Glossary

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

Search the glossary

A

Accuracy: The share of test cases an AI step gets right: correct answers divided by all cases. Kettle & Crane's first routing step scored 85% on its 40-message test set.
Agent: An AI assistant that carries out a multi-step task itself. It looks, plans, acts and checks in a loop, using tools like a browser, files and apps. Sam's agent built the Seattle café lead list this way.
Agent brief: The written instructions you give an agent, in six parts: goal, inputs and tools, method, boundaries (what it must never do, such as buy, send, delete or sign up), stop and ask (when to pause and check with you), and done means. It must cover what happens when things go off-plan, because the agent may take 50 steps before you look again.
Agent mode: The setting in an assistant, such as ChatGPT's agent or Gemini's agent mode, that lets it act as an agent. It often runs in a browser you can watch and pauses before important steps.
AI step: A step in a workflow where text is sent to an AI model with your instructions and the answer is passed to later steps. Typical jobs are extract, classify, summarize and draft. In order intake it turns a café's email wording into fields.
Approval step: A point where a workflow pauses until a person says yes before it acts. At Kettle & Crane, Sam approves each order row marked "To approve" before it goes to shipping.
Audit trail: The record of who approved what, when, and whether they changed anything. It lets you review decisions later and also gives you your error rate.
Automation ladder: A five-level scale of automation: saved prompt, custom assistant, scheduled task, workflow and agent. Each step up saves more time but carries more risk, so the sensible choice is the lowest level that does the job.
Automation platform: A no-code service such as Zapier or Make that connects apps and runs workflows between them, built in a visual editor. Others include n8n, Microsoft Power Automate and Microsoft Copilot Studio (which also builds agents).
Automation policy: A one-page set of rules for AI automation: approved tools, registering first, least access, people deciding key matters, data limits, testing first, owners and incidents. Kettle & Crane's has eight rules.
Automation register: A shared list of every automation and team assistant. It records the owner, backup, connections, data handled, people in the loop, log location and last review. In a crisis it tells you what's running and who to call.

C

CAN-SPAM Act: The US law on commercial email. It doesn't require prior consent, but each message needs honest sender details and subject, a statement that it's an ad, a postal address and a working opt-out honored within 10 business days.
CCPA (California Consumer Privacy Act): A California privacy law that governs how businesses collect, use and keep residents' personal data, and applies once a business passes its size thresholds. It is one example of the state privacy laws that apply to personal data handled in automations.
Champion: A respected, curious colleague chosen on each team to help others use AI well. They collect good prompts and problems and meet the other champions.
Citation check: A test for knowledge assistants. For each test answer, you check that the passage the assistant cites really says what it claims. Wrong or missing citations are an early warning of document problems.
Classify: An AI-step job that chooses one label from a fixed list, such as order, complaint, question or other. Kettle & Crane uses it to route contact-form messages.
Confidence: A field where the AI marks itself high (sure) or low (it had to interpret or guess). A rule sends low-confidence cases to a person, which is how Bloom Café's "swap" order reached Sam.
Connected search: A way to give AI company knowledge: the assistant searches your drive, SharePoint, wiki or email through connectors, with each user's own permissions. It suits large, changing collections with clean sharing settings.
Connector: A link that lets an assistant look things up, and sometimes act, in another app such as Gmail, Google Drive or Slack. Each connector gets permissions you approve when you connect it.
Cost per run: A month's running cost divided by that month's runs. For order intake, $34 ÷ 338 runs is about 10 cents per order, which you compare with the time saved per run.
Credits / tasks: The units automation platforms charge by. Zapier counts tasks (completed actions); Make counts credits (formerly operations), roughly one per step run, with AI steps using more.
CRM (customer relationship management): A system that holds customer records. Once connected, an assistant can answer questions such as which cafés haven't ordered in 30 days.
Custom assistant: An assistant with saved instructions and files that you or a team reuse. It's level 2 of the automation ladder, and business plans let you share it with a team.
Customer-facing AI: A chatbot or automatic reply that talks to customers with no colleague seeing each answer first. Tell people it's an AI, limit it to topics your files cover, offer a route to a person and review a sample of conversations weekly.

