Reading an AI tool’s website can feel like reading a foreign language. Tokens, context windows, fine-tuning, retrieval, grounding, agents, guardrails: the words come thick and fast, and they are often used loosely. When you are about to spend money or put company information into a product, vague vocabulary is a real problem, because you cannot compare what you do not understand.
This glossary explains the terms tool buyers meet most often, in plain language and grouped by theme rather than alphabet. For each group you will see what the words mean, why they matter when you are choosing a tool, and the question to ask a vendor. It avoids product claims and works for any tool. Use it as a reference while you read vendor pages and as a checklist for demos.
Key takeaways
- Many AI marketing terms describe ordinary ideas, so ask a vendor to explain any term in one plain sentence and to show it working.
- The terms that most affect buying decisions are about data handling, accuracy, limits, integration and control, not about model size or novelty.
- Words like “trained on your data”, “secure” and “private” can mean very different things, so ask exactly what happens to your information.
- Terms such as agent, automation and assistant are used inconsistently, so judge a tool by what it does, not what it is called.
- Keep a one-page list of the terms and questions that matter for your use case and use it on every vendor call.
Core ideas
These are the terms that sit under almost everything else. If you understand them, the rest of the vocabulary becomes much easier.
| Term | Plain meaning | Why a buyer cares |
|---|---|---|
| Artificial intelligence (AI) | A broad label for software that performs tasks that normally need human judgement, such as recognising speech or writing text. | The label alone tells you nothing. Ask what the tool actually does. |
| Machine learning | Software that learns patterns from examples instead of following only hand-written rules. | Explains why results can vary and why data quality matters. |
| Model | The trained system that produces the output, such as the text or the prediction. | Tools may use their own model or one from another provider; ask which and where it runs. |
| Large language model (LLM) | A model trained on vast amounts of text to predict and generate language. | Powers most chat and writing tools. See what a large language model is. |
| Generative AI | AI that produces new content such as text, images, audio or code. | Its output can be wrong, so it needs checking. |
| Training | The process of teaching a model from data, done before you use it. | Clarify whether your data is used for training. |
| Inference | The act of the model producing an output when you use it. | Relates to speed and cost per use. |
| Algorithm | A set of steps a computer follows. | A generic word; do not accept it as an explanation of how a tool works. |
Foundation model and wrapper
A foundation model is a large, general model that many products build on. A wrapper is a product built on top of someone else’s model, adding an interface, prompts and features. Neither is bad, but it matters: the underlying provider’s terms, reliability and data handling can affect you even if you never deal with them directly. Ask who provides the model.
Parameters
Parameters are the internal values a model learns during training. People sometimes quote their number as a sign of power. For buyers it is a weak guide: a larger model is not automatically better for your task. Judge by results on your own work.
Inputs, outputs and limits
This group covers how you talk to a model and the boundaries on what it can handle.
| Term | Plain meaning | Question to ask |
|---|---|---|
| Prompt | The instruction or question you give the model. | Can we save and share good prompts? |
| System prompt | Standing instructions that shape behaviour for every conversation. | Can we set our own? See how to write better prompts. |
| Token | A small chunk of text, often part of a word, that models read and write. | How does pricing and any limit relate to tokens? |
| Context window | The amount of text the model can consider at once, including your input and its reply. | Is it large enough for our documents? |
| Output length limit | The maximum size of a single reply. | Will long reports be cut off? |
| Temperature | A setting that controls how predictable or varied the output is. | Can we lower it for consistent results? |
| Multimodal | Able to work with more than one kind of input, such as text and images. | Which kinds do we actually need? |
| Latency | How long you wait for a response. | Is it fast enough for real-time use? |
| Rate limit | A cap on how many requests you can make in a period. | Will our team hit the cap? |
Why tokens and context windows matter
Most pricing and limits are expressed in tokens or related units, and the context window decides how much of a long document or conversation the tool can hold in mind. If a tool forgets earlier parts of a long chat or cannot take your whole report, the context window is usually the reason. For plan details, see how to read an AI tool’s pricing page.
Prompt engineering
Prompt engineering is the practice of writing prompts to get better results. It is a skill anyone can learn, not a secret craft. Treat claims that a tool needs special expert prompting with some caution.
Customising a tool
Vendors offer several ways to make a general model fit your business. The differences are important, and the words are easy to mix up.
| Approach | Plain meaning | Good for | Watch out for |
|---|---|---|---|
| Prompting | Giving instructions and examples in the request. | Quick adjustments to style and format. | Needs repeating unless saved. |
| Retrieval, often called RAG | The tool looks up relevant passages from your documents and gives them to the model with your question. | Answers based on your own content that changes over time. | Quality depends on your documents and the search step. |
| Fine-tuning | Further training of a model on examples to shape its behaviour or style. | Consistent tone or format at scale. | Needs good data and effort; does not reliably add facts. |
| Custom assistant | A saved set of instructions, documents and settings for a task. | Repeatable team tasks. | Check what data it can see and who can edit it. |
| Knowledge base | The collection of documents a tool draws on. | Support and internal help. | Out-of-date content leads to out-of-date answers. |
RAG in one paragraph
Retrieval-augmented generation means the tool first searches your documents for relevant passages and then asks the model to answer using them. It helps keep answers tied to your material. It does not make the system infallible, because the search can miss the right passage and the model can still misread it.
