Skip to main content

Overview

A UK-based automotive recruitment firm with more than 30 years in the industry had built up a database of over 50,000 candidates who had previously registered interest in roles such as technicians, panel beaters, service advisors, HGV mechanics, and sales executives. Most of these leads had gone cold. The firm’s recruiters did not have the time or capacity to work through the full database manually, which meant a large volume of potentially valuable candidates was sitting unused. RapidCall was used to deploy an AI voice agent that could re-engage these candidates at scale, identify who was still open to opportunities, and book callbacks directly with the right human recruiter. The objective was not to close candidates on the call. The objective was to surface warm interest and route qualified conversations back to the recruitment team.

The Goal

The company needed an agent that could:
  • call dormant candidates at scale,
  • reference their previous interest naturally,
  • identify who was still open to hearing about roles,
  • handle common objections and questions,
  • book callbacks with senior recruiters,
  • and route those bookings to the correct consultant based on geography and specialism.
This was a classic reactivation workflow: large database, low manual capacity, and high upside if the right people could be brought back into conversation.

Results After 2 Full Months

March 2026 performance

The headline result was the 38.9% booking rate from connected conversations. For context, human recruiters typically convert cold re-engagement lists at much lower rates. The AI agent performed strongly because it combined scale, consistency, and speed. It reached more people, called at better times, and followed the same proven structure every time.

Voice Stack and Configuration

Core stack

The call flow was structured and linear, so a lightweight model was the right fit. The agent did not need deep reasoning. It needed to follow the script naturally, respond quickly, and handle a small number of common branches reliably. A British female voice was chosen because it matched the client’s brand and audience. The tone needed to feel local, professional, and approachable. Call-centre background sound was added deliberately. It made the call feel like it was coming from a real busy office, which reduced early suspicion and helped the agent feel less synthetic.

Key speech settings

A few settings made a particularly large difference. The 0.4-second pause before speaking was critical. Without it, the agent spoke the moment the line connected, which made it sound like an autodialler. A very short pause made the interaction feel much more human. Response eagerness was kept very high because recruitment calls lose momentum quickly if the caller hesitates. Candidates are often busy, distracted, or between tasks. A slow agent feels awkward almost immediately. Interruption sensitivity was also kept high because candidates often cut in with questions. The agent needed to yield cleanly rather than talk over them.

Prompt Design

The prompt followed the standard five-part structure used across RapidCall deployments.

Role

The agent was named Beth and framed as an AI Candidate Manager. The role instructions focused heavily on tone and vocabulary. The agent was told to sound friendly, natural, and British, while using simple everyday wording. That mattered because the audience was not corporate. These were mechanics, technicians, bodyshop staff, and trade professionals. Language that felt too formal or too polished created distance immediately. Terms like job, role, and salary performed better than more corporate alternatives.

Context

Each call was personalised using CRM variables, including:
  • candidate first name,
  • date they last registered interest,
  • role type,
  • postcode,
  • email,
  • and assigned recruiter details.
This allowed the agent to reference prior activity naturally, which made the conversation feel like a real follow-up rather than a generic cold call.

Availability

The booking logic was restricted to Monday to Friday only. The agent checked real-time calendar availability through a Cal.com integration before offering any time. It never promised a slot without checking first, and it did not allow weekend callbacks.

Guidelines

The guidelines section carried a lot of the production value. The main rules included:
  • keep responses short,
  • never repeat the same sentence twice,
  • use casual British filler words naturally,
  • handle objections using a structured framework,
  • read contact information clearly,
  • never make guarantees about jobs,
  • and offer booking times one at a time rather than as a list.
One especially important rule covered candidate questions about role details. If a candidate asked about salary, location, or shifts, the agent was told to answer clearly, give realistic detail, and then guide the conversation back toward the callback booking. This prevented the trust loss that happens when an agent dodges a direct question.

Main script

The live call flow had six core steps: The script was designed to keep momentum high and decisions simple.

