meeting is booked or the next step triggered, with emotion analysis used to prioritise the prospects who sounded most engaged rather than simply the most recent.
The engineering problems worth naming
Latency is the whole experience. In text, a two-second pause is invisible. In conversation it is the point at which the person on the line decides they are talking to a machine and disengages. Everything in the voice path is built around holding response times inside the window where the exchange still feels like a conversation, which constrains the architecture far more than accuracy does.
Interruption is normal, not an edge case. People talk over each other. A voice agent that finishes its sentence regardless is immediately identifiable and immediately irritating. Handling barge-in properly — stopping, listening, picking up the thread — is a substantial share of the work and is what separates a system people stay on the line for from one they hang up on.
Qualification has to be conservative. An agent that marks a lukewarm prospect as qualified wastes a salesperson's time, and salespeople stop trusting the queue after a small number of those. The qualification threshold is deliberately cautious, and an ambiguous call is surfaced with its transcript rather than resolved by the agent guessing.
Cost is metered in minutes. Calling time is the unit that is actually consumed, which makes the economics unusually legible: every stage before the call exists to make sure the minutes are spent on someone worth calling. That framing is why verification and ICP refinement are first-class parts of the product rather than preprocessing.
What the agent does not do
It does not close. It qualifies and hands over, and the handover carries the transcript and the reasoning so the person picking it up knows what was said rather than inheriting a score.
It also does not pretend. Where a conversation goes somewhere the agent cannot handle, the correct behaviour is to say so and route to a person — not to improvise. A voice agent improvising on a sales call is making commitments on behalf of a business, which is the one thing it must never do.
The general lesson
The pattern generalises past sales. The impressive component in an agentic system is rarely the one that determines whether it works; the quality of what goes into it is. LeadChirp spends most of its engineering on discovery, profile refinement and verification precisely so that the conversational layer — the part anyone would demonstrate — is pointed at the right person when it runs.
Have a system that needs to do this?
500+ systems shipped since 2011, and we still maintain most of them. Tell us what you are trying to move.




