Zero-shot Learning

Zero-shot learning is what makes it possible to launch an agent on Monday for a business that did not exist in any training set. You describe the task in plain language, such as 'classify each message as booking, billing, support or other', and the model does it without a single labeled example. In a chat or voice agent, most of what the model does on day one is zero-shot: intent classification, entity extraction, summarization and reply writing from a description and your documents.

How it works. The model was pretrained on enough text that it has seen most tasks in some form. A description in the prompt (the task, the labels, the format of the answer) is enough to activate that general capability. The input is the instruction plus the live message; the output is the model's best interpretation. There is no training step, which is the point, and no feedback loop, which is the limit: if the description is ambiguous, the model resolves the ambiguity its own way, consistently and silently. Quality depends almost entirely on how precisely the task is described and how well the labels match the way customers actually talk.

Why it matters for launch speed and risk. Classic NLU platforms needed dozens of sample utterances per intent, collected and labeled before the first test. Zero-shot removes that, which is why a website chat agent can be live in minutes after documents are uploaded. The risk is the mirror image: because no examples were required, nobody checked whether the model's idea of 'billing' matches yours. A description that seems clear to the author can split in two directions on real traffic, and the error rate is invisible until someone reads transcripts.

Example. A restoration company adds a new intent, 'mold inspection', to its after-hours voice agent with a one-line description: 'Caller suspects mold or reports a musty smell and wants an inspection.' That evening the agent correctly routes 'there's black stuff on the bathroom ceiling' and 'the basement smells damp since the flood'. It also routes 'my tenant says the AC smells weird' to mold inspection, where the owner wanted a referral to the heating and cooling partner. Two transcripts later the description gains the clause 'not odors from appliances or air conditioning', and the misroute stops. No examples were needed, but the review was.

Zero-shot vs few-shot. Zero-shot describes; few-shot demonstrates. Use zero-shot for tasks the model already knows well and whose labels are unambiguous. Switch to few-shot when transcripts show a repeatable misreading, when the output format must be exact, or when house style matters more than general competence. Fine-tuning is a third step beyond both, justified only for stable, high-volume behaviors.

What to measure.

• Classification agreement: a reviewer labels 100 real messages and compares with the model's zero-shot labels.
• Description drift: how often an intent description has to be amended after launch, which reveals which ones were underspecified.
• Catch-all rate: a high share of messages falling into 'other' means the label set does not match customer language.
• Time from new intent to first correct routing in production.

Botcadence chat agents go live from uploaded documents and instructions, with the website widget typically live in under 5 minutes after upload; see chat agents and document training.