Every technical term you'll hear in an AI hiring conversation, defined in plain language. Organized alphabetically so you can pull it up during a vendor call and find what you need fast.

Each entry opens with a one-sentence definition, followed by brief context and, where relevant, a note on how the term applies in hiring specifically. The terms marked "(AgentR vocabulary)" are either coined by or most clearly defined within the Academy — use them precisely, because most of the industry still doesn't.

Agent (AI agent)

"An AI agent is a system that takes autonomous action toward a goal using available tools, making its own decisions about what to do next within defined boundaries." Unlike a chatbot (which waits for each instruction), an agent operates continuously until its goal is achieved or it needs to escalate to a human. In hiring, common agents include screening agents (read resumes, rank candidates, explain reasoning), sourcing agents (identify passive candidates, draft outreach, handle responses), and scheduling agents (coordinate interview panels without human intervention). The three ingredients of any real agent are a goal, a set of tools, and autonomy within bounded rules.

Agentic AI

"Agentic AI is a category of AI systems designed to take actions rather than just produce responses." The term describes a shift in AI deployment: from tools that need to be invoked turn-by-turn (type a prompt, get a response) to systems that operate across multiple steps to achieve an objective. In hiring, "agentic hiring" refers to deployments where multiple agents coordinate across stages — sourcing, screening, scheduling, interview support — with humans at the decision points. The phrase is increasingly used in vendor marketing; verify that a product actually has agentic capability by asking what decisions it makes without human input.

Applicant Tracking System (ATS)

"An Applicant Tracking System (ATS) is software that stores, parses, and filters job applications for an organization." The first modern ATS (Taleo) launched in 1999, and the core logic of most systems since — parse resume, extract keywords, rank by keyword density — has not meaningfully changed. Common ATS platforms include Workday, Greenhouse, Lever, iCIMS, and SAP SuccessFactors. The ATS was designed for an era of low application volume and hand-written resumes. It performs poorly under modern conditions (250+ applications per posting, AI-written resumes) because its filtering logic measures keyword presence rather than candidate fit.

ATS bloating (AgentR vocabulary)

"ATS bloating is the practice of inflating a resume with keywords and AI-generated content specifically designed to game the Applicant Tracking System's keyword-matching layer." Candidates use AI to mirror the job description's exact language, ensuring the resume ranks highly regardless of whether the candidate's actual experience matches. ATS bloating isn't a moral failing by candidates — it's a rational response to a system that rewards keyword density over capability. The phenomenon is the core reason keyword-based screening has become unreliable: the filter is now measuring how well the candidate used AI, not how well they fit the role.

Behavioral Anchor

"A behavioral anchor is a specific, observable description of what a given rating (e.g., 1, 2, 3, 4) looks like for a particular interview question." Instead of rating a candidate "good" or "bad," an interviewer rates against pre-defined descriptions: '3 = walks through the stakeholder's position with sophistication, describes specific actions taken, names at least one trade-off.' Behavioral anchors are what separate real structured interviewing from superficially-structured interviews that are just consistent question sets. Without anchors, scoring remains subjective and inter-rater agreement collapses.

Bias (algorithmic)

"Algorithmic bias is systematic unfairness in the outputs of an AI system, typically because the training data or the model architecture reproduces patterns of discrimination from the historical data it learned from." Three main channels in hiring: training data bias (models learn from hiring records that reflect past discrimination), proxy variables (the model uses a correlate of a protected characteristic, like zip code), and feedback loops (the model's decisions shape who gets hired, shaping future training data). Mitigation requires auditing inputs, auditing outputs across protected classes, and continuous monitoring — point-in-time audits decay as models retrain.

Boolean Search

"Boolean search is a method of searching databases using logical operators (AND, OR, NOT) to combine or exclude keywords." Standard recruiter technique on LinkedIn, Google, and candidate databases — for example, `("data engineer" OR "analytics engineer") AND python AND -manager`. Effective at finding candidates who match expected title and keyword patterns; ineffective at finding candidates whose capabilities don't match those patterns. AI-assisted sourcing uses AI to generate Boolean strings for non-obvious adjacent roles, broadening the pool to candidates competitors miss.

