Zero-shot

What is zero-shot?

Zero-shot means zero examples. You give the model the task and nothing else: “summarise this email”, “translate this into French”, “is this message a complaint?”. You don’t show it beforehand what a good answer looks like.

The fact that this works is one of the big changes brought by large language models. It used to be that sorting messages meant training a model on thousands of labelled examples. Now a general-purpose model understands the instruction and does a reasonable job first time, because it saw vast amounts of similar text during training.

How it works

The model has only the instruction and what it already knows. It works out what you’re asking and replies with what it judges most likely. That brings one advantage and one risk:

• Advantage: it’s quick to set up and uses few tokens, because the prompt is short.
• Risk: the model fills in anything you haven’t specified with its own judgement. If your way of sorting or writing is particular to you, it won’t guess it.

When zero-shot isn’t enough, the next step is usually few-shot: adding two or three examples. For tasks that need several steps of thinking, it also helps to ask the model to reason before answering, with chain of thought.

Example

An estate agency in Brighton gets enquiries from property portals and its own website. It wants to see at a glance whether each contact wants to buy, rent or sell. It tries without examples first:

Read the message and reply with one word only:
buy, rent or sell.
If it isn't clear, reply "unsure".
Message: "{message}"

It gets most messages right, because the difference is obvious. Where it slips is with some investors asking about flats that already have tenants, with a view to buying them. The agency adds an example of that case and moves to few-shot. Starting with zero-shot showed it where the real mistakes were.

What it means for your business

The practical rule is to start simple. Try the task without examples, look at what goes wrong and add examples only where it fails. That way you know what each change brings.

An AI agent like UTIO’s doesn’t rely on the instruction alone: it answers with what it knows about your business. Creator builds it by reading your website and the agent starts in draft mode, so you can review its first replies. If you want to try zero-shot tasks on your own text, the AI API generates and classifies text with your own prompts.