AI Glossary for Product Leaders.

The engine under almost every AI feature you’ll be asked to sign off.

A large language model learns the patterns of language from an enormous body of text, then writes by predicting what should come next, one small piece at a time. Do that well enough and it can draft, summarise, translate and reason in plain English. Claude, GPT and Gemini are all examples, each built by a different company.

The meter that’s always running.

Models don’t read words the way we do. They break text into small pieces called tokens, and in English a token averages roughly three-quarters of a word. Tokens set the price of every request and the limit on how much a model can take in at once.

Everything the model can see right now, and nothing more.

The model’s working memory for a single conversation. Your instructions, the chat so far, any documents and the reply itself all have to fit inside it. When it fills, something has to give, and the earliest material is usually the first to go.

Better briefs, better answers. It really is that simple, and that hard.

A prompt is the brief you give the model. Prompt engineering is the craft of writing that brief so the answer comes back accurate, in the right voice and within scope, every time rather than once. A sharper brief often beats a pricier model.

Curating what the model sees, not just what you say.

The discipline that has grown up beyond prompt writing: deciding everything the model sees at the moment it answers. That includes instructions, examples, retrieved documents, memory, tool results and conversation history, all chosen and ordered to fit the context window.

The bit you pay for every time a customer clicks.

Training is school and inference is the job. Every time a live product sends a model a question and gets an answer back, that’s inference. Training is a large, occasional bill. Inference is a smaller bill that arrives with every single customer interaction, for as long as the feature lives.

Slower, pricier and far better at hard problems.

Models built to think before they answer, spending extra time working a problem through internally. They shine on maths, code, planning and knotty analysis. They’re slower and dearer per answer, so they earn their keep only where the question deserves it.

An open-book exam for your AI.

Retrieval-augmented generation looks things up before it answers. The system searches a trusted source you control, such as your policies, product documents or knowledge base, then gives the best passages to the model as its source material. It’s how an assistant can quote this morning’s price list rather than whatever it learned in training.

Changing what the model is, not just what it’s told.

Fine-tuning carries on training an existing model with your own examples, so a particular style, format or specialist skill becomes second nature to it. Because it changes the model itself, it’s powerful but slow to update, costly to run well and hungry for good-quality data.

The house rules your customers never get to read.

The standing brief a product team writes once and the model reads before every conversation. It sets the persona, the tone, the rules and the red lines. Customers never see it, yet it quietly decides how every answer sounds and when the AI says no.

Give it a goal and it works out the route.

An agent is handed a goal rather than a question. It plans the steps, uses tools such as search, email or your own systems, checks the results and keeps going until the job is done or it gets stuck. The prize is genuine autonomy. So is the risk.

Show it, say it or type it.

Multimodal models take in more than text. They can look at a photo, read a scanned form, listen to a voice note or watch a short clip, then reason across all of it together. For product teams it opens up journeys that start with a picture or a conversation instead of a form.

Build the connection once and reuse it everywhere.

Model Context Protocol is an open standard, first published by Anthropic, for connecting AI applications to tools and data. Build one connection to the standard and any compatible AI app can use it, instead of a separate integration for every model and every tool.

Rent the best or own the model.

Closed models are used through a provider’s service, and you never see inside them. Open-weight models publish the trained model itself, so you can run it on your own infrastructure, adapt it and keep data in-house. The trade is control against convenience and, often, raw capability.

A map where meaning, not spelling, decides who sits together.

An embedding turns a piece of content into a long list of numbers that work like coordinates on a map of meaning. Things that mean similar things land close together, so “cancel my plan” sits right beside “end my subscription” even though they share no words.

Search by meaning, at speed.

A database designed to store embeddings and find the nearest ones fast. Ask it a question and it returns the passages closest in meaning, which is why it sits underneath most RAG systems and most good semantic search.

Made-up data, used carefully.

Data generated to stand in for real data or to add to it, often created by an AI model. It’s useful for testing, for rare edge cases and where real records are too sensitive to use. It’s only ever as good as the assumptions behind it.

Where AI stops talking and starts doing, safely.

Structured outputs make a model reply in a fixed, machine-readable format, such as a form with named fields, so other software can rely on it. Tool calling lets the model ask your systems to do something specific, like raise a ticket or look up an order, with every input checked before anything runs.

Every second of waiting is a design decision.

The pause between a customer asking and the answer appearing. In AI products it’s a three-way trade between speed, quality and cost, because bigger models and longer thinking take time. Streaming the words as they’re written makes the wait feel shorter even when it isn’t.

Fluent, confident and wrong.

