Generative AI at Work Glossary from A to Z — Generative AI at Work

Generative AI at Work Glossary — 127 definitions

All terms defined in the guide's 28 sections, listed alphabetically.
Tip: Ctrl + F to search for a term on the page.

Hook: The first 2 lines that decide the click: ask for 10 variants, keep the best.
Audience adaptation: The same substance, rephrased per audience: management (stakes, figures), technical team (precision), clients (benefits, simplicity).
Agent: The next step: an assistant that performs actions (search, create, notify) — see the Copilot 365 guide level 5 for the Microsoft ecosystem.
AI Act: The European AI regulation: prohibited practices (manipulation, social scoring…), high risk (hiring, credit, education…), transparency obligations, substantial sanctions.
Imposed angles: Forcing variety: "5 low-cost ideas, 5 digital ideas, 5 ideas a competitor would have".
Anonymization: Replacing every identifying element (names, amounts, places, exact dates) with fictitious ones — while checking that no cross-reference re-identifies.
Decision tree: Your personal flowchart: data sensitivity → authorized tool → task → method.
GDPR Article 22: The right not to be subject to a fully automated decision with significant effect (credit refusal, application rejection…): a human must stay in the loop.
Custom assistant (GPT, Gem…): A pre-configured assistant: permanent instructions + reference documents + name — reusable and shareable.
AI automation: AI inserted into a flow (Make, Zapier, Power Automate): e.g. each incoming request is automatically summarized and pre-classified.
Devil's advocate: "Attack this plan like a competitor / an unhappy client / a CFO": the stress-test for ideas.
Knowledge base: The documents attached to the assistant: it anchors its answers in them (your procedures, your editorial charter).
Bias: Distortions inherited from training data: stereotypes, overrepresentation of certain viewpoints, cultural angles.
Algorithmic bias: Systematic distortion inherited from data: e.g. a generated job ad that "masculinizes" technical roles.
Personal library: The file where your models and benchmark prompts live: the asset that capitalizes your AI learnings.
Contextual draft: Copilot writes the reply taking the whole thread history into account — always re-read before sending.
Editorial calendar: "Suggest a month of LinkedIn posts for [activity]: 2 posts/week, alternating advice, behind-the-scenes, client proof": assisted planning.
Gray case: Debatable situations: anonymized excerpt, already public information, structure without data — to be settled by the charter (5.5).
Live catch-up: During the meeting: "Summarize what I missed" — your questions to Copilot are private.
Whoever signs verifies: Responsibility for content belongs to whoever publishes it — "the AI said so" is no defense.
AI usage charter: The one-page internal framework: tools, data, verification, transparency, responsibility — signed and known to all.
Living charter: Revised on a fixed date (tools change fast) and fed by the gray cases the team encounters.
Contradiction hunt: "Which elements contradict this claim?": the fix for confirmation bias.
ChatGPT (OpenAI): The mainstream pioneer: versatile, rich ecosystem (custom GPTs), free and paid versions, enterprise offers.
Generated check-list: "Complete check-list for [task], grouped by phase, with frequent oversights": AI excels at exhaustiveness.
Classification: The upfront sorting: public (shareable), internal (organization-only), confidential (restricted access) — each level has its authorized tools.
Claude (Anthropic): Known for long documents, nuanced writing and code; also used in business via API.
Sycophancy: The model's tendency to agree with you: "my plan is good, right?" gets a yes. Fix: explicitly ask for criticism.
Minimal understanding: Have the code commented line by line: you must be able to say what it does, even without knowing how to write it.
Consent: Transcription is visible and announced: a legal requirement and professional politeness.
Context: The situation and audience: "40-employee SME, message for managers, following a quality incident". The richer it is, the less the AI invents.
Counter-argument: "What are the arguments AGAINST this thesis? Who disputes it?": the antidote to confirmation bias.
Convergence: Sorting with criteria: "Rank these ideas by impact and ease, in a 2×2 matrix".
Conversation (thread): A continuous exchange with context memory: the assistant remembers what came before WITHIN the thread, not across threads.
Copilot in Excel: Insights, formulas described in plain language, iterated charts — absolute prerequisite: data in a structured table (Ctrl + L).
Copilot in PowerPoint: Generating a presentation from a styled Word document (Heading 1/2), adding/reorganizing slides, summarizing a deck.
Copilot in Word: Writing a draft, summarizing, rewriting — and referencing your files as sources with "/" (e.g. "summary from /Report").
Cross-model check: Asking two assistants the same factual question: divergence signals shifting ground.
Cutoff date: The time limit of the model's knowledge. Assistants with web search work around it — while citing their sources.
