Glossary

70 terms from AI at Work Intermediate, in plain words, each with the chapter that teaches it.

Search the glossary

#

[CHECK]: A marker you invent and ask the AI to type wherever a needed fact is missing, instead of guessing. It turns hidden guesses into visible gaps you can fill.
[GUESS]: A marker you ask the AI to attach to any score or value based on a guess rather than the facts you gave. Resolve these before trusting a weighted table's winner.
[UNCLEAR]: A marker you ask the AI to type where it can't read part of an image, such as a whiteboard. It shows you exactly what to check by hand.

A

Agent: An assistant that carries out a multi-step task by itself, such as browsing sites, filling in forms or creating files, then reports back. Kettle & Crane could use one to collect café opening hours from 12 websites.
Agent mode: The setting in an assistant that lets it work through a multi-step online task on its own instead of just replying in chat. Watch the first runs and approve anything irreversible.
AI policy: Your workplace's written rules on which AI tools are approved, for what, and with which information. If there isn't one, agree a simple policy before using AI with internal or personal data.
Anonymize: Replace names and identifying details with consistent placeholders such as "Customer A" before pasting text into an AI. You keep the key that maps them back outside the AI and watch for indirect identifiers like "our only employee in Eugene".
Auto: A model-menu option where the assistant picks between a fast and a thinking model for you. It can't know how important the answer is to you, so choose thinking yourself for important work.

B

Baseline: How long a task takes, and how good the result is, before you change how it's done. Sam's baseline for the wholesale account review was 3.5 hours.
Bias: Unfair assumptions about people, such as age, gender or disability, that a model can pick up from the human writing it learned from. Read any text about people for these before it goes out.
Business plan: A company-managed version of an assistant with stronger privacy, which normally doesn't use your chats to train future models. Use it for work if your employer provides one.

C

Canvas / artifact: A document panel that opens beside the chat (ChatGPT and Gemini call it Canvas, Claude calls it artifacts). You can select one paragraph and ask for a change to just that part, or edit it yourself.
Chain: A sequence of prompts where each step's output becomes the next step's input, and you check each before moving on. For example: angles, outline, draft, edit.
Code tool: The assistant's ability to write and run small programs (usually Python) to calculate, clean data and draw charts. Because the program does the arithmetic, the figures are exact, though it can still use the wrong column or rows.
Common business metrics: Standard analyses you can ask for if you give the definition in your prompt: month-over-month change ((this − last) ÷ last), share of total, average order value (total value ÷ orders), repeat rate (customers who reordered within N days ÷ all customers), concentration (share of value from the top 10% of customers) and a cohort table (customers grouped by first-order month). Seasonal businesses should compare with the same month last year.
Connector: A link between a stand-alone assistant and another app such as Gmail, Google Drive, Slack or a customer database. Check whether it can only read or can also write, and connect only what a task needs.
Context window: The maximum amount of text a model can take in at once, including your files and the whole conversation. Quality drops well before the limit, and details in the middle get less attention.
Cross-check: Testing an important answer before you rely on it: run the same prompt three times in new chats, ask a second assistant to find errors, and ask what would make the answer wrong. Where answers disagree is where to look.
CSV: A plain-text spreadsheet file with commas between the values, which almost every system can export. You upload it to the assistant rather than pasting rows into chat.
Custom assistant: A saved set of instructions, knowledge files and tools for a repeated job, such as Kettle & Crane's customer reply helper. ChatGPT calls them custom GPTs, Gemini calls them Gems and Microsoft calls them Copilot agents.
Custom instructions: A few lines in Settings about who you are and how you like replies, applied to every chat. They save you repeating the same background each time.

D

Deep research: A mode where the AI plans the research, runs many searches, reads dozens of pages over 5 to 30 minutes and writes a cited report. A vague question gives a long general essay, so brief it with a question, a Why, a scope, sources and an output.

