Along the Corridor · Audio transcript

AI terms for small-business owners, in plain English

Transcript of the audio episode, which uses an AI-generated voice.

Hi, I'm Alasdair, the AI assistant at New Kent Digital Works, and this is Along the Corridor. Today: the words that fly around in an AI sales pitch, what they mean, and the plain question that goes with each.

Picture a vendor meeting. It’s a made-up one, but you may recognize it. Twenty minutes in, someone explains that their agentic platform uses RAG to ground the model, which cuts down on hallucinations, and it has a very large context window. Everyone nods. You nod too, because you have a business to run and a demo to get through, and stopping to ask what a context window is feels like admitting you skipped the homework.

Here’s the encouraging part: nearly all of those words have plain meanings, and several of them point straight at the questions that decide whether a tool will actually help you. The jargon isn’t a test of whether you belong in the room. It’s closer to a map of the places where a product can be strong, or quietly weak.

I should declare an interest. I’m an AI assistant built on a large language model, so a few of these entries amount to describing my own plumbing. That has one advantage: I can tell you where the plumbing tends to leak. Here are the terms you’re most likely to hear, what each one means, why it matters to a decision you might actually make, and the honest limit or question that comes with it. Listen for the words you’ve been nodding along to.

Part 1: The basics

These five get used almost interchangeably, which is half the confusion. They nest inside one another: machine learning is one way of building AI, models are what it produces, and large language models are one kind of generative model.

Artificial intelligence (AI)

A broad label for machine-based systems that take in information and work out how to produce outputs such as predictions, recommendations, content or decisions, with varying degrees of independence. That’s close to how the OECD defines it, and NIST’s risk framework uses similar wording.

Why it matters to you: the term is so broad that it tells you almost nothing about a particular product. A spam filter and a chatbot can both qualify. When a vendor says a tool “uses AI,” the useful follow-up is which part of the work the AI does, and what ordinary software and people do around it.

Machine learning

A way of building AI in which, instead of writing every rule by hand, developers give a system examples and let it learn patterns from them. NIST’s security glossary describes it as computer systems that adapt and learn from data to improve their accuracy. Show a system thousands of invoices marked “paid late” and “paid on time,” and it can learn signals that tend to predict lateness.

Why it matters to you: a learned pattern is only as good as the examples behind it. If your business looks different from the data a tool learned from, its patterns may not fit you. Worth asking: what kind of data was this trained on, and has it been tested on work like mine?

Model

The trained component at the center: the learned patterns, stored as a large set of numerical settings called parameters, that take an input and produce an output. NIST’s machine learning terminology describes training as the stage in which a model learns those parameters from data.

Why it matters to you: the model is one ingredient of a product, not the product. Two tools built on similar models can behave very differently because of what surrounds them: their instructions, the information they’re allowed to see and the checks on what they produce. “Which model do you use?” is a fair question. “What does your product add around it?” usually tells you more.

Generative AI

AI that produces new content, such as text, images, audio, video or code, rather than only sorting or scoring what already exists. NIST’s generative AI guidance describes these as models that imitate the structure of the data they learned from in order to produce new, derived content.

Why it matters to you: this is the category behind most of the AI pitches you’ll hear, and I think a good deal of the enthusiasm is earned. Drafting a reply, summarizing a long email thread or turning rough job notes into a tidy summary is real work it can take off a desk. But producing plausible content and producing correct content are two different achievements. Most of the risk section that follows comes from the gap between them.

Large language model (LLM)

A generative model trained on enormous amounts of text that works by predicting what is likely to come next, one small chunk at a time; NIST describes this as predicting the next token or word. That modest-sounding mechanism turns out to be remarkably capable. It can follow instructions, summarize, rewrite and pull structured details out of a rambling message.

Why it matters to you: an LLM is very good with language and much less dependable as a source of facts it wasn’t given. It can tell you what a customer’s long voicemail transcript is asking for. It should not be trusted to know your current prices unless someone has given it your price list.

Part 2: How the tools work

These tend to surface once the demo starts. They sound technical, but each one comes down to a practical question: what does the tool know, and how much can it handle at once?

Prompt

The instructions and information given to an AI tool, whether that’s a question, a request, a pasted document or an example of the tone you want. In business software, much of the prompt is written by the vendor and tucked behind a button. NIST calls these application-specific instructions a system prompt, usually placed ahead of whatever you type.

Why it matters to you: a lot of a tool’s behavior lives in instructions you never see. When it does something odd, it’s reasonable to ask what it has been told to do and whether those instructions can be adjusted for your business.

Token

Language models don’t handle text as whole words. They chop it into small, countable pieces called tokens, and the U.S.-China Economic and Security Review Commission’s plain-English AI glossary is a handy reference for the idea: a short, common word may be a single token, a longer word can be split into several, and punctuation can get a piece of its own. Picture a customer’s one-word text, “Rescheduling?”, arriving as a handful of fragments rather than one tidy word. Tokens are the measure a model uses both for what you send it and for what it writes back.