“Trained on your data” is a red flag phrase
This phrase can mean very different things: your data is only looked up at the time of a question, or it is used to retrain a model, or it is stored and shared in ways you have not agreed to. Always ask for a precise description. See how to protect company data when using AI tools.
Accuracy and reliability terms
These words describe how and why AI output goes wrong, and what vendors say they do about it.
| Term | Plain meaning | What to do with it |
|---|---|---|
| Hallucination | Output that sounds plausible but is false or unsupported. | Plan checks. See what AI hallucinations are. |
| Grounding | Tying answers to specific source material. | Ask whether sources are shown and can be checked. |
| Citation | A reference to where a claim came from. | Click through; a citation can still be wrong or irrelevant. |
| Bias | Systematic skew in outputs, often reflecting skew in training data. | Test with varied examples relevant to your users. |
| Benchmark | A standard test used to compare models. | Treat as a loose guide; test on your own tasks. |
| Accuracy | How often outputs are correct. | Ask how it was measured and on what data. |
| Confidence score | A number intended to show how sure the system is. | May not match real correctness; do not rely on it alone. |
| Guardrails | Rules or filters that limit what a tool will produce or do. | Ask what they cover and how they are tested. |
| Human in the loop | A person reviews or approves outputs or actions. | Confirm where in the process this happens. |
| Evaluation | Systematic testing of outputs against criteria. | Ask how the vendor and you can test. |
A caution on accuracy claims
A vendor quoting accuracy has measured it on some set of tests. Those tests may not resemble your work. The reliable approach is to run your own sample tasks and check the results, using the habits in how to check AI output for mistakes.
Explainability
Explainability refers to how well you can understand why a system produced an output. Plain-language explanations produced by a model are text like any other, not guaranteed windows into its inner workings. If decisions affect people, ask what records and reasoning the tool keeps.
Agents, automation and actions
This is the part of the vocabulary where words are most loosely used.
| Term | Common meaning | Careful reading |
|---|---|---|
| Assistant | A tool that helps by answering and drafting. | Usually a person stays in control. |
| Chatbot | A conversational interface. | May answer only, or may be linked to systems. |
| Agent | Software that pursues a goal and uses tools to take steps. | Ask what it can do, and what needs approval. See AI agents explained for non-technical readers. |
| Automation | Steps that run without manual effort. | May contain no AI at all. |
| Workflow | A defined sequence of steps, sometimes including AI. | Predictable and easier to test than free-form agents. |
| Tool use or function calling | The model triggering an action in another system. | Check permissions and logging. |
| Integration | A connection to another app. | Check which actions are supported, not just whether a logo is listed. |
| Orchestration | Coordinating several steps or models. | Ask where failures are caught. |
| Autonomy | How much the system does without approval. | Match it to the risk of the task. |
Judge by behaviour
Whatever the tool calls itself, ask what happens when you give it a task: which systems does it touch, what can it change, and where can a person step in?
Data, privacy and security terms
For many buyers this is the most important section, because it decides whether you can use a tool with real business information.
| Term | Plain meaning | Question to ask |
|---|---|---|
| Personal data | Information about an identifiable person. | Do we need to put it in at all? |
| Data retention | How long the provider keeps your inputs and outputs. | Can we set or shorten it? |
| Data residency | The location where data is stored and processed. | Does location matter for our obligations? |
| Encryption | Scrambling data so only authorised parties can read it. | Is it applied in transit and when stored? |
| Access controls | Rules about who can see or change things. | Can we restrict by role? |
| Single sign-on (SSO) | Logging in through a central company account. | Is it supported on the plan we need? |
| Audit log | A record of who did what and when. | Can we view and export it? |
| Opt out of training | A setting to stop your data being used to improve models. | Is it on by default, and can administrators enforce it? |
| Data processing agreement | A contract about how a provider handles personal data for you. | Do they offer one, and has our adviser read it? |
| Subprocessor | Another company the provider uses to handle your data. | Who are they, and are changes notified? |
| Compliance certification | An independent check against a security or privacy standard. | Which one, covering which services, and can we see the evidence? |
Public and private versions
Many vendors have free or consumer versions and business versions with different data terms. Make sure the plan you are testing is the plan you will use, and that staff do not use a personal account for company work. For policy help, see how to write an AI use policy for your team.
Legal wording
Terms such as data protection, compliance and certification have legal or formal meanings that vary by place. This is general information, not legal advice. Ask a qualified adviser to review contracts and data terms for anything sensitive.
