Voice Agents for Candidate Screening: UK Recruitment Ops

A 35-person UK staffing agency was running 35 first-round screening calls per day last autumn — availability checks, right-to-work confirmations, and salary bracket questions that resolved or disqualified in under 90 seconds, but still consumed 14 consultant hours per week. We rebuilt the first screen as a voice agent flow. The agent now handles 62% of calls end to end; consultants only dial when there is a qualified candidate to progress.

That outcome required mapping exactly which questions caused disqualification, designing prompts that extract structured data reliably, wiring verdicts back into the ATS, and auditing every question against the Equality Act before the first live call. This post covers all four in the order we built them.

Mapping the first-round screening call: which questions disqualify 60% of applicants before minute three

Three questions account for the bulk of disqualification outcomes. After reviewing 12 months of historical screening notes, the order that minimises average handle time is:

1. Right to work — "Do you have the right to work in the UK without employer sponsorship?" A "no" closes the call immediately for the 85% of SME clients who cannot offer a sponsor licence.
2. Availability — "The role requires [shift / hours / start date]. Does that work?" Mismatch on start date or shift pattern is the second-largest disqualifier in light-industrial and temp pipelines.
3. Salary bracket — "The role pays £[X]–£[Y] gross. Is that in your range?" Salary misalignment at this stage saves both sides from a 45-minute interview that goes nowhere.

FIRST-ROUND SCREENING CALL FLOW
─────────────────────────────────────────────────────
DIAL ATTEMPT
   │
   ├─ No answer / voicemail
   │      └─ voicemail drop + SMS link → retry queue (72h window)
   │
   └─ Live answer
          │
          ├─ CONSENT DISCLOSURE (0–10 sec)
          │      └─ Opt-out requested → transfer to consultant queue
          │
          ├─ RIGHT TO WORK gate
          │      No  → graceful close + ATS: disqualified_rtw
          │      Yes → continue
          │
          ├─ AVAILABILITY gate
          │      No  → graceful close + ATS: disqualified_avail
          │      Yes → continue
          │
          ├─ SALARY BRACKET gate
          │      No  → graceful close + ATS: disqualified_salary
          │      Yes → continue
          │
          └─ QUALIFIED → book slot or escalate to consultant
                         ATS: screening_passed

Average handle time for disqualified calls: 58 seconds. Qualified calls averaged 2 minutes 14 seconds.

Prompt architecture for structured candidate extraction: availability, salary bracket, location, and right to work

The voice agent has one job during extraction: get the answer into a typed field the ATS can consume. Free-text transcripts create matching problems downstream. Structure the LLM output as a JSON object from the first turn.

The system prompt below is the configuration used on the Vincere deployment:

{
  "role": "system",
  "content": "You are a screening assistant for [Agency Name]. Conduct a brief eligibility check in this exact order: right_to_work → availability → salary_bracket → location. Extract each answer into the schema below. If a candidate gives an ambiguous answer, ask one clarifying question before marking the field. Do not ask about age, health, disability, family status, or any characteristic protected under the Equality Act 2010.\n\nExtract to:\n{\n  \"right_to_work\": \"yes | no | unclear\",\n  \"available_start\": \"ISO-8601 date or null\",\n  \"shift_confirmed\": \"yes | no | unclear\",\n  \"salary_in_range\": \"yes | no | unclear\",\n  \"location_confirmed\": \"yes | no | unclear\",\n  \"disqualification_reason\": \"rtw | availability | salary | location | none\",\n  \"screening_verdict\": \"pass | fail | escalate\"\n}"
}

The escalate verdict fires when LLM confidence on any field drops below 0.7, or when the candidate raises a question the agent cannot resolve — notice period exceptions, visa type queries, apprenticeship eligibility. That signal routes the call or a follow-up task to a consultant.

