Glossary, plain-English AI terms

No jargon required. Here's every term the library uses, in one line each. Skim it once and the guides will read easily.

The AI tools

Claude / ChatGPT / Gemini / Copilot, the main AI chat assistants. Claude is the Lab's default.
Model, the underlying AI (e.g., "Claude Opus"). Newer/bigger models are more capable.
Consumer (free/Pro) account, a personal AI login. Do not put client info in these.
Team / Enterprise tier, a firm-level business plan whose terms keep your data private and out of model training. This is the kind of account to use for real client work.

Project (Claude) / Custom GPT (ChatGPT), a saved workspace that remembers your instructions and reference files, so you don't re-explain yourself each time.
Custom Instructions / system prompt, standing instructions you set once that apply to every chat.
Training, when a vendor uses your inputs to improve its model. You want this turned off for client data (business tiers exclude it by default).
Zero data retention (ZDR), the vendor doesn't keep your inputs after processing. Stronger privacy.
DPA (Data Processing Agreement), the contract that binds a vendor to protect your data. Part of using a tool with real client information.

Prompt, what you type to the AI; your instruction or question.
Hallucination, when AI confidently makes something up (a fake citation, a wrong number). The reason we verify before we rely.
Token (in this library), a stand-in label like [Person_1] that replaces a real name or number when you redact client data. (In AI billing, "tokens" also means chunks of text, different meaning; we always mean the redaction placeholder.)
Anonymize / pseudonymize, remove identifying details (or swap them for tokens) so the AI never sees who the client is.
Local key, the private list matching each token back to the real value. Stays on your machine.
RAG / retrieval, the AI looks up real source documents before answering, instead of relying on memory. More reliable for facts.

Skill, a saved set of instructions that makes the AI run a task the same way every time.
Claude Code, a version of Claude that runs in a terminal and can build files, scripts, and skills.
API, a way for software (not a person typing) to talk to the AI; powers custom automations.
Agent / agentic, AI that can take a sequence of actions on its own toward a goal (not just answer one question). Powerful, keep a human approving anything important.
MCP (Model Context Protocol), a standard way to let the AI read your other tools (email, calendar, drive, practice-management software).
No-code / low-code, automation tools you set up by clicking, not coding. Examples: Zapier, Make, n8n (n8n is just the name of one of them).

PII, personally identifiable information (names, SSNs, EINs, account numbers, addresses).
TRI (tax return information), anything a client gives you to prepare a return, plus what you derive from it. Protected by §7216.
IRC §7216 / §6713, the federal rules restricting how you disclose or use a client's tax return information; can require written consent before it goes to a third party.
FTC Safeguards Rule (under GLBA), the law requiring your firm to protect client data; satisfied via a written security plan.
WISP (Written Information Security Plan), your firm's required data-security plan; any AI tool you adopt should fit inside it.
SSTS, the AICPA's enforceable tax-practice standards; they say you stay responsible for the work product even when you use AI.
Circular 230, the rules governing practice before the IRS (competence and due diligence).
AICPA Code of Professional Conduct, the profession's ethics rules, including client confidentiality.
SSARS / AR-C and GAAS / AU-C, the standards for compilations & reviews (SSARS) and audits (GAAS), if you do attest work.
Reviewer of record, the human professional who reviews, decides, and signs. AI is never the reviewer of record; you are.

Part of the AI Lab for Accountants library. Don't see a term? It probably isn't needed to get started, focus on the guide in front of you.