Pricing and contract terms
Pricing language can hide the real cost. These terms come up on almost every pricing page.
| Term | Plain meaning | Watch for |
|---|---|---|
| Seat | One user licence. | Minimum seat counts and who counts as a user. |
| Usage-based pricing | You pay according to how much you use. | Costs that rise unpredictably. |
| Credits | Units you buy and spend on tasks. | How many credits a typical task uses, and whether they expire. |
| Fair use | An informal limit on heavy use. | What counts as heavy. |
| Service level agreement (SLA) | A promise about availability and support. | What remedy you get if it is missed. |
| Lock-in | Difficulty leaving once you rely on a tool. | Can you export your data and settings? |
| Proof of concept or pilot | A small, time-limited trial. | Agree success measures in advance. |
Total cost
The subscription is only part of the cost. Add training time, setup, oversight and the time spent checking outputs. The comparison method in how to choose an AI tool helps put all of this on one page.
Marketing phrases to question
Some phrases sound reassuring while saying little. When you meet them, ask for specifics.
| Phrase | What to ask |
|---|---|
| “Enterprise-grade security” | Which controls and certifications exactly, and can we see evidence? |
| “Your data stays private” | Private from whom, stored where, kept for how long, used for what? |
| “Human-level accuracy” | Measured on what tasks, by whom, against which human? |
| “Fully autonomous” | What can it do without approval, and what are the limits? |
| “No hallucinations” | How is that tested, and what happens when it is wrong? |
| “Works out of the box” | With our data and processes, in how many days, with whose effort? |
| “Learns from your business” | What exactly is learned, where it is stored and can we delete it? |
| “Unlimited” | Subject to what fair use or rate limits? |
Ask for a demonstration
The best antidote to vague wording is a live demonstration using your own sample material. If a vendor cannot show a claim working, treat it as unproven.
Using the glossary in a demo
A glossary is only useful if you use it. Here is a simple approach for a vendor meeting.
Before the meeting
- List your use case in one sentence and the data involved.
- Choose the five to eight terms from this guide that matter most for it.
- Write one question for each, using the tables above.
- Prepare two or three sample tasks from your real work.
During the meeting
- Ask the vendor to explain any term you do not know in one plain sentence.
- Ask them to show the feature, not describe it.
- Note which questions get clear answers and which get deflected.
- Ask for the answers to key questions in writing.
After the meeting
- Compare the written answers across vendors.
- Run your sample tasks yourself and check the output.
- Have a qualified person review the data and contract terms.
Templates and checklist
One-page buyer sheet
| Area | Our requirement | Vendor answer |
|---|---|---|
| Model and provider | [what we need to know] | [answer] |
| Context and limits | [document sizes, volumes] | [answer] |
| Customisation method | [prompting, retrieval, other] | [answer] |
| Accuracy testing | [how we will test] | [answer] |
| Data retention and training use | [our rules] | [answer] |
| Access and logging | [roles, audit needs] | [answer] |
| Actions and approvals | [what it may do] | [answer] |
| Pricing model | [expected usage] | [answer] |
| Exit and data export | [our needs] | [answer] |
Glossary checklist
- I can explain tokens, context window and retrieval in one sentence each.
- I know whether our data would be used for training and can get that in writing.
- I have asked what “agent” or “automation” means for this particular tool.
- I have tested accuracy on my own tasks.
- I understand how the tool is priced and how our usage would be measured.
- I have asked about retention, access controls and logging.
- I have noted any marketing phrase I want clarified.
Frequently asked questions
Do I need to understand the technical details to buy an AI tool?
No, but you do need to understand the practical ones: what happens to your data, how accurate the results are on your tasks, what the tool can do on its own and what it will really cost. This glossary is aimed at those.
What is the difference between an AI model and an AI tool?
The model is the trained system that produces output. The tool is the product around it, with the interface, features, settings and terms. Two tools can use the same model and behave quite differently.
Is a bigger model always better?
Not necessarily. The right choice depends on your task, speed needs, cost and data rules. Test on your own examples instead of relying on size claims.
What should I ask first about data?
Ask whether your inputs are used to train models, how long they are kept, who can access them and whether you can delete them. Get the answers in writing.
How can I tell if a term is just marketing?
Ask for a one-sentence plain definition and a live demonstration. If the explanation stays vague or the feature cannot be shown, be cautious.
Where should I start learning more?
Begin with a short explainer on how these models work, then practise writing clear prompts and checking outputs. The guides on large language models and prompts linked in this article are a good start.
Conclusion
AI vocabulary becomes manageable once you group it: core ideas, inputs and limits, customisation, accuracy, actions, data, content and pricing. For each group, there is a small set of questions that matter, and most are about what happens to your data, how you can verify the results, what the tool can do without you and what it will cost in total. Ask for plain explanations and live demonstrations, and be wary of phrases that sound reassuring without specifics.
Keep the one-page buyer sheet handy and reuse it on every vendor call, so you compare like with like. To build on this, read how to choose an AI tool and what a large language model is.
