The term AI gets slapped on everything from your fridge to the stock market, so let's cut through it. When people say AI today, they almost always mean machine learning: software that learns patterns from enormous amounts of data instead of following rules a programmer wrote by hand. Within that, a handful of categories cover nearly everything you will actually touch. Here they are in plain language.
Large language models
Large language models, or LLMs, are the engines behind tools like Claude and ChatGPT. They work with text: you write something, they respond with text that is relevant, coherent, and often genuinely insightful. Under the bonnet they are prediction machines trained on vast amounts of writing, and I'll unpack exactly how they work in the next article, because understanding it will make you dramatically better at using them.
What matters here is what they can do: draft, summarise, analyse, translate, brainstorm, answer questions, follow instructions, and reason through problems step by step. If your work involves words or decisions, an LLM can help with a chunk of it.
Vision AI
Vision models understand images and video. Show one a photo and it can tell you what is in it, read the text off a document, spot the defect on a production line, or describe a chart. Modern LLMs like Claude are multimodal, meaning vision is built in, you can drop a photo of a whiteboard, an invoice, or an error message straight into a chat and ask questions about it. Standalone vision systems also power things like quality inspection in manufacturing and safety monitoring on work sites.
Audio AI
Audio models work with sound in both directions. Speech to text turns spoken words into written ones, which is how meeting transcription works. Text to speech goes the other way, producing natural sounding voices. Put the two together with an LLM in the middle and you get voice agents, AI that can hold an actual phone conversation, answer questions, and book appointments. This category has improved staggeringly fast, and if you have not heard a modern AI voice recently, it will surprise you.
Generative media
Image and video generation models create pictures and footage from written descriptions. They have gone from producing curiosities to producing usable marketing assets in a couple of years. They are a different beast from the models above, you direct them with descriptions and taste rather than instructions and logic, but for content heavy work they are a serious productivity lever.
Coding harnesses and agents
This is the category most people have not heard of, and it is the one I am most excited about. A harness is scaffolding wrapped around a language model that lets it act rather than just talk: read and write files, run code, execute commands, browse, check its own results, and keep working in a loop until a task is done. Tools like Claude Code turn a model that can discuss software into something closer to a junior engineer who actually builds it. The same idea, a model plus tools plus a feedback loop, is what people mean by AI agents more broadly, and it is how AI goes from answering questions to completing work.
Which of these do I actually use?
In D3ploy, the overwhelming majority of what we build for clients runs on LLMs and coding harnesses. Language models do the understanding, drafting, and deciding, and coding harnesses let me build and ship the systems around them at a speed that still feels unfair. Voice agents, which are LLMs wearing an audio front end, come next for the businesses that live on their phones.
In L3ad, our lead generation venture, the mix broadens out, we lean on generative media for content alongside models that analyse performance data and trends, because marketing is a different problem to operations.
The point of this taxonomy is not trivia. It is so that when someone says "we should use AI", you can ask the sharper question: which kind, on which problem? The next article goes deep on the type you will use most, the LLM.