Glossary
The language of AI in HR and talent, in plain terms
What each term means, and why it matters when you are the one deciding. Only the words AI brought with it, across finding, hiring, developing and keeping people. The vocabulary that was already there is covered well enough elsewhere.
92 terms, A to Z
A
Agentic AI
AI that carries out a sequence of actions towards a goal rather than just answering a question. The difference between a chatbot and an agent is that the agent does the work.
AI Act (EU)
See EU AI Act.
AI-generated application
A CV or cover letter written or heavily polished by generative AI. Now common enough that document quality no longer signals candidate quality, which is pushing screening towards evidence of actual capability.
AI Interviewer
An AI that conducts a structured interview by voice or chat, asynchronously, and produces a scored report linked to what the candidate said. Candidates must be told they're speaking with an AI.
Try it yourselfAI inventory
A list of every AI system touching hiring and HR: who owns it, what it decides, and which obligations attach to it. No article of the AI Act names it, and it is still the first thing an auditor, a works council or a customer's procurement team asks to see. Nobody assembles one quickly under pressure.
Read our positionAI literacy
EU AI Act Article 4: anyone deploying AI must take measures to support AI literacy among the staff using it. In force since 2 February 2025 and softened from a stricter wording by the 2026 Digital Omnibus, though national authorities only began supervising it on 3 August 2026. In practice: training records, a written policy, and an inventory of what AI you use.
Test your setupAI screening
Using AI to read and score every application against the role's criteria instead of reviewing each one by hand, with a stated reason per criterion rather than a single opaque number. Keeping a human in the loop does not take you outside the EU AI Act, but it is what makes the obligations ordinary: a reviewer who can see the reasoning and overturn it is most of what oversight asks for.
How we do thisAlgorithmic bias
Systematic skew in a model's outputs, usually inherited from patterns in its training data rather than written in deliberately. If past hiring favoured one group, a model learning from it will too.
Audit log
An immutable, timestamped record of what a system did and what a human changed - including who overrode which AI score, and when.
Read our positionAutomated decision-making
A decision made about a person with no meaningful human involvement. Under GDPR Article 22 this is prohibited in principle for decisions with significant effects, such as a job rejection or a promotion that never happens, unless a narrow legal gateway applies. The practical answer is a design question rather than a legal one: the system recommends and a person decides, with the time, the information and the authority to disagree.
Automated sourcing
AI that finds candidates for a role without a recruiter running the search manually, usually triggered when a job goes live.
How we do thisAutomation bias
The human tendency to over-trust a machine's recommendation and stop scrutinising it. The main reason "human in the loop" can become a rubber stamp, and something the AI Act's oversight requirements explicitly ask you to counter.
B
Benchmarking
Calibrating an AI tool's criteria and thresholds against your own past hires rather than accepting generic defaults. Powerful and double-edged: past hires carry past bias, and too few of them teach the model your noise rather than your standard.
Bias audit
A structured, repeated test of whether a selection tool produces different outcomes for different groups. Not a one-time certificate - models and applicant pools both drift. On your own funnel data and on a set rhythm is the version that holds up. Once, at procurement, is the version that does not.
Read our positionBlind screening
Hiding names, photos and other identifying details during evaluation. Easier to apply consistently when a system does it, though field evidence on the effect is more mixed than commonly assumed.
Boolean search
Combining keywords with AND, OR and NOT to filter a candidate database. Worth knowing precisely for what it cannot do: it finds only people who used your exact words, which is the gap semantic search exists to close.
C
Competence mapping
Building a picture of the skills your organisation actually has. The practice is decades old and used to mean a workshop, a spreadsheet and a snapshot that covered a fraction of roles and went stale within the year. What AI changed is the cost of keeping it current, so it can cover everyone and stay true. Internal mobility and workforce planning were always waiting for that.
How we do thisConfidence score
How certain a model is about its own output, as distinct from how good the candidate is. A high fit score with low confidence means "look at this yourself".
Conformity assessment
The EU AI Act procedure for demonstrating a high-risk system meets its requirements before going to market. Your vendor's job, but you should ask to see it. A vendor who cannot show you theirs has answered the question.
Context window
How much text a model can consider at once. Determines whether a system can weigh a full CV, transcript and job description together or has to work in fragments.
D
Data enrichment
Expanding a thin candidate record with signals from public sources: skills, seniority, tools, tenure, context. Matching quality is capped by data quality.
