Natural language understanding (NLU) is the part of natural language processing that works out what a person means: their intent, key details, and context. In customer service, NLU turns a message such as "my parcel never arrived" into a request a system can act on.
On a phone call, speech recognition first turns the audio into text, and NLU then works on that text. In chat and email, it reads the customer's words directly.
How NLU works
Intent recognition. The system decides what the customer wants, such as tracking an order or blocking a card (see intent recognition).
Entity extraction. It picks out the details the task needs, such as an order number, a date, or an amount. These details are also called slots.
Context. It keeps track of earlier turns, so "change it to Friday" refers to the delivery mentioned a minute ago.
Sentiment and urgency. Some systems also score frustration or urgency, which can change routing or priority.
On voice, errors carry forward. A misheard digit becomes a wrong order number, so NLU accuracy on calls also depends on word error rate.
NLU, NLP, and NLG
Natural language processing (NLP) is the wider field of computing with human language. NLU is the part that interprets input. Natural language generation (NLG) is the part that produces text, such as a reply. A customer service agent uses both, NLU to read the request and NLG to answer it.
Research has long tested understanding across many tasks at once. GLUE, a benchmark published in 2018, scored models on a range of existing NLU tasks, on the view that useful understanding has to be general.
Classic NLU and large language models
Classic NLU. A team defines a list of intents and trains a classifier with example phrases for each. The system returns an intent with a confidence score, and low-confidence messages go to a fallback or a person. It is predictable, and every new intent needs new training data.
LLM-based understanding. A large language model can interpret requests without training for each intent, including messages that mix two requests or change topic halfway. It also copes with phrasing the team did not anticipate. Its output can vary between runs, so it needs testing and limits.
Many current systems combine the two. A language model interprets the message, and fixed rules or code decide what happens next, especially for payments, identity checks, and disclosures.
NLU in the contact center
Telecom. One sentence can carry two requests, such as a billing question and a fault report. Recognizing both can save the customer a second contact.
Banking. "I lost my card and there's a payment I don't recognize" combines urgency, a card block, and a possible dispute. The card block usually comes first, then the dispute.
Ecommerce. Return, refund, and exchange requests are phrased in many ways, often in several languages. Good NLU maps each phrasing to the same process in every language.
Frequently asked questions
What is the difference between NLU and NLP?
NLP is the whole field of processing human language by computer, including translation, summarization, and text generation. NLU is the part of NLP that interprets meaning: intent, entities, and context.
Is NLU still needed with large language models?
Yes, though the method has changed. Large language models now do much of the work that intent classifiers used to do. Many teams keep named intents for reporting, routing, and regulated flows, because a named intent is easy to count and test.
What is an intent in NLU?
An intent is the goal behind a message, such as "track order" or "report a lost card." The NLU system maps many phrasings of the same goal to one intent, which then starts a process.
How do you measure NLU accuracy?
A common method uses a labeled test set of real messages the system has not seen. Teams measure the share classified correctly, precision and recall per intent, and which intents get confused with each other. Detection of out-of-scope messages is worth measuring too, since a confident wrong answer often costs more than a handover.
How we approach this at Zowie
Zowie builds AI agents for voice, chat, and email. In each conversation, the language model speaks. A separate, deterministic Decision Engine returns every amount, rate, offer, disclosure, and escalation from your rules, with an audit log. You are welcome to book a demo.
Explore: Natural language processing, Intent recognition, Large language model, Conversational AI, Word error rate, AI accuracy