One failure mode we hit early: candidates on Graduate Route visas ("it expires in 14 months") were classified as unclear and escalated to consultants who then re-ran the same questions. Adding a clarification branch that confirmed visa expiry date against the role's contract length reduced false escalations by 34%.

For prompt patterns that keep turn latency under 800ms, see our post on prompt engineering for voice agents.

ATS integration: pushing screening verdicts to Bullhorn, Vincere, and Greenhouse automatically

All three platforms support verdict writes via REST API, but integration complexity varies:

The Bullhorn write requires session management: establish a session, refresh the BhRestToken every 10 minutes, and map verdict fields to Bullhorn's internal status codes. We used n8n for orchestration — a webhook node receives the call completion event, a code node formats the payload, and an HTTP node POSTs to Bullhorn. Total latency from call end to ATS update: under four seconds.

Greenhouse is the awkward one. The Harvest API does not support bulk status updates — each application is a separate PATCH. At 35 calls in a two-hour window you will hit the rate limit without a Redis write queue in front of the HTTP node.

Our voice AI and document analysis case study shows this same n8n webhook-to-API pattern applied to a document extraction pipeline, with code for session refresh handling.

Consent and disclosure design: what to say in the opening 10 seconds to stay on the right side of UK employment law

The ICO's guidance on AI and data protection is clear that automated decision-making producing legal or similarly significant effects on individuals requires explicit disclosure before processing. Under Article 22 of UK GDPR, candidates have the right to request human review of any automated decision.

The opening script we use runs to 12 seconds:

"Hi, this is [Agency Name]'s automated screening assistant calling about your application for [Role]. This call is recorded and processed by an automated system. You can speak to a person at any time by saying 'human' or pressing 0. Otherwise, I'll ask you three quick eligibility questions — about 90 seconds. Ready to start?"

This script identifies the call as automated, states that processing is automated (Article 13/14), and offers a human opt-out before any question is asked. What it avoids: naming the LLM vendor, or framing the call in terms that induce anxiety before questioning. Document those choices in your DPIA — treat the assessment as part of the build spec, not a post-launch retrospective.

Equality Act constraints a screening voice agent must never cross — and how to audit for them before launch

The Equality Act 2010 protects nine characteristics: age, disability, gender reassignment, marriage and civil partnership, pregnancy and maternity, race, religion or belief, sex, and sexual orientation. The obvious questions to avoid are clear. The traps are subtler:

• "Do you have a full driving licence?" is lawful if driving is a genuine occupational requirement (GOR), but unlawful if it operates as a proxy for excluding candidates from postcodes that correlate with ethnicity.
• "Are you available for a 6am start?" is lawful on its face but creates indirect sex discrimination risk if offer data shows it disproportionately filters out female candidates with childcare responsibilities.
• Salary history questions are not per se unlawful in the UK, but if they systematically suppress offers to women in equivalent roles they feed an equal pay claim.

Pre-launch audit process: list every question the agent can ask, including clarification branches; identify the protected characteristic most plausibly affected by each; run a GOR test; then after the first 200 calls, segment disqualification rates by observable proxies (name-based gender inference, postcode deprivation decile) and check for disproportionate outcomes.

The Equality and Human Rights Commission's employment statutory code of practice contains worked examples on pre-employment enquiries that map directly to the decisions you will face at step three.

Candidate experience design: call tone, opt-out handling, and re-engagement after a disqualifying result

Tone. Use a neutral UK-English ElevenLabs voice at 1.05x speed — fast enough to feel efficient, slow enough to be intelligible for candidates whose first language is not English. Avoid overly warm delivery; candidates who receive a disqualification find retrospective warmth condescending.

Opt-out handling. Any candidate who says "human" or presses 0 must reach a live answer within 30 seconds or receive a committed callback time within 2 business hours. Routing opt-outs into a full IVR tree is a fast route to Trustpilot complaints.