D

Data processing agreement: A contract with a vendor that sets out how it may handle personal data on your behalf. It is part of the paperwork to check before a tool or automation handles personal data.
Dedicated mailbox: An email inbox used for only one purpose, such as orders@kettleandcrane.com. The trigger then never sees unrelated email.
Drafting steps: AI steps that write rather than label, such as a reply or confirmation email. Make them reliable with a fixed template and gaps, facts only from fields, length and tone limits, and always a draft until the edit rate shows people rarely change it.

E

Edit rate: The share of AI drafts a person changed before using them, and how much they changed. At Kettle & Crane, 41% of review reply drafts were edited, showing they were too formal.
Error rate: The share of outputs that were wrong, tracked from approvals and samples marked right or wrong in the log. Order intake had 0.6% against about 2% for hand typing.
EU AI Act: European Union law that adds duties for some AI uses, notably hiring and decisions about people. It also requires telling users they're talking to an AI if you serve people in Europe.
Exception handling: A design where clear cases run automatically and unclear ones go to a person. Exceptions include low confidence, a failed rule, a first-time customer or an unusual amount.
Extract: An AI-step job that pulls named fields out of messy text, such as order lines, dates, names or amounts.

F

Failure alert: A message, usually email, that an automation platform sends when a run fails. It stops a workflow from silently dying.
Fallback path: A branch that catches unexpected results, such as an AI label that isn't an allowed value, and sends the item to a person instead of dropping it.
Fixed fields: AI-step instructions that ask for named fields with allowed values, one per line, and nothing else. This gives later steps predictable input.
Freshness check: A test question about something that changed recently. If the assistant gives the old answer, an outdated document is still being found.

G

GDPR (General Data Protection Regulation): The EU privacy law that applies if you serve customers in Europe. The habits are to have a reason to use personal data, use no more than needed, keep it secure and not keep it longer than necessary.
Green coffee: Unroasted coffee beans, which roasters like Kettle & Crane buy from importers. Arabica is the main species the company buys, and Jo's Monday digest tracks green coffee and arabica prices.

H

Hallucination: When an AI states something invented as if it were fact, such as an invented refund in a draft. Training should show how to catch it.
Hand-off: The moment work passes from an automation to a person. Give them the original message, what the AI concluded and why, what it would have done next, and one-click actions such as approve, edit, reject or reassign.
Human in the loop: A design where a person takes part in an automated process at chosen points. The four patterns are approve, review, exception and sample.

I

Incident plan: A short plan written before anything goes wrong, with five steps: stop, assess, contain (recall or correct what went out, revoke access if needed), tell and learn. It includes knowing where each automation's off switch is.

K

Knowledge files: Documents attached to a custom assistant or project, which it reads and quotes from. Anyone who can use the assistant can in practice read them.

L

Least privilege: Giving a tool or person no more access than the job needs. For example, a connector starts read-only and gets write access only when a task requires it.
Log: A record of each run of an automation: the input, the extracted fields, the rule results and the outcome. It lets you find out what went wrong and measure error rates.

M

Make: A no-code automation platform whose workflows are called scenarios. It charges by credits and gives more control over complex branching workflows.
Mapping: Telling each workflow step which earlier field goes where. For example, you map the sheet's Summary column to the AI step's summary. An empty value usually means a wrong mapping.
MCP (Model Context Protocol): An open standard for connecting AI assistants to other software, like a standard plug. An app that offers an MCP server can connect to any assistant that supports MCP without custom code.
MCP server: An app's side of an MCP connection, which offers tools to assistants. Zapier's MCP server, for example, gives an assistant access to thousands of apps through one connection.

N

Nothing-to-report rule: A line in a scheduled prompt telling the AI what to say on quiet runs, for example "Quiet week". It stops filler and invented significance.

O

Off switch: The way to turn an automation off fast. Every platform has one, and each owner should know where theirs is and write it into the incident plan.
Oversharing: Files shared more widely than intended, such as with everyone in the company or anyone with the link. AI search can surface them. At Kettle & Crane, a staff pay spreadsheet shared with everyone was found this way.

