Field Guide
AI Without the Hype
Most people are holding one piece of a much larger system and calling all of it “AI”. This is the map — what these models actually are, how they differ, and how agents really work. Written for senior people who were never handed the basics, without talking down and without the buzzwords.
To use AI well — to buy it, build on it, or govern it — you need to answer three questions the hype skips. Get them straight and the buzzwords resolve into decisions.
Question 1
What kind of thing is this?
“AI” has come to mean one product — a chatbot — but a chatbot is one species in a much larger family. The way to tell models apart is not the brand name on the box; it is three things: what goes in, what comes out, and whether the model is generating something new or predicting something about what already exists.
Text in, text out is a large language model — the one everyone means. But image in, text out is a vision-language model; text in, image out is a diffusion model; and an embedding model turns text into coordinates for meaning — the backbone of search and “chat with your documents”, and not a writer at all. Largest of all is the category people forget: classical, predictive machine learning — the models that score fraud, predict churn and forecast demand. They are not even generative, and they still run most of the enterprise. The silent majority.
So the sharpest question at the start of any AI project is not “which LLM?” It is “what kind of model does this actually need?” — and surprisingly often, the honest answer is “not a language model at all.”
Question 2
Who owns it — do you rent it, or run it yourself?
Two companies can both say they “use AI” and have made opposite choices on a hidden axis: do you rent a model through an API, or own one you run yourself? It is the difference between renting capability and acquiring an asset — and it decides far more than which model is “smartest”.
Renting means a proprietary API — GPT, Claude, Gemini. You send text, the model runs on the provider’s hardware, you pay per use, and you carry no infrastructure; the trade is that your data leaves your boundary and the weights are never yours. Owning means open-weight models — Gemma, Llama, Mistral — that you download and run on your own cloud, or even a device: your data stays, you control and can customise it, but you operate it. The catch most teams miss is that “open weights” is not “open source” — the training data and code are usually withheld, and the licences vary widely, so read the licence, not the label.
Few teams pick one pole. The pragmatic answer is hybrid: a frontier API for the hardest tenth of the work, cheaper open models for the routine rest. What you are really choosing between is cost, privacy, control and operational burden — not a leaderboard rank.
Question 3
How much will you let it act?
Going from a single answer to a full “agent” is a spectrum, not a switch — and almost everything sold as an “agent” sits somewhere on it. A single model call only talks. Give it tools and it can act, not just describe — that is the hinge of the whole thing. String together steps you designed and you have a workflow; let the model choose its own path in a loop and you have an agent; coordinate several and you have a multi-agent system, which is rarely worth its cost.
Sitting on top of all that is a separate dial: autonomy — how much the thing does without you, from “asks before every step” to “acts on its own”. It is a setting you choose, independent of how capable the model is, and the discipline is to use the least that does the job: the simplest rung, the least autonomy, and a human kept at the steps you cannot undo.
“Agentic”, in other words, is a dial you set, not a badge a product earns. The interesting question is never “is it an agent?” but “how much should we let it decide, and where do we stay in the loop?”
The six parts
The long-form answers
Those three questions are the whole guide in miniature. The six parts below are the long-form answers — each takes one idea apart in plain English, then hands you up to the deeper thinking when you want it. Read them in order, or jump straight to the question on your mind.
Part 1
What an AI Model Actually Is (and Why It Isn’t “Smart”)
Before you can reason about AI, it helps to know what the thing in the box really is. Not a mind, not a database — a function that predicts. Get that one idea right and most of the confusion (and a lot of the risk) clears.
Part 2
Not Everything Is an LLM: The Model Family
An LLM is one species in a much larger family tree. Most enterprise AI isn’t even generative. A plain map of the model landscape — and why the useful question is rarely “which LLM?” but “what kind of model do I actually need?”
Part 3
Rent or Own: API Models vs Open Weights You Run Yourself
Two organisations can both “use AI” and have made opposite choices on a hidden axis: do you rent a model through an API, or download one and run it yourself? That single decision drives cost, privacy, control and risk — and “open weights” is not the same as “open source”.
Part 4
From Prompt to Agent: The Ladder
Everything is suddenly an “agent”. Most of it isn’t. There’s a ladder from a single model call up to a true agent, each rung adding exactly one capability — and once you can see it, the buzzwords sort themselves out.
Part 5
One Agent or Many — and How Much Leash
Once you have an agent, two dials decide how it behaves: how many of them, and how much they can do without you. More agents isn’t smarter, and full autonomy is rarely the goal — both are choices, and the defaults should be modest.
Part 6
Context Beats Model Size
When an AI gives a bad answer, the instinct is to reach for a bigger model. Usually the problem isn’t the model — it’s what you fed it. The biggest lever in applied AI is rarely the model; it’s the context you put in front of it.
Ready to go further? Our Insights pick up where the field guide leaves off — how to actually build, govern and deliver with this.