The main types of AI
"AI" covers very different technologies. Knowing which kind you're dealing with tells you what it's good at — and how it's likely to go wrong.
Learning objectives
- Recognise six families of AI used at work
- Match each family to typical failures
- Explain what large language models are and what multimodal means
- Distinguish chatbots from agents, and AI users from AI builders
2.1 Six families of AI you'll meet at work
Most AI you'll meet falls into a handful of families. Knowing which one you're dealing with helps you predict how it might fail.
| Type | What it does | Workplace examples | Typical failure |
|---|---|---|---|
| Prediction & forecasting | Estimates a number or likelihood | Sales forecasts, demand planning, churn risk | Breaks when conditions change (e.g. a pandemic) |
| Classification & scoring | Puts things into categories or ranks them | Spam filtering, CV screening, fraud alerts | Unfair or inaccurate results for under-represented groups |
| Recommendation | Suggests what you might want next | Product suggestions, learning platforms, news feeds | Filter bubbles; optimising for clicks, not value |
| Computer vision | Interprets images and video | Document scanning, quality inspection, ID checks | Poor lighting, unusual angles, demographic accuracy gaps |
| Speech & language | Transcribes, translates and understands text or speech | Meeting transcription, chatbots, sentiment analysis | Accents, jargon, sarcasm |
| Generative AI | Creates new text, images, audio, video or code | Drafting, summarising, coding assistants, image creation | Hallucinations: plausible but false content |
2.2 Generative AI and large language models
The generative AI tools most people use — ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Microsoft Copilot — are built on large language models (LLMs). An LLM has been trained on vast amounts of text to predict what comes next. That simple-sounding skill, at enormous scale, lets it draft, summarise, translate, answer questions and write code.
Many of these models are now multimodal: they can take images, documents, audio and sometimes video as input, and produce more than text as output.
The next course in this path, How AI and Large Language Models Work, explains tokens, training and why LLMs hallucinate — no maths required.
2.3 From chatbots to agents
A chatbot answers when you ask. An AI agent is given a goal and can take several steps to reach it, using tools: searching the web, reading files, sending messages, filling in forms or running code. Agents can save a lot of time, but they also raise the stakes — a mistake can turn into an action, not just a wrong sentence.
Chatbot: "Draft a reply to this customer complaint." You read the draft and decide whether to send it.
Agent: "Handle new complaints in the inbox." The agent reads, categorises, drafts and sends replies, and issues refunds under £50. Now you need clear limits, logging and a human approval step for anything unusual.
2.4 AI you use vs AI you build
Most people at work are AI users (or, in EU AI Act terms, deployers): they use tools someone else built. A smaller group build or customise AI — training models, connecting LLMs to company data, or creating custom assistants. Both groups need AI literacy, but builders carry extra responsibilities for testing, documentation and safety.
Take three AI-powered tools from your work. For each, name its family from the table above and its most likely failure. Which one would cause the most harm if it went wrong — and who would notice?
Key takeaways
- Prediction, classification, recommendation, vision, language and generative AI each fail in characteristic ways.
- ChatGPT, Claude, Gemini and Copilot are built on large language models, increasingly multimodal.
- Agents turn AI mistakes into actions, so they need limits and approval steps.
- Most staff are AI users; builders have extra duties for testing and safety.