P

Parallel run: Running a new automation on real data alongside the manual process, writing to a test sheet, to compare results without risk. Kettle & Crane's week matched the typed sheet on 80 of 81 orders.
Path / router: A condition that checks something and splits a workflow into branches. It's called a path in Zapier and a router in Make. Contact-form routing uses one to send each message to a different place.
Pilot: A small, time-limited trial with a few keen users from different roles before a wider launch. Kettle & Crane piloted with four people for a month.
Playbook: A role's guide to which tasks to do with AI, with what prompt or assistant, and how to check the result.
Prompt injection: Hidden instructions planted in content an AI reads, such as white text on a web page, meant to hijack it. It's the main security risk of agents and of workflows that read outside content.

R

Read and write access: Read access lets an assistant search and read emails, files, calendars and records without changing anything. Write access lets it send, create, edit, post, update or delete; start every connection read-only and add write only when needed, tool by tool, preferably with approval each time.
Retrieval: The technique of searching documents for the passages most relevant to a question and giving only those to the AI to answer from. Messy documents therefore give messy answers.
Rubric: A checklist of Pass or Fail criteria for judging drafts, such as "under 80 words" or "no promise of refunds". It's used because drafts have no single right answer.
Rules: Fixed if-this-then-that checks that don't need AI, such as "quantity must be 1 to 50". They are free, fast and never vary.
Run history: The platform's record of each run of a workflow, step by step, showing each step's input and output. It's how you see where a run went wrong.

S

Sample: A human-in-the-loop pattern in which a person regularly checks a random selection of automatically handled items. A typical rate is 5 to 10% on an ongoing basis, because a reliable step can drift as models, inputs and apps change.
Sandbox: A safe place for an agent or workflow to work, such as a new tab, a copy of a sheet or a test folder, so the real thing isn't touched.
Scheduled task: A prompt an assistant runs automatically at set times, such as Jo's Monday market digest. It's usually the safest first automation because it mostly reads and reports.
Service account: A shared, business-owned account that automations run on instead of a person's login. This way the automation doesn't stop when that person leaves.
Seven parts of an automation: Every automation has a trigger, input, AI step, rules, action, check and log. Writing all seven down before building prevents most problems.
Source notebook: A tool such as NotebookLM that answers only from the sources you add, linking statements to their passages. It's good for a defined set of documents.
Status column: A column in a sheet, such as "To approve", "Approved" or "Rejected", that a second workflow can watch to continue. It's a simple, visible, auditable way to build an approval.
Structured output: An AI setting where you define the fields in a form and the AI must fill exactly those. It removes most format errors.

T

Task scoring: Scoring a task from 1 to 5 on frequency, time, pattern, checkability (can you quickly tell if the result is right?) and safety. A total of 20 or more is a strong candidate; 12 or less, leave it manual.
Team assistant: A custom assistant used by many people who didn't write it. It needs an owner, tests and stated limits.
Test set: Real past examples with correct answers written down before running anything, used to measure an AI step or assistant. You rerun it after every change.
Threshold: The rule in a scheduled prompt for what's worth reporting, such as only price changes of 5% or more.
Time saved: Time before minus the time spent checking, fixing errors and maintaining the automation. Many automations that "save hours" save much less once these are counted.
Tool check: Questions to ask before a tool or connector handles company data: who makes it, how it uses your data, what access it needs, admin controls, how to export and delete, and any paperwork for personal data.
Traffic-light rule (green, yellow, red data): A way of sorting data by sensitivity: green is public, yellow is internal business information, and red is personal or sensitive. Red data belongs in an automation only if policy specifically allows it, and then only the minimum needed.
Trigger: The event that starts a workflow, such as a new email, a new row, a form submission or a time. It's the first of the seven parts of any automation.

U

Untrusted input: The principle of treating AI output as untrusted by the next step. You validate it with rules, and never let an AI step choose where to send data.

W

Workflow: An automated chain of steps across apps (a trigger, steps and actions), with AI as one step. It's called a Zap in Zapier and a scenario in Make.

Z

Zap / scenario: Zapier calls its workflows Zaps, and Make calls its workflows scenarios.
Zapier: A no-code automation platform whose workflows are called Zaps. It charges by tasks, usually has the most app connections, and also offers an MCP server.