How we do thisData residency
Where your data is physically stored and processed, and under which jurisdiction. Usually the first question procurement asks about an AI vendor in Europe. For people data the answer worth having is that it is processed in the EU and stays there, sub-processors included, and that someone will put that in the contract.
Read our positionDeepfake candidate
Someone using AI-generated video, voice or identity documents to impersonate a different person in a hiring process - or to have a proxy sit the interview. A fast-growing fraud vector, and the reason identity verification is moving earlier in the funnel. Chasing the artefacts is a losing game as the models improve. Verifying identity once, early, and keeping the record of it, is the part that keeps working.
Deployer vs provider
The EU AI Act's two roles. The provider builds or sells the AI system; the deployer uses it under their own authority. If you buy AI screening, you are the deployer and carry your own obligations - a point many employers miss.
Test your setupDPA (Data Processing Agreement)
The GDPR-required contract setting out how an AI vendor may process candidate data on your behalf.
Read our positionDPIA (Data Protection Impact Assessment)
The GDPR assessment you run before processing that is likely to be high risk to the people involved. AI screening of applicants is a standard example, so this is usually the first document a data protection officer asks for.
Read our positionDSAR (Data Subject Access Request)
A candidate exercising their GDPR right to see what you hold on them. Where AI has scored them, the score and the reasoning behind it are part of the answer, so a tool that cannot show its working turns a routine request into an escalation.
Read our positionE
EU AI Act
The EU's regulation of artificial intelligence. AI used to filter applications and evaluate candidates is classed as high-risk (Annex III), bringing obligations for both the vendor and you as the employer deploying it. The compliance deadline moved to 2 December 2027 under the Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026). Most of what it asks for is documentation, oversight and record-keeping rather than different technology, so the work is usually organisational rather than technical.
Read our positionEvaluation (evals)
Structured testing of an AI system against known-correct examples, to measure how well it actually performs before and after you deploy it. The difference between a vendor's claim and a checkable number.
Explainability
Whether a system can show why it reached a conclusion, in terms a person can follow and challenge. An unexplainable score is unusable in hiring. Usable means a person can see which requirement drove the score and the words behind it, and can disagree with either.
F
Fairness metrics
The specific statistical definitions used to test whether a model treats groups equally - selection rates, error rates, calibration. They conflict with each other mathematically, so which one you optimise for is a policy choice, not a technical one.
Fine-tuning
Further training a general model on specific data to specialise it - for a domain, a role family, or an organisation's own patterns. Where "recruitment-trained AI" claims come from, and worth asking what the tuning data actually was.
Fit score
A single number expressing how well a candidate matches a role, ideally broken down into the criteria behind it. Useful for sorting, not as a verdict.
How we do thisFlight risk
The likelihood an employee leaves within a given period, estimated from engagement signals rather than guessed at.
How we do thisFoundation model
A large general-purpose AI model - GPT, Claude, Gemini - that specific tools are built on top of. Worth asking whether your candidate data is used to train one.
G
GDPR
The EU's data protection regulation, governing how candidate data is collected, stored, used and deleted - including how long you may keep a rejected applicant's record for future AI matching.
Read our positionGenerative AI
AI that produces new content rather than only classifying or predicting. Writes job ads and outreach - and, on the candidate's side, CVs and cover letters.
Governance (AI governance)
The policies, roles and review processes determining how AI is approved, monitored and overridden in your organisation. Under the AI Act this, not the software, is the deliverable.
Read our positionGround truth
The known-correct answer a model's output is measured against. In hiring it's genuinely hard to get: you only ever see how the people you hired worked out, never the ones you rejected.
Guardrails
Constraints placed around a model to stop it doing certain things - going off-topic, asking prohibited questions, making a decision it isn't allowed to make on its own.
H
Hallucination
When a model states something confidently that isn't true - an invented employer, a qualification that was never there. The argument for evidence-linked scoring: if every claim points to a line in the source, hallucination has nowhere to hide.
High-risk AI system
The EU AI Act's classification for AI that can significantly affect people's rights or life chances. Annex III point 4 covers recruitment: targeting ads, filtering applications, evaluating candidates, promotion and termination decisions. High-risk means regulated, not forbidden. The classification tells you what you have to document, oversee and be able to explain, not that you have to stop.
Read our positionHuman-in-the-loop
A design where AI recommends and a person decides, able to see the reasoning and override it.
Read our positionHuman oversight
The EU AI Act's Article 14 requirement that a high-risk system be usable under meaningful human supervision - including that the person can understand its output, spot when it's wrong, and stop it. Stronger than simply having a human present. In practice it means the reviewer has the time, the information and the authority to overturn a result, and that you can point at a case where someone did.