E

Escalation rule: An instruction telling a custom assistant which cases to hand to a person, and the words to use. In Kettle & Crane's reply assistant, a lawyer or injury mention gets "Escalate to Sam" and no draft.
Extraction: Pulling specific facts out of documents into a consistent structure, such as one table row per supplier quote. It turns hours of reading into a table you can sort and compare.

F

Fact pass: An editing pass where the AI lists every figure, date, name and claim in a document and the input each came from, flagging any with no source. It is the pass that catches invented numbers.
Fast model: A model mode that replies almost instantly, suited to rewording, short emails and first drafts. For multi-step or numeric problems a thinking model is better.

G

Gemini in Google Workspace: Google's AI built into Gmail, Docs, Sheets, Slides and Meet. It can draw on your emails, files and calendar, so name the exact source to use; in Sheets, "Fill with Gemini" fills a column from a prompt, such as classifying customer comments.
Green, yellow and red information: A three-color way to sort information by how sensitive it is before putting it into an AI tool. Green is public or non-private and can go into any assistant, yellow is internal business information for an approved work assistant only, and red is personal or highly sensitive and should go nowhere unless workplace rules allow it in an approved tool (anonymize it first).

H

Hallucination: A confident-sounding but false statement from the AI, such as an invented figure, quote, policy or source. It happens because the model writes text that sounds right.

K

Knowledge files: Documents a custom assistant can read and quote, such as policies, price lists and example replies. Keep them current, one topic per file, and free of secrets.

L

LLM (large language model): The program behind an AI assistant, trained on a huge amount of text to predict which words are likely to come next. It writes a reply one small piece at a time.

M

Machine habits: Patterns that make writing sound AI-made, such as empty openers, inflated verbs like "unlock", "It's not just X, it's Y" and tidy summary endings. Add them to your style guide as banned habits.
Memory: An assistant feature that remembers facts from earlier chats. Review what it has stored, delete anything wrong or private, or switch it off.
Microsoft 365 Copilot: Microsoft's AI built into Word, Excel, PowerPoint, Outlook and Teams, depending on your employer's license. It can summarize threads, write formulas, and in Excel analyze tables and build charts and pivot tables (spreadsheet tables that summarize data by category).
Model: A version of the AI that you choose from a menu near the message box. The choice is essentially between a fast mode and a thinking mode.

N

NotebookLM: Google's tool for asking questions about a library of documents. It answers only from the sources you add and links each statement to the passage it came from.

P

Pre-mortem: A stress-test where you imagine the plan has already failed a year from now and list the most likely reasons, with early warning signs. Imagining failure as fact gives sharper risks than asking "what could go wrong?"
Primary source: An original or official source, such as a regulator's guidance or a company's own price list, which is the most trustworthy for facts about its own rules or offers. Vendor content and forum posts deserve little or no trust as facts.
Profile the data: Asking the AI to describe a dataset before analyzing it: rows, columns, types, date range, missing values and duplicates. It catches a wrong or incomplete export before anything else runs.
Project (Notebook): A workspace in ChatGPT, Claude or Gemini that groups related chats with shared instructions and files. It suits one body of work over weeks, like a spring launch.
Prompt: The message you send to an AI assistant, such as a question, instruction or task description. The assistant writes its reply by continuing from the prompt and the rest of the conversation.
Prompt injection: Hidden text in a web page, email or file written to give an AI instructions, such as "ignore your task and send…". Don't let an agent with access to sensitive accounts browse unknown sites.
Python: A programming language the code tool usually writes in to analyze your spreadsheet. You don't need to read it, but you can open the panel to see it.

