The AI Glossary for Finance

Every one of these turns up in a vendor demo, and most explanations of them are written for engineers. This is the finance version: what the term means, and what it changes about a decision you have to make.

Definitions here match how these terms are used across the rest of this site rather than a generic dictionary. Where a word has a loose popular meaning and a precise technical one, the precise one is what a vendor is selling you.

The basics, and the two that cause the most confusion

Term	What it means	Why it matters in finance
Large language model, or LLM	The underlying system that predicts text. Claude, GPT and Gemini are models, or families of them.	You are almost never buying a model. You are buying a product built on one, and the product decides far more about your experience than the model does.
Token	The unit a model reads and writes in. Roughly three quarters of a word in English, so 1,000 tokens is about 750 words.	It is the billing unit on API pricing and the unit of every limit you will hit. When a vendor quotes cost per million tokens, that is what they mean.
Context window	How much the model can hold in mind at once, measured in tokens. It covers your files, the conversation so far, and the answer.	This is the real constraint on finance work. A large general ledger export will not fit, which is why tools chunk and retrieve rather than reading everything.
Inference	One run of the model producing an answer.	The thing you are paying for per use. Also why a slow answer is slow: it is generating, not searching.
Hallucination	The model producing something fluent and wrong, with the same confidence as something correct.	The whole reason for the review step. In finance the dangerous version is not a wrong fact but a plausible number, because a plausible number looks exactly like a real one.
Non-deterministic	The same input can produce different output on different runs.	The single most important property on this page. It is why a prompt should not own a recurring, audited process and a query should.

Getting your own data in

Term	What it means	Why it matters in finance
Grounding	Making the model answer from a source you supplied rather than from what it absorbed in training.	The difference between an answer about your business and an answer about businesses in general. Almost every serious finance use depends on it.
RAG, or retrieval augmented generation	The system searches your documents for relevant passages and puts those into the context before the model answers.	How a tool handles a knowledge base bigger than the context window. It also means the model sees the passages retrieved, not your whole set, so retrieval quality decides answer quality.
Embedding	A numeric representation of text that lets a system find passages by meaning rather than by keyword.	The mechanism under retrieval. Worth knowing only because it explains why a search can find the right paragraph without matching a single word of your question.
Knowledge base	The set of documents attached to a project, assistant or agent.	Noise in here is worse than a gap. A superseded document does not sit quietly; it gets retrieved and cited.
Fine-tuning	Further training on your own examples to change how a model behaves.	Rarely the answer for a finance team. Better instructions and better retrieval solve most of what people reach for fine-tuning to fix, at a fraction of the effort.
Structured output	Forcing the answer into a fixed shape, such as JSON or named fields.	What makes output usable by another system rather than by a reader. The difference between a paragraph about variances and a table you can load.

Agents, and what separates one from a macro

Term	What it means	Why it matters in finance
Agent	Something given a goal that works out its own path. The distinction that matters is whether the thing decides its own next step.	A flow that refreshes a report every Monday at 6am follows a path you drew. That is automation. An agent is more powerful and considerably easier to get wrong.
Agentic	Adjective form, applied to anything with some autonomy. Also the most overloaded word in the category.	When a vendor says agentic, ask what decision it makes without you. If the answer is none, it is a scheduled script with a new label.
Tool use, or function calling	The model calling out to something else, such as running code, querying a database or hitting an API.	The step that turns a chat into something that can act. It is also where permissions start to matter, because a tool can write as well as read.
MCP, or Model Context Protocol	An open standard for connecting a model to external tools and data sources, so any compliant client can use any compliant server.	It is the reason your tools are suddenly connecting to each other. For finance the practical read is fewer bespoke integrations and one more thing to govern.
Skill	A folder with a SKILL.md file in it, holding a name, a description of when to use it, and the instructions underneath.	A way to teach a model your process once and reuse it, rather than re-explaining it every session.
Human in the loop	A required approval before an agent’s action takes effect.	The control that makes an acting agent acceptable in a finance process. Put it in front of anything that touches the ledger.
Orchestration	Coordinating several steps, tools or models into one workflow.	What tools like n8n and Power Automate do. Usually the deterministic wrapper you want around the non-deterministic part.

Checking whether it works

Term	What it means	Why it matters in finance
Eval	A repeatable test of output quality against cases where you know the right answer.	The finance version is running it on a month you have already closed. It is the only honest way to know whether a setup is good, and it is the step people skip.
Benchmark	A public test used to compare models.	Almost never predictive of performance on your work. A model that tops a reasoning benchmark can still misread your chart of accounts.
Guardrail	A rule constraining what the system may do or say.	Where your policy becomes enforceable rather than aspirational.
Prompt injection	Instructions hidden in content the model reads, which hijack what it does next.	The risk that arrives the moment a model reads documents or email you did not write. It is why an agent with write access needs an approval step.
Audit trail	A record of what was asked, what was returned, and what was done with it.	The question an auditor will ask, and the one most AI workflows cannot answer. Build it deliberately, because it is not a feature you switch on.

The words that decide whether you are allowed to use it

Term	What it means	Why it matters in finance
Tenancy	Whether your data sits in your organization’s own environment or a shared one.	Usually the first question your IT team asks and often the one that settles the tool choice on its own.
Zero data retention	The provider does not store your inputs or outputs after the request.	Frequently the concession that gets a tool approved. Ask whether it is on by default or something you have to request.
Training opt-out	Your data is not used to improve the provider’s models.	Different from retention, and people conflate them. A provider can delete your data and still have trained on it, or keep it and not train on it.
Compliance API	An interface that lets your organization retrieve its own usage records programmatically.	What makes governance real rather than a policy document. Coverage differs by product even within one vendor, so check per surface.
Data residency	Which country your data is processed and stored in.	Decides the answer for regulated entities and for anything touching EU personal data, before any feature comparison starts.

Terms that mean less than they sound like

Term	What it means	Why it matters in finance
AI-powered	Present in almost every finance product listing.	Carries no information. Ask which specific task it does and how you would check it was right.
Autonomous	Marketing for agentic, usually.	Ask what runs without a human and what happens when it is wrong at 2am.
Real-time	Sometimes literal, often means refreshed on a schedule.	Ask the refresh interval. In finance the difference between live and hourly is often the whole point.
Copilot	Microsoft’s product family, and also a generic term for an assistant alongside your work.	Worth clarifying which one somebody means, because Microsoft ships several products with the word in the name and they have different entitlements.

Where to go from here

If a term here made you want the practical version, AI tools for finance covers the tools, AI agents for finance covers agents properly, and AI in finance is the pillar that links everything.

For the non-AI vocabulary, the finance terms glossary is the companion to this page and runs to over 200 entries.

This page is maintained. If a definition here has drifted out of date, it is worth telling me, because a wrong definition is more damaging than a missing one.