Screen every applicant in minutes, in their language, on the channel they actually answer
Screening is usually the single largest line in cost per filled shift, and the least differentiated work a recruiter does. The AI Recruiter takes the dialling, the six qualifying questions and the note-taking. Your recruiters take the conversations that need a person.
What this work looks like today, and what changes.
Today, per applicant
- Recruiter pulls the next name from the ATS or a spreadsheet.
- Dials. No answer. Leaves a voicemail. Dials again after lunch.
- Reaches the candidate, asks the same six questions: shifts, location, right to work, licence, pay, start date.
- Discovers the candidate cannot work weekends. Ends the call politely.
- Types a note into the ATS. Often incompletely.
- Repeats for the next 40 names. Qualified candidates wait a day or two for a callback.
With the AI Recruiter
- Applicant arrives from a job board, your website or WhatsApp. The agent contacts them within minutes, by voice or WhatsApp, in their language.
- The agent runs your gates, verifies answers, asks follow-ups when something is unclear.
- Answers, transcript and audio are written to the candidate record, structured, field by field.
- Passes are transferred warm to a recruiter or booked straight into an induction slot.
- Fails and unclear cases land in a recruiter's review queue with the reason attached.
- Retries run across time windows and stop the moment a worker asks.
Step by step: what goes in, what the automation does, where a person decides.
Green steps run on their own. Amber steps are where a human approves or decides, by design.
Job board, website form, referral or an inbound WhatsApp. Any source that reaches your ATS or a shared inbox.
The agent calls or messages in the applicant's language. AI disclosure at the start of every call. Retries by time window, not by hammering.
AutomatedShift availability, distance and travel mode, right-to-work or licence, equipment, language level, pay expectation. Your questions, in your order.
AutomatedStructured answers, transcript and audio attached to the candidate record in Bullhorn, zvoove, your in-house system or a spreadsheet.
AutomatedPasses are transferred warm or booked into induction. Fails and edge cases surface to a recruiter with the reason. Nobody is rejected silently.
Human approvesA qualified, documented shortlist the same day, with recruiter minutes spent only on people worth talking to.
One screen, three panes: the input, the reasoning, the result.
- Shift pattern 06:00–14:00 Mon–Fripass
- Saturdays alternate weeks onlyflag
- Distance 40 min by bus, under 60-min limitpass
- ID and bank account confirmed verbally, upload requestedpass
- Pay expectation within bandpass
- Start date asking now…
- Status
- Qualified · recruiter review (Saturday flag)
- Language
- Hindi · transcript + audio attached
- Availability
- Mon–Fri 06:00–14:00, alt. Saturdays
- Travel
- 40 min · bus · Bhiwandi
- Documents
- Aadhaar, bank: upload link sent on WhatsApp
- Next
- Warm transfer to Rahul (recruiter) · 2 min
Every answer is a field, not a paragraph. A recruiter sees the flag and the reason before saying hello.
A composite of a real screening flow. The agent gathers and verifies. Deterministic rules apply your gates. A recruiter makes the call on anything borderline.
Where the advantage comes from, step by step.
| Step | Today, by hand | With ShiftMargin | Why it matters |
|---|---|---|---|
| First contact | Next-day callback if the recruiter has time. Applicants who applied to five agencies take the first one that calls. | Within minutes, 24/7. Voice and WhatsApp in the applicant's language, with polite retries by time-of-day window. | Speed to first contact is the strongest predictor of who you actually hire. |
| Qualifying questions | Six questions asked from memory, in a different order by each recruiter, with answers summarised in a free-text note. | Your gates, verbatim, every time. Follow-up questions when an answer is vague. Answers stored field by field. | Consistent screening is also fairer screening, which matters under EU AI Act and local equality rules. |
| Documents | "Can you send your ID?" followed by a chase. | Upload link sent in-conversation, checked for legibility, attached to the record. | Onboarding starts during screening instead of after it. |
| Recording the result | Typed into the ATS later, if at all. | Structured write-back plus transcript and audio. Bullhorn or zvoove via API, or a scheduled import where there is no API. | Auditable decisions, and a searchable history of why someone was or was not progressed. |
| Decision | Recruiter judgement, undocumented. | Rules decide pass or flag. A human decides hire. Fails surface with the failing gate, never a silent drop. | Human approval on anything that affects whether a person works. By design, not by policy. |
| Handoff | Recruiter calls back to arrange the next step. | Warm transfer with answers on screen, or a self-booked induction slot. | The recruiter's first minute is the interview, not the intake. |
The components, in plain language, and why each one matters to you.