Career Patterns (AgentR vocabulary)

"Career patterns are the signals in a candidate's work history that reveal fit beyond keyword match — trajectory, intentionality, adaptability, industry fit, and context fit." AgentR's platform analyzes 30+ such patterns across every application. Career pattern analysis is what reasoning-based AI makes possible at scale: reading the 'shape' of a career rather than the vocabulary on the resume. Strong careers with non-obvious paths — career pivoters, high-adaptability generalists, intentional outsiders — surface under pattern analysis and get rejected under keyword filtering.

ChatGPT

"ChatGPT is a consumer-facing conversational AI product from OpenAI, launched in November 2022, that produces text responses to user prompts." ChatGPT popularized generative AI in the workplace and is frequently the user's mental model of "AI." It's important to note what ChatGPT is and isn't: it's a chatbot (responds to prompts), not an agent (takes autonomous action). ChatGPT by itself doesn't screen resumes, call candidates, or schedule interviews — it produces text. Hiring-specific AI platforms may use the same underlying models (GPT, Claude) but wrap them in agent architectures that can act.

Context Window

"A context window is the maximum amount of text a language model can process at one time, measured in tokens (roughly 0.75 words per token)." Modern models have context windows ranging from 8,000 to over 1,000,000 tokens — enough to hold entire books. For hiring, context window matters because analyzing a resume, a job description, a hiring manager's notes, and a company's historical hiring data simultaneously requires the model to fit all of that in context. Systems with small context windows have to compromise — analyzing resumes without the JD context, or the JD without the role history.

Explainability

"Explainability is the ability of an AI system to surface, in human-readable form, the reasoning behind a specific decision or output." In hiring, this means the system can answer: "Why was this candidate ranked above that one? Which factors contributed to this score, and how were they weighted?" Explainability is a regulatory requirement under the EU AI Act (effective August 2026 for high-risk systems) and implied by US state laws giving candidates the right to request the basis for AI-influenced decisions. Deep learning systems that produce scores without explanation fail this requirement. Reasoning-based systems that produce natural-language rationale alongside outputs are structurally easier to audit.

Fine-tuning

"Fine-tuning is the process of taking a pre-trained AI model and further training it on a smaller, task-specific dataset to improve performance on that specific task." In hiring, a general-purpose language model might be fine-tuned on recruiting-specific data — resumes, JDs, hiring outcomes — to perform better on hiring tasks than a generic model would. Fine-tuning is expensive and requires quality training data. Most modern AI hiring tools don't fine-tune large models; they use general-purpose models with careful prompting and reasoning layers built on top.

Generative AI (GenAI)

"Generative AI is a category of AI systems that produce new content — text, images, code, or audio — in response to prompts." The term distinguishes these systems from earlier AI that classified or predicted without generating. Generative AI entered hiring from two directions simultaneously in 2022–2023: candidates started writing resumes with it, and screening tools started reading resumes with it. The result is an AI-vs-AI dynamic where both sides optimize for each other while the human signal — whether the candidate can actually do the work — gets washed out unless screening evolves beyond keyword matching.

Hallucination

"Hallucination is the tendency of AI models to produce output that is fluent and confident but factually wrong." In hiring, a hallucination might be a model claiming a candidate has experience they don't, inventing a company the candidate never worked at, or citing a regulation that doesn't exist. Hallucination rates have dropped substantially in modern models but not to zero. Mitigation requires grounding (giving the model specific source material to draw from), verification (cross-checking outputs against source data), and human review at consequential decision points.

High-risk AI System

"A high-risk AI system, under the EU AI Act, is any AI system whose use could significantly affect people's rights, safety, or livelihoods — a category that explicitly includes AI used for recruitment, candidate evaluation, and employment decisions." High-risk systems face obligations including risk assessments, technical documentation, bias testing, human oversight, logging, and transparency disclosures to candidates. Full obligations apply from August 2, 2026 (though the Digital Omnibus proposal may push this date). Penalties for violations: up to €15 million or 3% of worldwide annual turnover. HR and recruitment AI is one of the most consistently cited high-risk categories in regulatory discussion.

Humans + Agents (AgentR vocabulary)

"Humans + Agents is a design philosophy where neither humans nor AI agents operate alone — each does what it does best, coordinated through deliberate handoffs." Agents handle bounded tasks at scale: screening batches of resumes, generating structured outreach, coordinating schedules. Humans handle judgment under ambiguity, relationship work, negotiation, and final decisions. The phrase describes the specific orchestration model AgentR's platform is built on: not AI-only (which fails on judgment) and not human-only (which fails on scale), but a deliberate division of labor. See the agent orchestration lesson for what this looks like in practice.