Individual decision: Hiring, sanctioning, evaluating a person: AI never decides — this is also a legal requirement (GDPR art. 22, AI Act).
Multichannel rollout: One pillar content rolled out: LinkedIn post, thread, newsletter, short video script — each in its channel's codes.
Framed script request: Specify: the environment (Windows, Excel, browser), the input (which files/data), the expected output, an example.
Multiple request: Several questions in one prompt: shallow answers to each. Fix: split, one question per message.
Vague request: No goal or angle ("tell me about X"): the AI produces a lukewarm encyclopedia. Fix: specify task + intended use.
Description without data: Giving the STRUCTURE (column names, types) without the values: help without exposing real figures.
Divergence: Producing lots of ideas without judging: "20 ideas, from classic to wacky" — sort afterwards.
Company data: Anything non-public: internal figures, clients, projects, organization — the #1 criterion for tool choice.
Personal data: Anything identifying a person (name, exact role, salary, review, health…): forbidden in unvetted tools — GDPR (see 5.4).
Rights over outputs: Per major tools' ToS, you can generally use commercially what you generate — but the protectable originality of a 100% AI work remains legally debated.
Benchmark example: YOUR best specimen of the task (the perfect template email, the ideal sheet): the investment that gets reused endlessly.
Error explanation: Paste the full error message + context: "what does this error mean and how to fix it?" — the safest use.
Formula explanation: Paste a formula (=INDEX(MATCH(…))…) and ask for a step-by-step breakdown: reverse-engineering inherited files.
Context window: The conversation's "working memory": beyond a certain length, the model loses the start of the exchange.
Few-shot: Providing one or more examples of the expected result in the prompt: the model imitates the demonstrated format.
Bogged-down thread: After too many contradictory back-and-forths, answers degrade: restart from a fresh thread with a consolidated prompt.
Format: The output shape: "150 words, 3 paragraphs, neutral tone" or "3-column table" — the most-often-forgotten block, the most profitable.
Formula described in plain language: "A formula that shows Follow-up if the date exceeds 30 days and status is not Paid": name the exact columns.
Writing templates: Office archetypes: announcement, follow-up, polite refusal, apology, deadline request — each deserves its benchmark prompt.
Imposed glossary: Your official terms: "translate collaborateur as employee, don't translate product names" — terminological consistency.
Google Gemini: Google's equivalent: built into Gmail, Docs and the search engine; strong on multimedia.
Hallucination: An answer that is wrong but stated with confidence: the #1 risk, covered in detail in 0.4 and 5.2.
Legal hallucination: Invented code article, case law or rate: the domain where AI hallucinates with the most confidence — systematic verification against official sources.
High risk: Heavily regulated listed uses — shortlisting candidates and employee evaluation are among them: heavy compliance required.
Generative AI: Family of AI that produces new content (text, image, code, audio) from a prompt, instead of only classifying or predicting.
Style imitation: Generating "in the manner of" a living artist/author: high legal and reputational risk zone — avoid commercially.
Refinement instructions: The classics: shorter/longer, more formal/simpler, as list/table, another angle, add/remove X.
Permanent instructions: The framing written once and for all: role, tone, rules, limits — the equivalent of an engraved RCPF prompt.
Targeted question: "What does this document say about costs?": extracting THE answer without summarizing everything.
Iteration: Refining the answer through successive instructions instead of rephrasing the whole request.
Iteration on error: Code failing? Paste the error: "here is the error, fix it" — the standard assisted-debugging loop.
Plain language: Having regulatory text rephrased in simple terms: excellent use — the rephrasing gets re-read against the source text.
Recipient test: Before sending, re-read asking: "if I received this message, how would I take it?" — AI can run this test for you.
Microsoft 365 Copilot license: Full integration requires the enterprise license (separate from the classic Microsoft 365 subscription).
Tool allowlist: The organization-vetted tools, by authorized data level: the operational core of the charter.
LLM (large language model): The engine of assistants (GPT, Claude, Gemini…): a neural network trained on billions of texts, generating word by word the most probable continuation.
AI literacy (art. 4): The AI Act asks organizations to ensure sufficient AI competence in their teams — this guide contributes to it.
Lived material: What AI cannot invent: your client case, your figure, your expert opinion — the differentiating ingredient.
Microsoft Copilot: The assistant built into Windows and Microsoft 365: its strength is access (with an enterprise license) to your files, emails and meetings — see our Copilot 365 guide.
"Work" mode: Copilot chat connected to the Microsoft Graph: YOUR emails, files, meetings — never beyond your permissions.
"Web" mode: The same chat over general knowledge and the Internet: for anything non-internal.
Template: A fill-in-the-blanks document — meeting-report frame, template email, evaluation grid — created once, served a hundred times.