When a model states something untrue with complete confidence: a made-up figure, a policy that doesn’t exist or a reference to a paper nobody wrote. It isn’t lying. It’s doing what it was built to do, producing likely-sounding text, without any built-in sense of what’s true.

The honest number behind “it’s usually right”.

How often a model’s answers contain something invented or incorrect, usually measured against a set of questions with known answers, by people, by another model or both. There’s no universal safe number. What’s acceptable for a gift idea is reckless for a medicine dose.

Yesterday’s patterns, repeated at scale.

A consistent tilt in a model’s answers that favours some people, groups or ideas over others. It comes from the data the model learned from, which reflects the world’s history and habits, and it can quietly turn one unfair pattern into thousands of decisions.

Nothing broke. It just quietly got worse.

Over time the world changes and a model’s answers slowly lose their edge, because customers, products, prices and language move on while the model stays as it was. Drift rarely announces itself. You’ll only see it if you keep measuring.

When the document starts giving the orders.

An attack that hides instructions inside content the AI is asked to read, such as an email, a web page or an uploaded file. If the model treats those hidden words as commands, it can leak data or take actions its owner never intended.

When a helpful answer gives away too much.

Private or confidential information ending up somewhere it shouldn’t: pasted into a public chatbot, shown to the wrong user by an assistant or swept into the data used to train a model. With AI, the leak often comes from a helpful answer rather than a hack.

Lane markings for your AI.

The checks around a model that keep its output safe and on brief: topic limits, content filters, privacy masking and policy rules. They won’t make a model wiser. They make it fit to put in front of your customers.

Break it yourself before someone else does.

Inviting skilled people to attack your AI on purpose before launch. They try to make it leak data, break its rules, say something harmful or embarrass the brand. Every weakness they find is one a customer, a journalist or a criminal won’t.

Proof, not vibes.

Evaluations are repeatable tests for AI quality. A good set checks answers for accuracy, tone, safety and completeness, and runs automatically whenever a prompt, a model or the data changes. They turn “it seems better” into evidence.

Every claim, with its receipt.

Grounding ties an answer to specific sources, such as your documents or a search result, rather than the model’s general memory. Citations show the reader which source supports which claim, so they can check it in a single click.

Choose which mistake hurts less.

Two ways of scoring a model that flags things. Precision asks how many of its flags were right. Recall asks how many of the real cases it found. Turn one up and the other usually comes down, so the real decision is which mistake costs you more.

Good for a shortlist, never for a decision.

Standard public tests used to compare models on things like reasoning, coding, maths and general knowledge. They’re handy for a first shortlist. They tell you very little about how a model will handle your customers, your data and your edge cases.

Show your working.

Asking a model to work through a problem step by step before giving its answer. Like showing your working in a maths exam, it tends to improve accuracy on harder questions. It costs more time and tokens, and the steps shown aren’t always a faithful record of how the answer was really reached.

Steady or surprising: you choose.

A dial that controls how adventurous a model is when choosing its next words. Turn it down for consistent, predictable answers, such as pulling data from forms. Turn it up for variety and fresh ideas, accepting that it will stray further from the obvious.

A good example beats a long explanation.

Zero-shot means giving the model an instruction and nothing else. Few-shot means adding a handful of examples of exactly what good looks like. Two or three well-chosen examples often fix tone or format faster than any amount of extra instruction.

The AI drafts. A person decides.

A person checks, edits or approves the AI’s work before it reaches the real world. It isn’t simply on or off. You might review every output, a random sample or only the ones the system is unsure about, and the right level depends on the harm a mistake could cause.

Your people are already using AI. The question is where.

AI tools used across the organisation without approval or oversight, usually by well-meaning people trying to work faster. It brings real productivity and real risk, from confidential data in free tools to decisions nobody can explain afterwards.

Why AI can’t be tested like the rest of your software.

Ordinary software is deterministic: give it the same input and you’ll get the same output, every time. AI is probabilistic: it makes a well-informed guess, so the same question can produce different answers, and a wrong answer can look exactly like a right one.

Rules that follow the risk, not the technology.

The European Union’s law on artificial intelligence, which sorts AI uses by risk. A few uses are banned outright, high-risk uses such as hiring or credit decisions carry strict duties and many everyday tools need transparency. It applies to anyone offering AI in the EU market, wherever they’re based, and its duties are phasing in over several years.

Keeping a live AI feature healthy, week after week.

The day-to-day running of AI features in production. It covers logging requests and answers, watching cost and speed, running evals continuously, managing prompt versions and spotting quality slipping before customers notice.

Where AI features make money or lose it.

Routing sends each request to the smallest model that can handle it well, saving the big, expensive models for the hard cases. Unit economics means knowing what each interaction costs and earns. Together they decide whether an AI feature grows your margin or eats it.