The terms vendors use loosely, defined precisely.

Half of AI sales works because the buyer and the seller mean different things by the same word. This glossary fixes definitions for the terms that actually change a purchase decision, and says where the industry uses them inconsistently so you can ask a sharper question.

The terms that most often mislead

• Agent — vendors use it for anything from a decision tree to a tool-using model.
• RAG — retrieval quality, not model size, is what decides answer accuracy.
• Barge-in — whether a caller can interrupt without the agent talking over them.
• Deflection rate — meaningless unless you know what counted as deflected.
• Eval harness — the regression suite that gates a deploy, not a one-off test.

Latency, and which number to ask for

Vendors quote model latency. What a caller experiences is time from their silence to first audio out, which includes speech recognition, reasoning, and the start of synthesis. Ask for that number, measured at the ninety-fifth percentile, on live traffic. The difference between the two figures is often more than a second.

Why the definitions matter commercially

A per-minute price is only comparable if both vendors include the reasoning layer. A deflection rate is only comparable if both count escalated conversations the same way. Nailing the vocabulary before the demo is the cheapest leverage a buyer has in this category.