Read our positionI
Internal mobility
Filling open roles with people you already employ. The idea was never the hard part: nobody could see the skills sitting in the building, so roles went outside by default and people left to get the job they could have had at home. Matching against a live skills picture is what makes it work at any scale, and the internal candidate is usually the fastest and cheapest hire available.
How we do thisK
Knockout criteria
The hard requirements a candidate either meets or does not: a licence, a work permit, a language. Most teams call them must-haves. Automating a rejection on one is about as defensible as automated rejection gets, because the fact is objective and checkable. It is still a decision with significant effect on a person, so GDPR Article 22 applies and someone has to be able to explain it and reverse it.
How we do thisL
Large language model (LLM)
An AI model trained on very large amounts of text, able to read, summarise and generate language. CVs, applications and interview transcripts are all text, which is why recruitment changed quickly once LLMs arrived.
Lawful basis
The GDPR ground you rely on to process someone's data. Recruitment usually runs on legitimate interest rather than consent, because consent a candidate is not free to refuse is not consent. Worth settling before sourcing starts rather than after a complaint.
Read our positionM
Matching engine
The component that compares a role's requirements against candidate profiles and ranks them by fit.
How we do thisMCP (Model Context Protocol)
An open standard that lets an AI assistant connect to a system and act in it, rather than only being told about it. In hiring it is what allows a general assistant to search your candidates or open a role without a bespoke integration being built first.
How we do thisModel card
A short standardised document describing what a model does, what it was trained on, how it performs and where it shouldn't be used. Ask for one.
Model drift
The gradual decay in a model's accuracy as the world changes around it - new titles, new tools, a shifting market. "Set and forget" is a warning sign. Ask what gets re-measured, how often, and what happens when a number moves.
N
Nurturing campaign
Automated, ongoing contact with candidates you're not hiring right now, so the relationship stays warm for the next role.
How we do thisO
Open-web sourcing
AI agents searching for candidates across the public internet - professional networks, code repositories, community sites - rather than one platform or your own database.
How we do thisOverfitting
When a model learns the noise in its training data rather than the pattern, and performs well on examples it has seen but badly on new ones. Why calibrating on a handful of past hires can backfire.
P
People analytics
Using workforce data to answer questions about hiring, performance, and retention. The discipline AI in HR plugs into, and the reason skills data has become the asset worth building.
Post-market monitoring
The EU AI Act requirement to keep watching how a high-risk system performs after deployment, not just certify it once. For hiring, this is what turns bias testing into an ongoing obligation.
Precision and recall
Two measures of a matching system. Precision: of the candidates it surfaced, how many were actually relevant. Recall: of the relevant candidates out there, how many it found. Improving one usually costs the other.
Predictive validity
Whether a selection method actually predicts job performance. The question every AI scoring claim ultimately rests on, and the one almost nobody measures.
Proactive hiring
Building pipelines and relationships before a vacancy exists - economically viable mainly because automation makes the continuous work cheap.
How we do thisProhibited AI practices (Article 5)
Uses of AI the EU AI Act bans outright. Two matter in hiring: emotion recognition in the workplace, and biometric categorisation to infer protected characteristics. Banned since 2 February 2025 and not deferred. Both are narrower than they sound, and a tool that reads what a candidate wrote or said does neither, though the GDPR and discrimination law still apply to it.
Read our positionPrompt injection
Hidden instructions planted in a document to hijack an AI reading it - for example, white text in a CV telling a screening model to rate the candidate highly. A live risk for anything that reads candidate-supplied files. The defence is a system that treats a document as evidence to be read rather than instructions to be followed, and that shows the wording behind every score, so an inflated one has somewhere to be caught.
Proxy variable
A neutral-looking data point that stands in for a protected characteristic - a postcode, a school, a career gap. How bias re-enters a model that never saw gender or ethnicity at all.
Pseudonymisation
Replacing identifying details with a separately held key. A GDPR safeguard - but pseudonymised data is still personal data.
Q
Quality of hire
How well a hire actually performs and stays, measured after the fact. The only real test of whether AI matching improved anything, and the least measured metric in hiring.
R
Ranked shortlist
A small set of candidates ordered by fit, with the reasoning attached, rather than an undifferentiated list of applicants.
How we do thisRecruitment automation
Software that performs recruitment tasks without human initiation. Rules-based automation does what you told it; agentic automation works out what to do.