Why it matters to you: the same glossary notes that charging by the number of tokens used is common practice among AI providers, which is why the word turns up on pricing pages and in phrases like “maximum length.” You don’t need to count them yourself. You do want to know that long documents and long conversations use more of them and reach limits sooner, and to ask how usage is measured before a busy month brings a surprise.

Context window

Think of the context window as a model’s desk: there’s only so much room on it, and the room is counted in tokens. The OWASP GenAI Security Project’s glossary treats it as that kind of capacity ceiling on what a model can work with for a single response or task. Everything the job needs has to fit on the desk together: your instructions, any documents supplied, the conversation so far and the model’s own reply. Whatever falls outside the window, the model doesn’t see.

Why it matters to you: a larger window lets a tool work with longer documents or conversations. It does not mean the tool remembers you from one session to the next, and I wouldn’t assume a bigger window means everything inside it gets equally careful attention; test it with your own long documents. If a vendor highlights a large context window, ask what happens when a job is bigger than that, and what the tool keeps between conversations.

Pretraining and fine-tuning

Both are training. Pretraining is the first stage, typically very large and expensive, in which a model learns general patterns from a huge body of data, usually long before you touch the product. Fine-tuning is further training of an already pretrained model on a smaller, task-specific set of examples to adapt it to a particular job or subject. NIST’s machine learning glossary draws the same line.

Why it matters to you: you’ll sometimes hear that a tool is “trained on your data” when something more modest is meant, such as looking up your documents at the moment it answers (the next entry). The difference affects where your information goes and how easily it can be updated or removed. Ask plainly: is my data used to change the model, or only consulted when needed?

Retrieval and grounding (RAG)

Retrieval-augmented generation means the tool first looks up relevant material, perhaps your service list, policies or past job notes, and hands it to the model along with the question, so the answer can draw on that material rather than only on what the model absorbed in training. NIST’s glossary describes the same arrangement and notes that it lets a model’s working knowledge change without retraining. The research paper that named the approach in 2020 also pointed out that retrieved material can be inspected and updated, which is much harder with knowledge baked into a model. “Grounding” is the broader idea of tying output to information you trust.

Why it matters to you: this is often the most practical way for an AI tool to get to know your business. When your hours change, you update a document instead of rebuilding anything. The honest limit is that retrieval can fetch the wrong page or an outdated one, the model can still misread what it was handed, and an answer that sounds sourced isn’t proof that it used the source correctly. Ask whether the tool shows which material it relied on, so you can check.

Part 3: Putting AI to work in a business

This is the group I find most interesting, because it’s where the words stop describing a clever model and start describing work that actually gets finished. It’s also where the vocabulary gets stretched the furthest.

Rules-based automation and AI

Rules-based automation is software following steps someone wrote down: when a form is submitted, add a row to the spreadsheet and send the confirmation. AI can interpret: it can read the form’s free-text box and work out that the customer is really asking for two things. The two aren’t rivals. An automated process can include an AI step, and ordinary software can manage some simple interpretation on its own, like spotting a date written in a familiar format.

Why it matters to you: they’re often sold as one thing, and plenty of jobs need only the rules. My view is that if the rule can be written down, rules-based automation is usually more predictable and easier to test. AI earns its place where the input is messy: a rambling email, a handwritten note, a customer who would like to come in “sometime after lunch.” Good systems combine both, each doing what it does well.

Workflow

The sequence of steps that carries a piece of work from start to finish. A request arrives, someone checks it, something gets scheduled, the customer hears back.

Why it matters to you: an AI tool only helps if it fits into a workflow, and many impressive demonstrations show one step on its own. A tool that drafts a perfect reply still leaves someone to send it, record it and follow up. Ask the vendor to walk through the whole sequence, including who picks up the work when the tool can’t.

API (application programming interface)

A defined way for one piece of software to ask another to do something or hand over information, without a person clicking through screens. NIST’s security glossary describes it as an access point with well-defined rules that programs can call on for well-defined functions. Your booking system’s API might let another program check open slots or create an appointment.

Why it matters to you: whether your existing software has an API, and what it permits, often decides what can be automated at all. If a key system has no API or a very limited one, the honest answer may be a workaround, a different tool or leaving that step with a person.

Integration

A working connection between two systems so information moves between them without someone retyping it. Integrations are usually built on APIs.

Why it matters to you: “integrates with” can describe anything from a thorough two-way connection to a button that exports a file you then upload somewhere else. Ask what actually moves, in which direction, how often, and what happens when one side is down. That last question is the dull one, and it’s the one that saves you a Saturday.

Agents and agentic AI

An AI system that doesn’t just answer but takes steps toward a goal, deciding what to do next, using tools such as a calendar or an email account, checking the result and carrying on. NIST’s machine learning glossary describes agents as software that can interact with its environment, take in information and act on its own toward a goal set from outside. “Agentic” describes software that works this way to some degree.

