AI recruiting: how the pipeline actually works

Most agencies that call themselves “AI-powered” have bolted a sourcing tool onto a manual desk. The recruiter is still the operating layer. ON3 Works is structured differently: AI is the operating layer that connects the database, the matching logic, the screening process, and the shortlist. The human is deployed at the point where judgement matters most — not spread across every step.

ON3 Works was designed this way from September 2022. It owns its sourcing infrastructure — the ON3 database of 1.7M+ structured profiles and hiron.ai (which ON3 also owns) — and runs the entire pipeline in-house. No rented databases. No third-party matching APIs.

Stage 1 · Keyword and constraint filtering

The pipeline starts with hard constraints: location, salary band, availability, visa status, required certifications, and non-negotiable technical requirements. This is fast, deterministic filtering — it eliminates candidates who cannot fit the role regardless of how talented they are. On a typical search, this stage reduces the candidate pool from hundreds of thousands to a focused long-list of several thousand.

Stage 2 · Vector embeddings — semantic matching

This is where ON3's pipeline diverges from traditional keyword search. Candidate profiles and role requirements are embedded into the same vector space, allowing the system to understand that a “distributed systems engineer” and a “backend platform engineer building event-driven microservices” may describe the same person — even if neither CV uses the other's keywords.

Embeddings capture what candidates have actually done rather than which buzzwords appear on their profile. A candidate who built a high-throughput data pipeline in Kafka and Go will surface for a “streaming infrastructure” role, even if neither “streaming” nor “infrastructure” appears in their CV. Traditional keyword matching misses this entirely.

Stage 3 · Fine-tuned LLM — contextual vetting

The embedding stage produces strong matches. The LLM stage evaluates them in context. ON3's fine-tuned model reads each remaining candidate's full profile against the role requirements and assesses fit beyond what structured matching captures: career trajectory, project complexity, team size and leadership signals, seniority consistency, and whether the candidate's experience pattern actually predicts success in this specific role.

The LLM produces a multi-dimensional weighted score for each candidate, with dimensions weighted to the role's priorities. A DevOps search weights infrastructure and incident response higher; a product engineering search weights end-to-end ownership and user-facing delivery. The weighting adapts to the brief. After this stage, approximately the top 10% of the original pool remains.

Stage 4 · AI screening conversations

CVs are unreliable narrators. “English — professional level” could mean fluent or functional. A project listing “Node.js, React, PostgreSQL” does not tell you whether the candidate wrote the Node.js service or sat two desks away from someone who did. “5 years experience” might mean five years of the same year.

ON3's AI agents conduct live structured conversations with candidates via WhatsApp, phone calls, and LinkedIn to extract the data points that profiles leave ambiguous. The conversations are multi-format — a candidate might receive a WhatsApp message, then a follow-up call — and the AI adapts its questions based on what it has already learned from the profile and earlier stages.

Stage 5 · Senior engineer interview — the human layer

For Atelier placements, a senior domain-matched engineer from ON3's vetted panel — typically 8–15+ years of hands-on delivery experience, matched to the candidate's stack (a backend engineer vets backend candidates, an ML engineer vets ML candidates) — conducts a structured technical interview. The interview focuses on system design, production debugging, code review, and technical ownership — not algorithmic puzzles.

Approximately 25% of AI-screened candidates pass the engineer interview. Interviews are recorded and validated through pretested.io to ensure scoring consistency across the panel. The engineer writes the assessment; the AI validates calibration. Net selectivity: approximately the top 2.5% of the original pool reach the client — with structured scorecards, interview transcripts, reference checks, and salary benchmarking.

The AI screening conversation is not a chatbot asking “tell me about yourself.” It is a structured extraction process, calibrated per role, that resolves the specific ambiguities a profile leaves behind. Every data point becomes a scoreable field in the candidate's structured profile.

• ·Technology ownership — did you write the Node.js service, or was it another team on the project? What did you personally build, and what was the surrounding architecture?
• ·Real experience depth — five years on a CV might be five years of the same year. The AI probes project complexity, team size, and progressive responsibility.
• ·Language proficiency — “English — professional” is meaningless. The conversation itself reveals fluency, technical vocabulary, and communication style.
• ·Motivation & timeline — why are they looking? What's the notice period? Counteroffer risk? This prevents wasted interviews with candidates who won't move.
• ·Salary reality — current compensation, expectations, flexibility. Aligned to ON3's salary data from 1.7M+ profiles so both sides start with market context.
• ·Technical depth — what scale did the system handle? What architecture decisions did you make and why? What would you change?
• ·Availability & logistics — location preferences, remote/hybrid/onsite, relocation willingness, visa status. Hard constraints that waste everyone's time if misaligned.
• ·Culture & communication — how they explain technical concepts. Whether they credit teams or only themselves. Responsiveness and professionalism through the process.

Every signal becomes a structured data point in ON3's database. Profiles are not static documents — they are continuously enriched as new interactions occur and re-scored as new roles arrive.

ON3 Works does not start every search from zero. The ON3 database contains 1.7M+ technical profiles across the UK and Spain, parsed into structured fields: skills, frameworks, seniority, experience history, salary data, location, availability, and signals from prior screening conversations. The database is refreshed daily through AI agents that scan LinkedIn and other public sources, plus hiron.ai (ON3's own sourcing platform).

Critically, profiles are continuously re-scored as new roles arrive. When a client submits a brief, ON3's pipeline does not just search the database — it re-evaluates every relevant profile against the new requirements. A candidate who was a marginal match for a previous role might be a strong match for the current one, because the weighting has changed. This is fundamentally different from static database search, where profiles sit until someone queries the right keyword.

24–48h to a first qualified shortlist. Median time-to-hire under 30 days. 90.7% twelve-month retention. ~2.5% net selectivity.

What the client receives: a shortlist of candidates who have passed three stages of AI filtering, a structured screening conversation, and (for Atelier) a senior engineer interview. Each candidate comes with a multi-dimensional weighted scorecard, interview transcript, reference checks, and salary benchmarking against ON3's 1.7M-profile index.

3.2 client interviews per offer (Atelier). 83% offer-accept rate. 60%+ fill rate on contingent searches, 85%+ on exclusive. 90-day replacement guarantee on every placement.

Agentic — 7% success fee

Stages 1–4 of the pipeline, fully AI-driven. For urgent IC roles, backfills, and searches where speed matters. No engineer interview — the AI screening conversation and weighted scoring are the quality gate. 60-day replacement window. No deposit.

Atelier — 20% success fee

The full five-stage pipeline. Senior engineer interview, structured scorecard, reference checks. For staff, principal, founding, and team-defining hires where the cost of a wrong hire is too high. 90-day replacement guarantee. No deposit.

For worked fee examples and a full side-by-side of what each tier includes, see the pricing page.

The market has three models. They price differently and they work differently.

Each guide below explores a specific angle of AI recruiting — costs, alternatives, and market data. Written for hiring managers and founders making real trade-offs.

• ·AI recruiter cost vs agency fees: what you actually pay (2026)
• ·Toptal alternative for permanent hiring in Europe (2026)
• ·Recruitment agency vs in-house hiring: costs, speed & risk (2026)
• ·Recruitment agency fees UK: full breakdown (2026)
• ·Hiring engineers in the UK — salaries & market
• ·Tech recruitment in London

Frequently asked