Now taking on new projects · Free call, AI Roadmap, Build, SupportSee services & pricing
Resources · Glossary

AI terms, in plain English

35 terms a business owner is likely to meet, each with what it means and, where it helps, why it matters for your business.

AI agent

A system built around a language model that can take steps toward a goal rather than only answer a single question. It might read an email, look up a record, draft a reply and create a task, deciding the order as it goes. Because agents act, the permissions they are given and the approval steps around them matter as much as how well they reason.

API

An application programming interface: a defined way for one piece of software to request data or actions from another. Most AI models are used through an API, and AI workflows connect to business systems through their APIs. Whether a system has a usable API often decides how easily it can be included in an automation.

Audit log

A record of who or what did what, and when, in a system. For AI workflows, an audit log shows which inputs the system received, what it produced and which actions it took. It is how you investigate a mistake, answer a customer question about what happened, and show that controls are working.

Context window

The amount of text a language model can consider at once, including instructions, documents provided and the conversation so far. It is measured in tokens. Anything outside the context window is not visible to the model, which is one reason systems retrieve only the relevant parts of large document collections.

Data residency

Where data is physically stored and processed, usually described by country or region. Some contracts, industries and regulations require data to stay in particular locations. If this applies to you, check where each service in an AI workflow processes data, and have a qualified adviser confirm what your obligations are.

Data retention and zero retention

A provider's terms for how long it keeps the data you send it, and for what purposes, such as abuse monitoring or model training. "Zero retention" or similar terms mean the provider commits not to store inputs and outputs beyond processing the request, usually under specific plans or agreements. These terms often differ between free, business and enterprise plans, so read the terms for the plan you actually use.

Embeddings

Lists of numbers that represent the meaning of a piece of text, so that passages with similar meaning have similar numbers. They let a system find relevant documents by meaning rather than exact keywords. Embeddings are a building block of most retrieval systems used with language models.

Evaluation (evals)

Structured testing of an AI system against examples with known correct results, producing measurable outcomes rather than impressions. Evaluations are run before launch and again after any change to the model, prompts, tools or data. They are the main evidence that a system is good enough to use.

Fine-tuning

Further training of an existing model on a set of examples so it adapts to a particular style or task. It can help with consistent formats or specialized language, but it requires good training data and must be repeated or re-checked when base models change. For many business tasks, clear instructions plus retrieval of relevant documents is simpler and easier to update.

Grounding and citations

Grounding means requiring a model to base its answer on specific source material, such as your documents, rather than on general knowledge. Citations show which source passages support each part of the answer. They make answers checkable, though a citation can still be misapplied, so reviewers should confirm the source says what is claimed.

Guardrails

Controls that limit what an AI system can receive, produce or do: input filters, output checks, blocked topics, spending limits, required approvals and restricted permissions. The most reliable guardrails are enforced by the surrounding software and permissions, not only by instructions to the model, because instructions can be ignored or overridden.

Hallucination

When a model produces information that is false or unsupported but presented confidently, such as an invented policy, figure or citation. It is a known behavior of language models, not a rare fault. Grounding, citations, testing on real examples and human review all reduce the risk, but none removes it entirely.

Human-in-the-loop

A design in which a person reviews, approves or corrects an AI system's output before it takes effect. It is most important for actions that are hard to undo or that reach customers. A review step only works if the reviewer has the evidence and time to check properly.

Knowledge base

A collection of documents an AI system draws on to answer questions, such as policies, procedures, product information and FAQs. Its quality sets a ceiling on the quality of answers: outdated or contradictory documents produce outdated or contradictory answers. Each document should have an owner who keeps it current.

Large language model (LLM)

A type of AI model trained on very large amounts of text to predict and generate language. LLMs can draft, summarize, classify, extract information and follow instructions. They do not look facts up unless connected to sources, and their output can be wrong, so how they are used and checked matters.

Latency

The time between sending a request and getting a response. For a customer chatting live, delays are noticeable; for a nightly batch of documents, they rarely matter. Larger models and multi-step agents tend to be slower, which is one of the trade-offs when choosing a design.

Least privilege

The security principle of giving any user or system only the access it needs for its job. For AI workflows, this means a dedicated account with narrow permissions, such as read-only where reading is enough. It limits the damage if the system makes a mistake or is manipulated by a malicious input.