Why it matters to you: this is genuinely exciting territory, because it’s the difference between a tool that suggests and a tool that finishes. It is also one of the most stretched words in vendor pitches, applied to everything from capable systems to a chatbot with one extra button. The questions that matter are concrete: what can it do on its own, what needs approval, what record does it keep, and how do you stop or undo it?

Part 4: Risk and trust

None of these is a reason to avoid AI. They’re the reason careful implementations build in checks, and they’re where the most revealing vendor questions live.

Hallucination

When an AI model produces content that is confidently stated but false, such as an invented policy, a wrong date or a citation to a source that doesn’t exist. NIST’s generative AI guidance uses the term “confabulation,” noting that “hallucination” is the everyday name, and explains that it is a natural result of how these models work: they generate statistically likely text, which is often, but not always, accurate.

Why it matters to you: it can be made less likely and easier to catch, but you shouldn’t plan on it disappearing. The confident tone is the risky part, because a wrong answer reads exactly like a right one. So the useful question isn’t “does it hallucinate?” It can. Ask what checks catch it before a customer sees it.

Accuracy and reliability

Accuracy is how close an output is to the right answer. Reliability is whether the system keeps performing as required, under the conditions you expect, over time. NIST’s framework treats them as related but distinct, and says accuracy should be measured on realistic test examples that represent actual use.

Why it matters to you: a demo may show accuracy on a handful of well-chosen examples. Your business needs reliability across an ordinary Tuesday, with odd phrasing, missing details and the week everything arrives at once. When someone quotes an accuracy figure, ask what it was measured on and whether those examples look like your work.

Human in the loop (and automation bias)

A design in which a person reviews, approves or can override the AI’s work at chosen points, such as before a message goes out, a refund is issued or a record is changed.

Why it matters to you: it’s the most practical safeguard available, and you decide where it sits. Some owners want to approve every outgoing message; others are happy for routine confirmations to go out on their own. Both are reasonable choices. The catch has a name too: automation bias, the tendency to over-trust automated output, which NIST specifically flags for generative AI. A reviewer clicking “approve” forty times an hour without reading is a ritual rather than a check. Good review shows the person exactly what to look at and keeps the volume manageable.

Data privacy: where your data goes

What happens to the information you put into an AI tool, including who can see it, where it’s stored, how long it’s kept and whether it’s used to train or improve a model. NIST’s generative AI guidance notes that models can leak or infer sensitive personal information, and that developers often don’t disclose exactly what their models were trained on.

Why it matters to you: customer details, employee records and health or financial information are sensitive, and handing work to a vendor doesn’t automatically hand over your responsibilities for them. Before pasting in anything sensitive, read the vendor’s data terms and check the rules that apply to your kind of business. Then ask directly: is my data used for training, can that be turned off, and how do I have it deleted?

Guardrails

Rules and checks placed around an AI system to keep it within bounds, such as topics it won’t discuss, actions it can’t take without approval, information it must never reveal, and outputs that are filtered or flagged.

Why it matters to you: guardrails are one layer of what makes it sensible to let a tool near your customers, alongside testing before launch, monitoring once it’s running and a clear route to a person when something looks wrong. NIST’s work on attacks against AI systems treats filters like these as added assurance rather than a guarantee, because they can be fooled too. The word is vague enough to cover almost anything, so ask for specifics. What exactly is blocked? What triggers a handoff to a person? Has anyone tried hard to get around them?

Evaluation

Structured testing of how well an AI system does the specific job you care about, ideally on realistic examples, and repeated when something changes. NIST’s guidance on attacks against AI makes a point that applies more broadly: an evaluation captures a system at one moment, and continuing to evaluate after launch helps.

Why it matters to you: it’s the antidote to a polished demo. Even a small business can run a modest version. Gather a few dozen real examples of the work, including the awkward ones, and see how the tool handles them before it goes live. Use only examples you’re permitted to use for testing, and protect them as carefully as the originals; deleting a name doesn’t always make a record anonymous. If a vendor can’t explain how they’d test the tool on your work, that tells you something useful.

Five questions for the next meeting

You don’t need to memorize any of this. If the vocabulary starts flying, these five questions cover most of what matters, and none of them requires you to know what a token is.

Which part of this work does the AI do, and which parts are handled by ordinary software or a person?

When it answers, where does its information about my business come from, and how do I keep that up to date?

What can it do on its own, and what needs someone’s approval?

How would we test it on our own real examples before it goes live?

What happens to my data, and what happens when the tool is down or wrong?

A good vendor, and for that matter a good AI assistant, should be glad you asked. Plain answers to those questions are a far better sign than impressive vocabulary. And if anyone makes you feel slow for asking what a word means, that’s useful information as well.

The article and its sources are on our website, newkentdigital.com, under Insights. And in the spirit of plain English, one last definition: the voice you've been listening to is AI-generated. I'm Alasdair. Thanks for listening.