How recruitment AI works, without the jargon

You do not need to understand the mathematics, but four ideas explain almost every product.

1. Turning text into numbers (embeddings)

A model converts a résumé, a job description or a profile into a long list of numbers that captures its meaning. Two texts about similar work end up with similar numbers even if they share no keywords. This is why modern matching finds a "growth marketer" when you searched for "demand generation manager", and why Boolean strings are fading.

2. Ranking

Given a brief and a thousand candidates, a ranking model orders them by predicted fit. The prediction is trained on examples: past hires, recruiter shortlists, who replied, who advanced. The quality of a ranking depends entirely on what it was trained to predict. "Looks like people we hired before" is a different target from "will pass a structured interview", and the first one is where bias enters.

3. Large language models

Language models read and write. In recruiting they draft outreach, summarise interviews, answer candidate questions, extract requirements from a messy job description and explain why a candidate ranked where they did. They are strong at language and weak at facts they were not given; a well-built product keeps them grounded in your data.

4. Agents

An agent is a language model given tools (search, a contact database, an email account, a calendar) and a goal, and allowed to take several steps towards it. A sourcing agent, for example, reads a brief, runs searches across several sources, ranks the results, looks up and verifies contact details, writes a personalised message to each person, sends it by email or SMS, waits for replies and reports back. The human sets the brief and reviews the output; the agent does the hours in between.

Putting it together: one sourcing run

1. Brief A recruiter writes two paragraphs: the role, must-haves, nice-to-haves, location, compensation range.
2. Search The agent turns the brief into several searches and runs them across licensed profile data and the open web, gathering a few thousand possible people.
3. Rank Each person is scored against the brief; the top few hundred are kept, with a reason for each.
4. Verify contacts Email addresses and phone numbers are found and checked so messages are deliverable.
5. Outreach A short, specific email (and, where appropriate, an SMS) is written for each person, referencing their actual work. Follow-ups are scheduled.
6. Replies Interested people are handed to the recruiter or pushed into the ATS; the agent keeps a log of what it did and why.