Large Language Model (LLM)

"A Large Language Model (LLM) is a type of AI trained on massive text datasets to predict and generate natural language." Examples include GPT-5 (OpenAI), Claude (Anthropic), and Gemini (Google). LLMs are the underlying technology powering most of what people call "AI" in hiring — from chatbots to screening tools to interview agents. An LLM by itself is not a hiring system; it's a capability that hiring systems are built on top of. Understanding this distinction helps when evaluating vendor claims: "AI-powered" often means "calls an LLM behind the scenes," which is increasingly table stakes rather than a differentiator.

Machine Learning

"Machine learning (ML) is a broad category of AI techniques where systems learn patterns from data rather than being explicitly programmed with rules." Most modern AI is machine learning, including LLMs. In hiring, ML is used to score resumes based on patterns in historical hiring data, predict candidate fit from profile signals, and identify retention risk. The quality of ML outputs depends entirely on the quality of training data — biased inputs produce biased outputs. This is why auditing training data is a core part of ethical AI deployment.

Non-standard Talent (AgentR vocabulary)

"Non-standard talent refers to qualified candidates whose resumes don't fit the pattern a keyword-based ATS was designed to recognize — career pivoters, candidates from adjacent industries, people with non-linear paths, and high-adaptability generalists." These candidates typically have the capability the role requires but lack the specific vocabulary the filter is looking for. They get rejected at the keyword-match layer. Reasoning-based AI surfaces them by reading career story rather than matching strings. In most pipelines, non-standard talent is where the best hires are hiding — because traditional filters systematically miss them, the competition for these candidates is lower.

Orchestration

"Orchestration is the coordination of multiple AI agents across a workflow, including the handoffs between agents, the escalation points to humans, and the feedback loops that let the system improve." The term distinguishes orchestrated deployments from isolated automation (single rule, single trigger) or standalone AI (one model doing one task). Hiring orchestration typically involves screening, sourcing, scheduling, and interview agents coordinated so that output from one feeds into another, with human decision points at consequential moments. Good orchestration design is increasingly the TA leader's most important skill.

Paper Tiger Hire (AgentR vocabulary)

"A Paper Tiger is a candidate whose resume looks like an A Player but who performs like a B or C Player on the job." They pass every filter, clear every phone screen, and handle the interview loop with rehearsed confidence. Then they start the role and the gap between the resume and the reality opens up. Paper Tiger hires are the most expensive kind of hire a company can make, because they survive the entire process before revealing themselves. Historically they required 3+ rounds of interviews to catch; reasoning-based AI can flag most of them before the first phone call by checking whether resume claims hold up under verification. Companies using this approach report 59% fewer Paper Tiger hires.

Prompt Engineering

"Prompt engineering is the practice of crafting the input to an AI model to get useful output." Four elements of good prompts: context (who this is for, what the situation is), constraints (boundaries on the output), examples (what good looks like), and format (how the output should be structured). Weak prompts produce generic outputs; strong prompts produce specific, usable outputs. For recruiters, prompt engineering is the single highest-leverage AI skill to develop — not because it's technical, but because it compounds: every hour spent getting better at it pays back across all future AI-assisted work.

Quality of Hire

"Quality of hire is a metric that measures how well a hired candidate performs in the role over time, typically rated at 6 and 12 months post-hire." It's the most important hiring metric and the least commonly measured — partly because collecting the data requires discipline (hiring manager ratings at regular intervals) and partly because the data is noisy at the individual level. At the cohort level, quality of hire is hugely informative: it reveals which sources, hiring managers, and selection methods are producing strong hires versus weak ones. Organizations that track quality of hire make different decisions than organizations that optimize on time-to-hire alone.

Reasoning Models

"Reasoning models are AI systems designed to work through problems step by step, explaining their logic, rather than producing single-shot outputs." Examples include GPT-5's reasoning mode, Claude's extended thinking, and DeepSeek-R1. In hiring, reasoning models are what make pattern-based resume analysis feasible — they can articulate 'why' a candidate ranks where they rank, citing specific evidence from the resume. This is both more accurate than score-only models and more auditable, which matters increasingly as explainability becomes a regulatory requirement.