Everything here is chosen so your ops team can maintain it after we leave. The only pieces we keep maintaining are the ones that move money.
ElevenLabs Agents or Retell class. Visual flow builder, published consent tooling, 29+ languages. You can change a question without calling us. For clients above roughly 20,000 calls a month we run a self-hosted stack that is 25–45% cheaper per minute.
Meta Cloud API via 360dialog in the EU or Gupshup in India and Southeast Asia. Same conversation logic as voice, for workers who do not pick up unknown numbers.
Twilio or Telnyx in the US and Europe, Exotel in India, a regional carrier in Africa. Licensed interconnect only. No grey-route VoIP.
A Haiku-class model runs the dialogue and extracts answers. It costs about two cents per call. It never applies a gate.
Your pass and flag criteria live in versioned code, not in a prompt. Same input, same result, every time.
Bullhorn via your own OAuth credentials, zvoove via its connector, or scheduled export and import. Postgres holds the audit log and every transcript.
What stays human, and what happens when something goes wrong.
US law treats an AI voice as an artificial voice. Disclosure is the first turn of every call, and prior express consent is captured in the application flow. Cold lists are out of scope.
A candidate who fails a gate is surfaced to a recruiter with the reason. The agent never closes a door on its own.
Noisy line, code-switching, an unexpected answer: the agent hands to a human rather than guessing, and the WhatsApp path catches workers who cannot be screened by voice.
Your existing screening process stays live during the parallel run and can take over in one switch. Your clients never feel a bad week.
What it costs by hand, what it costs to run, and where the margin goes.
for 15 loaded minutes in the US. $1.50–3 in India or South Africa. Half of those minutes go to voicemail.
all-in for a four-minute call: voice platform, telephony, transcription and model. 5,000 calls a month runs $1,250–2,000.
Platform choice is a handover decision, not a cost one, until volumes pass roughly 20,000 calls a month. The saving comes from the workflow, not the vendor.
Six weeks to production, behind a manual fallback.
Answer rates, minutes per screen, time to first contact, and the questions your best recruiter actually asks. This becomes the gate list.
Voice and WhatsApp flows, gate rules, ATS write-back, consent capture, escalation thresholds. Tested against recorded real calls.
The agent screens live while your recruiters still do. Every disagreement is reviewed. Fallback stays on.
Minutes per qualified candidate, time to hire, recruiter hours redirected. If the number stops moving, we stop building.
Before they let this near the business.
Will an AI recruiter reject candidates automatically?
Not in our design. The agent gathers and verifies answers, deterministic rules apply your gates, and anything affecting whether a person works is approved by a human and logged. Candidates who fail a gate are surfaced to a recruiter rather than silently dropped.
Is it legal to call candidates with an AI voice?
It depends on the market. In the US the FCC treats AI voices as robocalls, so prior express consent is required: workable for applicants who gave their number with proper disclosure, not for cold lists. We build consent capture into the application flow and adjust the channel mix per market.
Which languages does it screen in?
The managed platforms we deploy publish 29 to 32 languages, including Hindi, Arabic, Urdu, Bengali and Tagalog. Nobody publishes per-language accuracy, so we pilot-test any language that matters to you before committing.
How does it connect to our ATS?
Through the ATS API where one exists. Bullhorn lets your admin issue credentials for your tenant with no partner programme. Where there is no API, we start with scheduled export and import. Legacy systems are normal and are not a reason to delay.
What happens to recruiters whose calls are automated?
The goal is capacity, not indiscriminate headcount reduction. Recruiters move to client conversations, difficult candidates and the exception queue. Where a firm wants to hold headcount flat while growing volume, that is a decision the numbers make explicit rather than one the software makes quietly.
Start with your number, not our demo.
Bring the volumes and the time your team spends on this workflow today. In thirty minutes we will tell you whether it is your most expensive line, and if it is not, which one is.
Book a free 30-minute teardown →