Model Context Protocol (MCP)

An open standard for connecting AI applications to tools and data sources in a consistent way. A system that supports MCP can expose its functions, such as searching records or creating tickets, to compatible AI applications without a custom integration for each one. As with any connection, the permissions granted through it should be kept narrow.

Model provider

A company that offers access to AI models, usually through an API and often through its own apps. The provider's terms govern data retention, data location, usage limits and pricing. Choosing a provider is partly a technical decision and partly a contract and data-handling decision.

Multimodal model

A model that can work with more than one kind of input or output, such as text, images, audio or scanned documents. This allows tasks like reading a photographed form or describing a diagram. Accuracy on images and poor scans varies and should be tested on your own examples.

OCR and document extraction

Optical character recognition turns images of text, such as scans and photos, into machine-readable text. Document extraction goes further, pulling specific fields like invoice numbers, dates and totals into structured form. Both can make errors on poor scans, handwriting and complex tables, so extracted values that matter should be checked.

Orchestration

The logic that coordinates the steps of an AI workflow: what runs in which order, which tools or models are called, how errors are handled and when a person is asked to approve. Good orchestration makes a system predictable and easier to troubleshoot. It is often where most of the work in a custom build goes.

Personal data

Information that identifies or relates to an individual, such as names, contact details, account numbers or health information; sometimes called personally identifiable information (PII). Many laws and contracts restrict how it is collected, shared and kept. Before sending personal data to any AI service, a qualified adviser should confirm what is permitted in your situation.

Prompt

The input given to a language model: instructions, context, examples and the question or task. In a business system, most of the prompt is written once and reused, with each new case inserted. Small changes to a prompt can change behavior, so prompt changes should be tested like any other change.

Prompt injection

An attack in which text inside an input, such as an email, document or web page, contains instructions intended to make an AI system do something its owner did not intend. Models can follow such instructions, and no technique currently prevents this completely. The main defenses are limited permissions, approval for consequential actions, and testing with adversarial inputs.

Retrieval-augmented generation (RAG)

A technique in which a system first searches a set of documents for passages relevant to a question, then gives those passages to the model to base its answer on. It lets a model answer from your current documents without retraining. Answer quality depends heavily on whether the right passages are found and whether the documents are accurate.

Service account

A login created for a system or automation rather than a person. Giving an AI workflow its own service account makes its actions traceable in logs, lets you set its permissions precisely, and lets you revoke its access without affecting any staff member.

Structured output

Model output in a fixed, machine-readable format, such as a set of named fields, instead of free text. It allows results to be checked automatically and passed reliably into other systems. Correct format does not guarantee correct content, so the values still need validation.

System prompt

Standing instructions given to a model that apply to every interaction in a given application, such as its role, rules, tone and limits. It shapes behavior but is not a security boundary: users or malicious inputs may be able to get around it. Anything that must never happen should also be blocked by permissions or software checks.

Test set

A collection of real example inputs paired with the correct outputs or behavior, used to measure an AI system's accuracy. It should include typical cases, difficult cases and cases the system should refuse or escalate. It should be kept separate from examples used while building, and grow as new failures are found.

Token

The unit in which language models read and produce text; a token is typically a word or part of a word. Model usage is usually priced and limited by tokens, and context windows are measured in them. Long documents and long conversations use more tokens, which affects both cost and speed.

Tool use and function calling

The ability of a model to request that the surrounding software perform a defined action, such as searching a database, checking a calendar or creating a record, and then use the result. This is what lets an assistant become an agent. Each tool should be limited to what the workflow needs, and consequential tools should require approval.

Vector database

A database designed to store embeddings and quickly find the entries most similar in meaning to a query. It is commonly used to power retrieval in RAG systems. For smaller document collections, simpler storage can sometimes do the same job.

Webhook

A way for one system to notify another automatically when something happens, such as a new order, form submission or payment. Webhooks let an AI workflow start as soon as an event occurs instead of checking on a schedule. Whether your systems offer webhooks affects how responsive an automation can be.

Workflow automation

Using software to carry out a sequence of business steps automatically, such as moving data between systems, sending notifications or creating tasks. Traditional automation follows fixed rules; AI can be added for steps that involve reading or writing language. Many useful projects combine both, using fixed rules wherever they are enough.

§02Next step

Find your first AI win

Book a free 20-minute call. You'll leave knowing exactly where AI pays off in your business, and what to do first.