Post-disqualification close. The closing script confirms what the candidate did not qualify on (availability, salary, or right to work — never "you failed the screening"), the data retention period from the agency's privacy notice, and that they can re-apply when circumstances change. Hold the rejection email for 4 hours after the call, giving consultants a window to override the agent verdict before formal notification is sent. In our deployments the override rate runs at 3–4%, mostly right-to-work edge cases involving post-Brexit settled status documentation.

For broader call UX patterns, the voice agents for inbound customer service post covers opt-out flow design in more depth.

Scheduling and no-show handling in an outbound screening flow at 30+ calls per day

Our observed first-attempt answer rate for UK recruitment screening calls is 34–41%. The retry logic we run:

• Attempt 1: 9am–11am on the business day after application submission
• Attempt 2: 3pm–5pm, 24 hours after attempt 1 (voicemail drop on this attempt only)
• Attempt 3: 10am next business day with an SMS pre-notification sent 30 minutes prior

After three unanswered attempts within a 72-hour window, the candidate status moves to no_contact. A consultant reviews no_contact records weekly and extends the window for high-priority roles.

No-show rate for confirmed callback slots (candidate replied "yes" to the SMS pre-notification): 18%. Unconfirmed outbound calls no-showed at 28% — a 10-percentage-point improvement from candidate-confirmed slots alone. Day-by-day and hour-by-hour answer rate data is in our outbound call timing analysis for UK voice agents.

What changed in 2025–2026: AI scoring layers in Greenhouse and Bullhorn reducing manual screen volume

Two platform changes have shifted the landscape in the past 12 months.

Greenhouse launched its AI Candidate Scoring layer in late 2025, which scores applications against job requirements before any human or agent interaction. In our clients' Greenhouse instances this reduces the inbound screening queue by 15–25% — the lowest-scoring applications are auto-declined at the application stage, so the voice agent only calls candidates who cleared an initial threshold. This means each call has a higher prior probability of qualifying, which cuts telephony cost per hire.

Bullhorn extended its Automation module with a native voice-call scheduling capability in Q1 2026. For straightforward temp roles requiring only an availability confirmation, it places those calls directly from the ATS. For agencies already deep in Bullhorn, evaluate it before commissioning a custom stack.

One important counterpoint: a 2024 Ada Lovelace Institute report on algorithmic hiring found that automated scoring systems disproportionately disadvantaged candidates with non-linear career histories and those from underrepresented groups. If you layer AI application scoring with a voice screening agent, your DPIA must address compound automated decision-making — two filters in sequence without human review at either stage is harder to defend under Article 22.

Good / Bad / Ugly: three candidate screening deployments and the compliance risk each one created

Good — light-industrial temp agency, 40 calls/day, Vincere. Clean GOR-based question set: right to work, shift availability, site location within 10 miles. DPIA completed before launch. Consent script reviewed by an employment solicitor. Disqualification rates reviewed monthly against gender and postcode data; no disproportionate outcome detected in eight months. Cost: £0.11 per completed screen, down from £4.80 per consultant-run screen.

Bad — professional services recruiter, 15 calls/day, Greenhouse. The agent brief included "gaps in employment history exceeding six months" as a disqualification gate. This is not a GOR for professional roles and correlates with parental leave, carer responsibilities, and health absence — all of which map to protected characteristics. We flagged it before launch and the question was removed. It had cleared two internal sign-offs without anyone raising the issue. The fix took 90 minutes; the exposure had been real.

Ugly — national care staffing agency, 80 calls/day, Bullhorn. The agent went live for three weeks asking candidates whether they had "full availability without caring responsibilities at home" — a direct breach of the Equality Act on sex, disability, and carer grounds. The question was lifted verbatim from a legacy paper screening form. The agency received a formal ICO complaint; the agent was suspended, a retrospective DPIA commissioned, and 1,200 candidate records reviewed manually. The rebuild and legal review cost significantly more than the original build.

The failure pattern across all three is identical: the technology worked; the question set was the problem. Build your compliance audit into the brief, not the post-launch retrospective.