Appendix A

Glossary

TermDefinition
Adverse impactA substantially different rate of selection that works to the disadvantage of a group. Also called disparate impact.
AgentAn AI system that takes actions through tools, often over several steps.
Applicant tracking system (ATS)Software for managing job openings and candidates.
AugmentA mode of use in which AI drafts or analyses and a person reviews and decides.
Automated decision system / toolLegal terms (with varying definitions) for software that makes or substantially assists decisions about people.
Automation biasThe tendency to over-trust automated output.
BaselineThe value of a measure before a change, used for comparison.
Bias auditA test of whether a tool’s outcomes differ across demographic groups.
ChatbotSoftware that converses with users, by script or by language model.
Containment rateShare of assistant conversations resolved without handoff to a person.
Context windowThe amount of text a model can consider at once.
Data protection impact assessment (DPIA)A documented assessment of privacy risks, required under GDPR-style laws for high-risk processing.
DeployerUnder the EU AI Act, an organisation that uses an AI system (as distinct from the provider that develops it).
eNPSEmployee Net Promoter Score: % promoters minus % detractors.
Few-shot promptingIncluding examples in a prompt for the model to imitate.
Four-fifths ruleA rule of thumb: an impact ratio below 0.80 indicates possible adverse impact.
Generative AIAI that produces new content such as text or images.
GroundingSupplying source material and instructing the model to answer only from it.
HallucinationFluent output that is false or unsupported.
Hiring managerThe manager to whom a new hire will report.
HRBPHR business partner: an HR adviser assigned to a business unit.
HRISHuman resources information system: the system of record for employee data.
Human in the loopA person reviews, approves or can stop an automated process.
Impact ratioOne group’s selection rate divided by the highest group’s selection rate.
Instructions (system prompt)Standing directions an assistant follows in every conversation.
TermDefinition
KPIKey performance indicator.
Large language model (LLM)A model trained on large amounts of text to generate language.
OnboardingThe process of integrating a new hire, typically over 90 days.
Personal dataInformation relating to an identifiable person.
PromptThe input given to a model.
ProxyA variable that stands in for a protected characteristic, such as postcode for race.
Regression testingRe-running a full test set after any change.
RequisitionAn approved request to fill a position.
Retrieval-augmented generation (RAG)Finding relevant passages in documents and supplying them to the model to answer from.
Selection rateNumber selected divided by number who applied, for a group.
Task (Zapier)A successfully completed action, counted for billing.
Thematic analysisGrouping qualitative comments into themes.
TokenThe unit of text a model processes.
TriggerThe event that starts an automation.
Turnover rateSeparations divided by average headcount, as a percentage.
Voice cardA short description of an organisation’s writing style, used in prompts.
VoicebotA conversational AI system that works by speech.
ZapAn automated workflow in Zapier.