AI for recruiters, A to Z.

An agent is an AI given a goal, a set of tools, and the ability to plan multiple steps. In recruiting: an agent might source candidates, score them, draft outreach, and queue messages for your approval, all without you typing each step.

The EU AI Act phases in through 2026. Most AI used in hiring falls into the 'high risk' category, which requires risk management, bias testing, technical documentation, human oversight, and post-market monitoring.

A bias audit measures the rate at which an AI selects candidates across demographic groups. The 'four-fifths rule' flags any group selected at less than 80% of the most-selected group's rate.

Boolean strings match exact words. They are precise but brittle. The candidate who describes themselves differently than your search will not be found.

Instead of paying the tool a marked-up price for AI usage, you connect your own Anthropic, OpenAI, or other provider keys. You pay providers directly. The tool charges only for the platform.

Deduplication uses a mix of email, phone, name fuzzy-matching, and profile signals to merge duplicate records. Critical when sourcing across multiple data sources.

ChatGPT is one of the most-used AI tools in recruiting. Most recruiters use it for outreach drafting, JD writing, and research. The Pro tier (GPT-4o) handles longer context and more complex tasks.

Claude is increasingly used for screening summaries, structured scoring, and any task that requires reading a lot of context. Some recruiters prefer it for nuanced writing tasks.

Custom workflows let you specify what should happen when. New candidate received? Score them. Score above 80? Notify the recruiter. Etc.

Embeddings are how semantic search works under the hood. The AI converts every candidate profile and every search query into a vector and compares them mathematically.

Enrichment uses third-party data providers to fill in fields you do not have. Common sources: Apollo, Cognism, Contact Out, RocketReach.

EEO compliance requires anonymised demographic capture and audit-ready reporting. Most ATSs offer dedicated EEO survey flows.

Fine-tuning takes a model like GPT-4o and trains it further on your hiring data so it scores the way your team scores. Available on enterprise tiers of most AI products.

The base AI everything else is built on. Foundation models are trained on massive datasets and can be specialised through prompting or fine-tuning.

GDPR requires lawful basis for processing candidate data, retention limits, right-to-be-forgotten flows, and data portability. Most modern ATSs handle this for you.

Hallucinations happen when an AI generates plausible-sounding but false information. In recruiting: an AI might invent a candidate's previous job. Always verify any factual claim by clicking through to the source.

Not a recruiter. The hiring manager owns the role's success and usually has final say on offers. AI workflows often have a 'send to hiring manager' step.

Required by most regulations and good practice anyway. AI drafts, human approves. AI suggests, human decides. Especially important for sends, declines, and any candidate-facing action.

The ICP is the foundation of any AI search. Two sentences max: what does this person do, and what have they done before? Clearer ICP = better shortlist.

Integrations let agents read and write to external systems: Slack, Workday, Gmail, calendar, sourcing APIs like Exa or CoreSignal, your own internal tools.

Most JDs are too long, too vague, and too full of requirements that are not real deal-breakers. AI can rewrite them. Keep the impact at the top.

Recruiting knowledge bases store rubrics, exemplars from past hires, company context. Better knowledge base = more accurate AI scoring.

LLMs predict the next word in a sequence. They are general-purpose, can read and write any text, and are the substrate everything else is built on.

MCP is what makes 'bring your own AI' real. It lets Claude or ChatGPT drive your ATS natively, with no glue code. Vitae exposes every action as an MCP tool.

Single-channel outreach (email only) gets ~5% reply rate. Multi-channel sequencing can lift it to 15-20%. AI helps pick the best channel per candidate.

In effect since 2023. Applies to any role that will be performed in NYC. Requires a bias audit, public posting of results, and candidate notice that AI is being used.

Founded 2015, became the household name in AI in 2022. GPT-4o is the current flagship. Used by most AI recruiting tools either as the default or as one of several model options.

A sequence might be: email day 1, LinkedIn connect day 4, follow-up day 9, value-add day 16. AI writes the drafts; the recruiter approves the sends.

Personas humanise the ICP. 'Léa, the staff engineer who left Stripe to join Mistral' is more useful than 'senior backend engineer with payments experience'.

A clear prompt is just a clear brief. The best recruiters using AI build a personal prompt library of instructions that consistently produce useful output.

More art than science. Patterns: give context first, give examples, ask for structure, ask for reasoning, iterate.

A rubric lists must-haves, nice-to-haves, and deal-breakers, each weighted. Rubrics make AI scoring auditable and consistent across recruiters.

Semantic search reads intent. 'Senior backend engineer with payments experience' matches candidates whose history shows that shape, even if they never used those exact words.

AI sourcing now produces a scored, reasoned shortlist instead of a raw search dump. The recruiter spends time on the top 10, not the top 200.

Signals: recent profile update, new connections to recruiters, posts about job-hunting, tenure crossing 18 months. Stronger signals = higher reply rates.

Silver medalists are the highest-quality re-engagement pool you have. AI workflows can re-score and re-engage them quarterly without manual effort.

Type I assesses controls at a point in time. Type II assesses them over a period (usually 6-12 months). Most SaaS vendors are SOC 2 Type II compliant.

Outbound recruiting. Most senior roles fill through sourcing, not applications. AI is reshaping this category fastest because the brute-force version was always painful.

Required by most enterprise IT. SAML and OIDC are the common standards. SCIM handles user provisioning. Most ATSs offer SSO on mid-tier plans and above.

Your talent pool is your most valuable asset. AI can rescore the whole pool against any new role in seconds, surfacing matches you forgot existed.

The most common headline metric in TA. Industry average: 30-45 days. AI-assisted teams often run 20-30 days for the same role complexity.

An agent's tools might include: search for candidates, score against rubric, send email, book meeting, post to Slack. Each tool is a function the AI can choose to invoke.

A workflow has a trigger (something that starts it), conditions (gates), and actions (what it does). Modern workflow tools let you drop AI judgement into any step.