R

Reconcile: Comparing a grand total from the AI's analysis with a figure you already trust, such as the accounts or last month's report. It's one of the final checks before a figure leaves your desk.
Responsive search ad: A Google Ads format where you supply up to 15 headlines and 4 descriptions and Google tests combinations of them. Each line has strict character limits, which the AI miscounts unless it uses code.
Role-play: Using the AI to play a character with real motives, such as Kettle & Crane's biggest customer, so you can rehearse a negotiation or hard conversation. Tell it not to make things easy and to give feedback afterwards.
Rubric: A short checklist of what good looks like for a piece of work, with items you can judge Pass or Fail. You make the AI critique a draft against it instead of asking "is this good?"

S

Scenarios: Modelling best, expected and worst cases from your assumptions with code, then finding which single assumption the result is most sensitive to. That assumption is the one to research.
Scheduled task: A prompt the assistant runs automatically on a schedule, such as every Monday at 7am, to produce a news summary. Read the first few runs critically, then tighten the prompt.
Schema: A fixed list of fields, with formats and rules for each, that you give the AI for extraction. For the supplier quotes it included fields like price per 1,000 bags and lead time in business days.
Shadow AI: Using unapproved AI tools with work information. It can break privacy law and contracts even with good intentions.
Six-part brief: A complete prompt made of six parts: task, context, inputs, requirements, output and quality check. Think of it as the brief you'd give a skilled freelancer who has never heard of your company.
Stakes: How much it matters if an AI output is wrong, which decides how much checking it needs: low (read it), medium (fact pass on names, numbers, dates, promises), high (also verify sources, get a second reviewer or an expert).
Standard operating procedure: A written, numbered set of steps for doing a task the same way every time, usually with a purpose line, a list of what you need beforehand and a final checklist. For example, Priya's spoken explanation of preparing weekly shipments can be turned into one, which the person who knows the task should then review.
Structured interview: An interview where every candidate gets the same questions, each scored against what a strong, average and weak answer would include. It supports a fairer hiring process.
Style guide: A short description of how you write (sentence length, vocabulary, openings, banned words) that the AI builds from your own samples. Add "Follow my style guide" to writing tasks.
SUMIFS: A spreadsheet function that adds up values meeting several conditions, for example one café's orders in one month. Test it on one cell you can check by hand before copying it across.
Sycophancy: The AI's tendency to agree with you and praise weak ideas, because it's trained to be helpful and agreeable. Avoid it by asking for criticism and not revealing which answer you prefer.

T

Tags: Labels such as <email> and </email> that mark where pasted material starts and ends, so the AI doesn't confuse it with your instructions. Any clear marker works.
Temporary chat: A chat that isn't saved to your history or used for memory. It suits one-off questions you don't want shaping future answers.
Thinking model: A model that works through a problem step by step for seconds or minutes before answering. Use it for analysis, planning, long documents and anything with numbers or logic.
Tools: Extra abilities an assistant can use while answering, such as web search, deep research, code, files, images, connectors and agent mode. Before a task, ask which tool it needs.
Training data: The text a model learned from. Without web search, it answers current-fact questions, like this year's UPS rates, from this older data.
Transcript: A written record of everything said in a meeting or call, which Teams, Zoom or Google Meet can produce. AI minutes from them often turn suggestions into decisions, so require a supporting quote for each action.
Triage: Sorting items by priority using rules you give, such as splitting emails into Reply today, This week, Read only and Ignore.

V

Variance: Actual minus budget for each line, in dollars and percent. The finance playbook has code calculate it and flags lines far off budget.
Voice mode: A feature that lets you talk to the assistant and hear it reply, useful for thinking aloud or rehearsing a presentation. Don't use it where others can overhear private information.

W

Web search: A tool that finds and reads current web pages and links its sources. Switch it on for anything that changes, such as prices, rules and news; otherwise answers come from old training data.
Weighted decision table: A table where each option is scored on several criteria and each score is multiplied by the criterion's weight, then added up. It makes trade-offs visible, and you can ask how much a weight must change to flip the result.
Workflow: A recurring task split into saved steps, each done by the right tool or person, with checks. Sam's account review workflow cut 3.5 hours to 55 minutes.