The words vendors use, in the words you use.

The trained system that turns text in into text out: the thing behind ChatGPT, Claude, Copilot and Gemini. It has no memory of your business unless something gives it one, and it will produce a fluent answer whether or not the answer is right. For a business, the model is the smallest part of any result; the process around it is the rest.

The instruction and the material given to a model in one go. A good prompt is a good brief: the job, the audience, the facts, the format, and what to leave out. Most poor AI output in small businesses is a poor brief, not a poor model.

A model with a chat interface a person types into. Useful for drafting and summarising when a person reads the result. Not an automation, because a person is doing the work of moving the answer anywhere.

A model that has been given tools and permission to act: read a file, send an email, update a record, browse a page. The word is used loosely by vendors to mean anything from a chatbot to software that can change your systems. The question to ask is always: what can it read, what can it do, and what is the record afterwards. An agent runs as the person who signed in, with that person’s access.

A workflow that runs without a person carrying the work between steps. Much of it needs no AI at all (a won job creating a draft invoice). An AI step is added where something has to be read, extracted or drafted. Done properly it carries a human check wherever a mistake would matter.

A connection between two systems you already run, so information entered once appears where it is needed. The unglamorous work that removes the most hours in most small businesses.

A small, real, time boxed test of one workflow on your actual data, judged on a measure agreed before it starts. A demonstration is not a pilot; a pilot has a number, a budget cap and a point at which it stops.

The number a pilot is judged on, taken from your own systems: time from enquiry to quote, error rate on invoices, cost per published article, revenue per enquiry. Never hours saved on its own, because hours are only a saving if you would otherwise have paid for them.

The condition, agreed in writing before work starts, under which the work stops. It exists so that success cannot be redefined afterwards to match whatever happened.

Giving a model your documents to read at the moment it answers, rather than hoping it remembers them. It is how an assistant answers from your price list instead of the internet’s. Retrieval is why your AI keeps forgetting things: it never remembered them; it was never handed them.

A confident, fluent, wrong answer. Models do not know what they do not know. The defence is not a better model; it is a source for every figure and a person who checks before anything is sent.

Text that a model reads (a web page, an email, a filename) that contains instructions the model then follows as if they came from you. The reason an agent that can read the internet and act on your systems needs a boundary and a record.

The written rules for what AI may and may not do in your business, who checks its output, and what record is kept. For a small business it is one page, a tool list and a log, not a committee.

The specific statement of what an AI tool may read, which systems it may reach and what must never happen without someone being told. A boundary is a file a compliance officer can read; a policy is the promise to keep one.

A record of what happened, in order, kept so that it can be checked later. A ledger is an audit trail where each entry is sealed over the one before, so that altering, removing or back dating an entry is detectable, and in a good one, locatable.

A statement, at a point in time, of what a record holds and whether it is intact, made by someone who can be held to it. In ProjxAI Attest it is one page per machine per month; in accounting it is what an auditor signs.

A decision with the reasoning attached: GO, FIX or STOP, made against criteria agreed before the question was worked. The opposite of a recommendation, which is advice with the decision left to you and the criteria left unstated.

Assistants that answer a question directly instead of listing pages. Being cited inside the answer is the new version of ranking, and it rewards plain, factual, well structured pages over clever ones. No agency can guarantee placement inside an AI generated answer; anyone who says otherwise is selling something else.