New conversation: The reflex for each new topic: a clean thread keeps old contexts from polluting answers.
"Enterprise" offer: An assistant version with contractual guarantees (no training on your data, administration): the decisive criterion at work.
Vetted tool: The tool covered by an enterprise contract (Copilot 365, ChatGPT Enterprise…): the list belongs to IT/DPO, not intuition.
Critical passage: Binding sections (amounts, obligations, deadlines): to be re-read live, summarized or not.
Attachments: Most assistants accept documents and images: the basis of "summarize this file" uses.
Analysis leads: "Here are my columns: [list]. Which relevant analyses and pitfalls?": AI as research director.
Critical points: What gets verified first: figures, dates, names, quotes, legal references, surprising claims.
Assisted first draft: AI produces structure and substance; you adjust facts, tone and commitments — the standard professional-writing flow.
HR first draft: Job ad, job description, interview frame, onboarding plan: TYPE structures, without individual data.
Meeting prep: "Get me ready for my [time] meeting": context, attendees, documents, open items — daily use #1.
Prompt: Your instruction to the AI. All answer quality depends on it — the subject of level 1.
Pseudonymization: Replacing names with codes (Client A…): weaker than anonymization — structure alone can identify.
Leading question: Wording that contains the expected answer: it biases the model. Fix: phrase neutrally, or ask for both for AND against.
Transverse question: "Where does [project] stand?": the answer crosses sources that classic search keeps separate.
Funnel questions: From broad to precise: overview → players → figures → controversies — each answer feeds the next question.
Meeting recap (Teams): Summary, decisions, assigned actions — generated if transcription was enabled (and announced).
Connected search: The assistant performs real web searches and cites sources — the mandatory mode for facts and news.
Directed rephrasing: "Rewrite more diplomatic / firmer / simpler": changing tone without changing substance.
Privacy settings: Account options: disable use of your conversations for training, manage history, delete threads.
Data rules: The plain translation of classification: "client and HR data: only [vetted tool]; never in a public tool".
Anti-bias review: Targeted checking of people-related content: gendered wording, stereotypes, implicit exclusions.
Decisions/actions record: The meeting format: table of decisions made / actions / owners / deadlines — ask for it explicitly.
Thread summary (Outlook): The "Summary" banner of a long thread: who said what, where the decision stands — with clickable references.
Executive summary: 3-5 sentences for a busy decision-maker: conclusion first, essentials next.
Back-translation: Having the result retranslated into French (ideally in a fresh thread): meaning gaps jump out.
GDPR: The European personal-data regulation: fully applies when AI processes information about people (clients, employees, candidates).
Role: The hat given to the AI ("You are a French labor-law attorney"): it calibrates vocabulary, level and angle of the answer.
Shadow AI: Clandestine use of unvetted tools: the real risk of bans without alternatives — the answer is a vetted tool + a charter.
Structuring: From scattered notes to a plan: "Organize these items into a logical 3-part plan with headings".
Revealed oversharing: Unexpected content in an answer = a document shared too broadly upstream: flag it, don't exploit it.
Comparison table: Specify columns: "Compare X, Y, Z in a table: price, timeline, pros, limits, one-sentence verdict".
Task: The requested action, with a precise verb: write, compare, list, critique, rephrase, suggest 3 options.
Test on known cases: Any generated formula gets validated on 2-3 rows whose result you know — before adoption.
Test on a copy: Any script is tried on test data or a copy — never directly on production files.
Brand tone: Your documented editorial style (or demonstrated by examples): imperative to avoid generic content.
Translation with register: "Translate into professional British English, formal yet warm register": tone travels with meaning.
Data journey: Free public tool: possible use for training (per settings); enterprise offer: contractual framework, no training.
Style transfer: "Write in the style of this text: [sample]": tone, rhythm and vocabulary imitated — your texts keep your voice.
Usage transparency: Saying when content is substantially AI-generated, per context and charter — the AI Act mandates it for some content (5.4).
Triangulation: Cross-checking the AI synthesis against 2-3 primary sources consulted directly: the analysts' method.
Human validation: Who re-reads what before release: the explicit responsibility chain.
Variants: "Give 3 different versions": exploiting model variability to choose rather than endure.
Assisted monitoring: A reusable templated query run periodically: "What's new this month on [topic]? Dated, cited sources".
Proportionate verification: The golden rule: internal email = quick re-read; binding document = systematic verification of every fact.
Zero-shot: Asking without examples (the default mode): fine for simple tasks, risky for precise formats.
Risk zone: Requests where hallucination is likely and costly: exact figures, quotes, case law, bibliographic references, news.