Handler Design

The agent included handlers for the most common edge cases: These handlers mattered because most live conversations do not stay perfectly on script. The agent needed to handle resistance without sounding robotic or policy-driven. The AI disclosure handler is worth noting. Very few candidates asked whether they were speaking to AI. When they did, direct honesty did not hurt performance. The agent answered plainly and immediately moved back to the reason for the call.

Workflow Architecture

The agent sat inside a fully automated outreach workflow.

Data source

Candidate data was stored in a Google Sheet acting as a lightweight CRM. Each row included:
  • name,
  • phone,
  • email,
  • postcode,
  • role type,
  • registration date,
  • and assigned recruiter email.

Trigger logic

A Make.com scenario scanned the sheet every 10 minutes between 8:00 AM and 4:00 PM. If it found uncalled leads, it triggered up to 10 concurrent calls through the RapidCall API.

Calling windows

The workflow used two main calling windows:
  • First attempt: 8:00 AM to 2:00 PM
  • Second attempt: 2:00 PM to 4:00 PM
Leads not reached on the first call were retried later the same day.

Outcome logging

Every call outcome was written back into the sheet automatically, including:
  • call status,
  • outcome category,
  • sentiment,
  • summary,
  • recording URL,
  • AI question flag,
  • and do-not-contact flag.
This gave the team a live operations view without needing a heavier CRM setup.

Email follow-up

Based on call outcome, the workflow triggered one of several automated email templates, including:
  • no answer,
  • voicemail reached,
  • not interested,
  • referral interest,
  • and second attempt unsuccessful.

Booking flow

When a candidate agreed to a callback, the agent checked live availability via Cal.com, offered a slot, and booked it directly. The candidate and recruiter both received notifications, and the team manager was also informed.

Territory routing

Postcode prefixes were mapped to recruiter territories and specialisations using a lookup table. This ensured callbacks were booked directly with the right recruiter based on both geography and sector.

What We Learned

1. The opening line matters more than anything else

The first version of the opening was too long. It tried to explain the purpose of the call before earning attention. Candidates dropped before the agent finished the sentence. Reducing the opening to a simple identity check improved performance immediately.

2. One booking slot at a time works better than a list

When the agent offered several times at once, calls stalled. Candidates paused, hesitated, or asked the agent to repeat them. Offering one time at a time made decisions easier and improved booking speed.

3. The referral handler recovered value from “no” calls

The referral offer turned a meaningful number of otherwise dead conversations into follow-up opportunities. Without it, those calls would simply have ended.

4. Honest AI disclosure was not a problem

Only a very small number of candidates asked if the caller was AI. Those who did were handled directly, and the disclosure itself did not cause drop-off.

5. Voicemails helped extend reach

The agent left short personalised voicemail messages, and these were paired with automated follow-up emails. This meant candidates who did not answer still received a contact point.

6. The second attempt was worth it

The second pass later in the day generated additional conversations and bookings from candidates who simply were not available in the first window.

7. Routing by postcode removed a major manual step

Automatic recruiter assignment meant bookings landed on the correct calendar immediately. That eliminated manual triage and saved the team time every day.

8. Simple language performed better than formal language

Trade audiences responded better to plain, familiar wording. When the script sounded too corporate, engagement dropped.

When to Use This Pattern

This playbook is a strong fit for any business with a large database of previously interested leads that has gone underworked or dormant. It works especially well when:
  • the leads originally expressed genuine interest,
  • the business has something new to offer them now,
  • the goal is to qualify and hand off, not close directly,
  • and the human team does not have capacity to work the full list manually.
This pattern is useful well beyond recruitment. Any business sitting on a database of past enquiries, old leads, or dormant prospects can use the same structure: re-engage → qualify → book human follow-up The AI handles the high-volume filtering.
The human team handles the high-value conversation.

Prompt Template Availability

A stripped-down version of this prompt structure, with placeholders instead of client-specific details, can be included in the Appendix as a reusable template for:
  • candidate reactivation,
  • dormant lead follow-up,
  • callback booking,
  • and referral-led outreach.
That gives readers a starting point they can adapt to their own sector and workflow. THE EXACT PROMPT WE USED TO GET THESE RESULTS: