AI & Machine Learning
The Machine Learning Glossary Every Business Owner Needs
Last updated:
How to use this glossary
Jargon is not always a smokescreen, but it makes it hard to ask sharp questions. The terms below are the ones we hear business owners stumble over most, arranged by when they tend to come up rather than alphabetically. Each has a plain definition and a line on why you should care.
For deeper reading, several terms link to longer explainers in this series. If you want the big picture first, start with our business guide to machine learning.
Terms you will hear in the pitch
| Term | Plain meaning | Why it matters to you |
|---|---|---|
| Artificial intelligence (AI) | Broad label for software doing tasks that seem to need human judgement | So broad it tells you little; ask what the system actually does |
| Machine learning (ML) | Software that learns patterns from examples rather than following hand-written rules | Needs historical data and ongoing maintenance |
| Deep learning | Machine learning using large neural networks | Strong on images, audio and text; often overkill for spreadsheets |
| Large language model (LLM) | A very large model trained on text that can read and write | Arrives pre-trained; you steer it rather than train it |
| Generative AI | Models that produce new text, images or audio | Useful for drafting; needs checks because outputs can be wrong |
| AI agent | A model that uses tools in a loop to complete a multi-step task | Needs firm limits on what it may do unsupervised |
| Predictive analytics | Using data, often with ML, to forecast outcomes | Frequently the same thing as ML, marketed differently |
Terms about the data
| Term | Plain meaning | Why it matters to you |
|---|---|---|
| Training data | Past examples the model learns from | Its quality caps the model's quality |
| Label | The known answer for each training example | No labels, no supervised learning |
| Feature | One input the model uses, such as days since last order | Good features usually matter more than the algorithm |
| Supervised learning | Learning from examples with labels | Measurable, but needs recorded outcomes |
| Unsupervised learning | Finding structure in data without labels | Useful for grouping; harder to prove value |
| Test set | Data held back to check the model on unseen cases | Test figures are only honest if this was kept separate |
| Data leakage | Using information during training that would not exist at prediction time | Produces great test scores and poor real results |
| Synthetic data | Artificially generated data mimicking real data | Can fill gaps; risky as the only source |
| Embedding | A list of numbers representing the meaning of text or images | Powers search by meaning and recommendations |
More on labels and data in training data explained.
Terms from the build
| Term | Plain meaning | Why it matters to you |
|---|---|---|
| Model | The trained file that turns inputs into predictions | One part of a larger system you are paying for |
| Algorithm | The method used to learn, such as a decision tree | Less important than data and features for most projects |
| Classification | Predicting a category, such as yes or no | Output is usually a probability you set a threshold on |
| Regression | Predicting a number, such as units sold | Ask for a range as well as a single figure |
| Clustering | Grouping similar items without predefined categories | Always finds groups; someone must judge if they mean anything |
| Baseline | A simple rule the model must beat | If the model cannot beat it clearly, do not buy the model |
| Overfitting | The model memorised training data instead of learning patterns | Main reason models disappoint after launch |
| Fine-tuning | Adapting a pre-trained model with your own examples | Much cheaper than training from scratch |
| Retrieval-augmented generation (RAG) | Giving a language model relevant documents when it answers | The usual way to make an AI assistant use your own information |
| Hyperparameters | Settings chosen before training, such as model size | Technical detail; tuning matters less than data |
For RAG specifically, this explainer goes further.
Terms for measuring results
| Term | Plain meaning | Why it matters to you |
|---|---|---|
| Accuracy | Share of predictions that were right | Misleading when the outcome is rare |
| Precision | Of the cases flagged, how many were correct | Low precision means many false alarms |
| Recall | Of the real cases, how many were found | Low recall means many misses |
| Confusion matrix | Table of right and wrong predictions by type | Put costs on it to get a business case |
| Threshold | The score above which the model says yes | A business decision, not a technical default |
| False positive | A false alarm | Costs staff time or customer goodwill |
| False negative | A miss | Costs whatever the model was meant to prevent |
| Evaluation set | A fixed set of real examples used to score a model or AI feature | Lets you compare versions fairly over time |
Terms after launch
| Term | Plain meaning | Why it matters to you |
|---|---|---|
| Deployment | Putting a model into live use | Where most of the engineering cost sits |
| Inference | The model making a prediction | Drives running costs at volume |
| Drift | The world changing so the model's patterns go stale | Silent loss of accuracy without monitoring |
| Monitoring | Tracking inputs, predictions and accuracy over time | How you find drift before customers do |
| Retraining | Rebuilding the model on newer data | A recurring cost to budget for |
| MLOps | Practices and tools for running models reliably | The difference between a demo and a dependable system |
| Explainability | Being able to show why a prediction was made | Needed for trust, audits and some regulations |
| Human in the loop | A person reviews some or all predictions | Lets a less accurate model still be very useful |
Two regulatory terms worth knowing
- EU AI Act. European regulation that classifies AI uses by risk and places duties on providers and users, with obligations phasing in over several years. It affects UK firms that sell into the EU.
- Automated decision-making. Under UK and EU data protection law, decisions with significant effects on people made solely by automated means come with extra rights, including in many cases the right to human review.
If a project touches either, get advice early. SpiderHunts designs enterprise AI systems with audit logs and human review built in, because retrofitting them is painful. None of this glossary is legal advice, though, and a specialist solicitor is the right person for the final word.
Frequently asked questions
What is the difference between AI and machine learning?
AI is the broad goal of software performing tasks that seem to need human judgement. Machine learning is the main technique used to achieve it today, where software learns from examples rather than following hand-written rules.
What does it mean to train a model?
It means showing a learning algorithm many past examples with known answers so it adjusts its internal numbers to predict those answers well. The result is a model that can make predictions on new cases.
What is an algorithm versus a model?
An algorithm is the recipe for learning. A model is what you get after applying that recipe to your data: a specific trained object that makes predictions for your business.
Which terms should I understand before hiring a machine learning supplier?
At minimum: training data, baseline, overfitting, precision and recall, drift and retraining. Those six cover whether the model will work, how you will know, and what it will cost to keep working.
Heard a term in a proposal that is not on this list?
Send it to us with the sentence it appeared in. We will explain what it means for your project, and whether it matters.