Retrieval-Augmented Generation (RAG)

"Retrieval-Augmented Generation (RAG) is an AI technique where the model retrieves relevant information from a specific knowledge base before generating a response, rather than relying only on its training data." In hiring, RAG is used to let models reason over company-specific data — past hires, role definitions, hiring manager preferences — that weren't in the model's training set. RAG makes outputs more accurate (grounded in specific source material) and more auditable (the retrieved sources can be shown). Most modern hiring AI uses some form of RAG.

Reinforcement Learning from Human Feedback (RLHF)

"Reinforcement Learning from Human Feedback (RLHF) is a training technique where humans rate AI outputs and the model learns to produce responses rated higher." RLHF is the technique that made modern chatbots like ChatGPT and Claude feel conversational and helpful — earlier models produced technically competent but often unhelpful outputs. In hiring tools, RLHF influences things like the tone of AI-generated outreach, the structure of screening question output, and the kind of explanation the model produces alongside scores. It's mostly invisible to the end user but shapes the entire user experience.

Resume Parsing

"Resume parsing is the process of extracting structured data (name, titles, employers, dates, skills) from the unstructured text of a resume." Every ATS does resume parsing; quality varies significantly. Poor parsing misreads titles, misses dates, or fails on non-standard formats — which then cascades into worse filtering downstream. Modern resume parsing uses machine learning rather than rule-based extraction and handles formatting variation better than systems from even five years ago. Still, "the resume parsed correctly" is a real failure mode when troubleshooting why a qualified candidate didn't appear in a search.

Semantic Search

"Semantic search is a search method that matches the meaning of a query against documents, rather than exact keyword matches." Where keyword search would only surface resumes containing "Python," semantic search also surfaces resumes containing related concepts — "backend engineer," "Django," "data pipelines" — that suggest Python capability. Semantic search is a step up from keyword matching but still falls short of full reasoning: it understands that two terms are related, but it doesn't necessarily understand whether a candidate's career actually demonstrates the capability. Most modern "AI-powered" ATS tools are doing semantic search, often marketed as if it's reasoning.

Sourcing

"Sourcing is the proactive search for candidates — typically passive candidates who aren't actively looking — rather than waiting for applications." Sourcing is traditionally done through LinkedIn Recruiter, candidate databases, and referral networks using Boolean searches. AI-assisted sourcing adds the ability to find adjacent-role candidates (who don't match obvious keyword patterns), identify passive signals of openness to move, and personalize outreach at scale. The distinction between sourcing and inbound recruiting matters because the candidate pool, messaging, and conversion dynamics are all meaningfully different.

Structured Interview

"A structured interview is an interview where every candidate for the same role is asked the same consistent questions, scored independently against pre-defined behavioral anchors, and compared through a structured debrief process." Structured interviews have operational validity r = .42 versus r = .19 for unstructured interviews (Sackett et al. 2022) — more than double the predictive power for job performance. Structured interviews were historically expensive to build (question sets + scorecards + debrief templates took hours per role); AI now builds the whole kit in minutes, removing the last practical argument for running unstructured loops.

Training Data

"Training data is the dataset an AI model learns from during its training phase." For hiring AI, training data typically includes resumes, job descriptions, hiring outcomes, and various forms of recruitment text. The quality and composition of training data determines what the model can do and what biases it inherits. A model trained primarily on tech-industry hiring data will perform worse on retail or healthcare hiring. A model trained on historical data that reflects past hiring discrimination will reproduce that discrimination unless specifically corrected for. Auditing training data is a core compliance obligation under most modern AI hiring regulations.

Zero-shot

"Zero-shot is a capability of modern AI models to perform a task they were not specifically trained on, with no examples provided in the prompt." Example: a general-purpose language model can write an interview scorecard even if it was never trained specifically on interview scorecards, because its broad training gave it enough context to reason about the task. Zero-shot capability is what makes modern AI useful for the wide variety of hiring tasks — the same model can audit a JD, draft outreach, analyze a resume, and design an interview without requiring task-specific training for each one.

Spot a term missing from this glossary? The Academy evolves with its readers — send suggestions to [agentr.global/academy](https://agentr.global/academy)

Next: Lesson 23 — Where Hiring Goes From Here