Recycling non-hires
Automatically returning strong candidates who weren't hired to a talent pool and re-matching them to future roles. The cheapest source of qualified candidates most companies have.
How we do thisRetention period
How long you keep an applicant's record once the role closes. It is the quiet limit on re-matching people to future roles: a talent pool is only as deep as your retention policy allows, and "forever, in case something comes up" is not a policy.
Read our positionRetrieval-Augmented Generation (RAG)
A technique where an AI retrieves relevant documents - your candidate records, your job descriptions - and reasons over them, rather than relying only on training data. How a tool can answer questions about your candidates and cite evidence.
Risk management system
The EU AI Act's requirement for a documented, continuous process identifying and mitigating the risks a high-risk system poses. The paperwork backbone of compliance.
S
SCIM
A standard for automatically creating and removing user accounts from your identity system. How a leaver actually loses access to candidate data on their last day.
Read our positionSemantic search
Search that matches on meaning rather than exact words, so "built our billing platform" surfaces for "payments engineer". Finds the candidates whose vocabulary differs from yours.
How we do thisShadow AI
Employees using AI tools the organisation hasn't approved - recruiters pasting CVs into a public chatbot, for instance. A GDPR and confidentiality problem, and usually a sign the sanctioned tooling isn't good enough.
Skills-based hiring
Selecting on demonstrated capability rather than proxies like degrees or employer names. Long recommended, rarely done - assessing skills at volume was too expensive before AI.
How we do thisSkills taxonomy
A structured vocabulary of skills and how they relate. Without one, "React", "ReactJS" and "front-end experience" are three unrelated things. The quiet infrastructure behind decent matching.
How we do thisSSO (Single Sign-On)
Logging into applications through your organisation's central identity provider rather than separate credentials.
Read our positionStructured interview
An interview where every candidate gets the same core questions, scored against defined criteria. One of the better predictors of job performance - and exactly what an AI interviewer is good at running at scale.
Try it yourselfT
Talent pool
A segmented, maintained group of previously sourced or interviewed candidates, each carrying a fit score, re-matched automatically against new roles.
How we do thisTechnical documentation
The record the EU AI Act requires a high-risk system to have: what it does, how it was built, how it was tested, how it's monitored. Annex IV sets the contents.
Time-to-shortlist
Days from opening a role to having a credible set of candidates to review. The number automated sourcing moves most directly.
How we do thisTraining data
The data a model learned from. Two questions for any vendor: what was it trained on, and will our candidate data be used to train it?
Transparency obligation
Two distinct EU AI Act duties. Article 50 requires telling people they are interacting with an AI, and has applied since 2 August 2026. Articles 26(7) and 26(11) require informing workers and candidates about high-risk systems, and moved to 2 December 2027. In Sweden the co-determination rules can bite well before either: sections 11 and 38 of MBL apply to a significant change and to engaging an external tool.
Read our positionTrigger-based sourcing
Sourcing that starts automatically on an event - a job going live, a requisition opening, a request arriving by email - rather than on someone remembering to brief a tool.
How we do thisTwo-way integration
An integration that both reads from and writes back to your system, so AI scores and statuses appear in the ATS rather than in a separate tool.
How we do thisU
Unified API
A single integration layer connecting to many ATS and HRIS systems at once, instead of building each connector individually.
How we do thisUnstructured data
Information without a fixed format - CVs, cover letters, interview notes. The majority of what recruitment runs on, and effectively unusable at scale until LLMs.
V
Voice AI
Speech-based AI that listens and responds in real time. What makes a spoken AI interview possible; distinct from analysing tone or emotion, which is prohibited in the workplace under the AI Act.
Try it yourselfW
White-label
Delivering an AI product under your own brand rather than the vendor's, so candidates see you. If AI runs your first-round interviews, it is your employer brand for that moment.
Workforce planning
Forecasting what capabilities the organisation will need and how to get them: hire, develop, borrow or automate. Requires a skills picture most companies only get once AI builds one.
How we do thisX
XAI (explainable AI)
See Explainability.
Y
Yield ratio
The percentage of candidates passing from one stage to the next, cut by source. How you check whether an AI sourcing channel produces candidates that actually progress, or just volume.
How this is kept
Regulatory dates here are the same ones behind our EU AI Act readiness test, and they move when the regulation moves rather than when we get round to it. Every article number and deadline is checkable against the text of the law. If an entry reads as out of date, or plain wrong, tell us and we will fix it.
