NLU (Natural Language Understanding)

Natural language understanding is the part of an agent that converts a customer's words into something a workflow can act on. 'Can you stop my shipment' and 'I want to cancel the order I placed yesterday' are different strings with the same meaning; NLU is what makes the agent treat them the same way. In a voice or chat agent it runs on every incoming turn, after speech-to-text on calls, and before any reasoning, tool call or reply.

How it works. The input is the latest message plus the conversation so far. The output is structured: an intent label, a set of entities with normalized values (a spoken 'next Tuesday' becomes a date, 'two thirty' becomes 14:30), resolved references ('the first one' mapped to a specific option offered earlier), and often a confidence estimate. Earlier NLU engines did this with classifiers trained on dozens of sample utterances per intent; language models do it from a description of the intents and entities in the prompt, which removes most of the setup work. Where precision matters, such as phone numbers and amounts, the agent reads the value back for confirmation rather than trusting the parse.

Why it matters for containment and data quality. Every field an agent writes into your CRM or calendar passes through NLU first. Weak understanding shows up as wrong dates on bookings, duplicate contacts from misparsed names, and loops where the agent asks for information the customer already gave. Strong understanding shows up as short conversations: the customer says what they want once and the agent fills every slot it can from that message, asking only for what is missing.

Example. A clinic's chat agent receives: 'Need to book my daughter Maya for a flu shot, she's 7, any day after 3pm this week except thursday.' NLU returns intent book, patient name Maya, relationship child, age 7, service flu shot, time constraint after 15:00, date range this week, exclusion Thursday. The agent calls the scheduling tool with those constraints, gets two slots, and asks which one, without re-asking anything the parent already typed. On a call, the same message would also pass through speech-to-text, and the agent would confirm the spelling of Maya before creating the record.

NLU vs NLG. NLU reads; NLG writes. NLU takes unstructured text and produces structured meaning; NLG takes structured data and a decision and produces fluent text. In a language-model agent the same model does both, but they fail differently: a failure in understanding misroutes or misparses, a failure in generation answers correctly in the wrong tone or at the wrong length. Diagnose which one you are looking at before changing the prompt.

What to measure.

Botcadence agents extract booking and lead details across chat channels and voice and pass them to connected systems such as HubSpot and Google Calendar; see chat agents and workflow automation.