AI glossary for business owners

18 terms

AI agent

A system that takes a sequence of actions toward a goal with some autonomy, such as reading an inbox and filing items. Needs tight limits, logging and stop conditions.

AI assistant

A chat-style tool built on a large language model that drafts, summarises and answers questions from what you give it. It predicts plausible text; it does not know your business unless you provide context.

Automation

Software that carries out steps automatically when something happens (a form is submitted, an email arrives). May or may not involve AI.

Baseline

A measurement of how a process performs before you change it — hours, error rate, turnaround. Without one you cannot show improvement.

Data terms

A provider's rules about how they store, use and retain what you enter. Differ between plans; check the current terms for your account.

Grounding

Giving the assistant your own source material — procedures, past documents, data — and asking it to answer only from that, so outputs can be checked against a source.

Hallucination

When an AI tool produces information that sounds right but is false or unsupported. The main reason outputs need review, especially numbers, quotes and facts.

Human in the loop

A defined step where a named person reviews or approves AI output before it is used. Only effective when the reviewer knows what to check.

Judgement work

Steps that depend on experience, context or accountability. AI can help prepare; people decide.

Knowledge base

An organised, searchable collection of your procedures and answers, which people — and grounded assistants — can rely on.

KPI

Key performance indicator: a small number of measures you track regularly to see whether the business, or a change, is working.

Large language model (LLM)

The type of model behind most AI assistants, trained on large amounts of text to predict likely next words. Strong at language tasks; can state wrong things confidently.

Pilot

A time-boxed test of a change on one workflow, measured against a baseline, ending in a keep, adjust or stop decision.

Prompt

The instructions and context you give an AI tool. Clear role, task, inputs, format and constraints usually produce better results than a short request.

Redaction

Removing names, contact details, account numbers and other identifiers from material before sharing it with a tool or another person.

Rule-based work

Steps someone could complete from written instructions alone. The best candidates for AI assistance or automation.

SOP (standard operating procedure)

A written, step-by-step description of how a recurring task is done, including decisions and quality checks.

Stop condition

A rule written before a pilot that ends or pauses it, for example an error rate above an agreed limit.