What is natural language understanding?
Natural language understanding (NLU) is how a computer works out what someone means. People rarely say exactly what they want in a tidy, predictable way. A customer might say “the power’s gone off again on our street”, “is there an outage in Henderson?” or “my lights just died”. NLU is the technology that recognises all three as the same underlying need and picks out the useful details.
It is a core building block of conversational AI. In voice channels, automatic speech recognition turns speech into text first, and NLU then interprets that text. In text channels such as chatbots and email, NLU works on the customer’s words directly.
What NLU extracts
NLU usually produces three kinds of output:
- Intent. The purpose of the message: report a fault, ask for an update, change a payment date. This is covered in more detail under intent recognition.
- Entities. The specific details: an address, account number, ICP number, date, amount or vehicle registration.
- Context and tone. Signals such as urgency, negation (“it’s not the meter, it’s the whole house”) and emotion. Emotion is often handled by sentiment analysis.
How NLU has changed
Early NLU systems were trained on a fixed list of intents with many example phrases for each. They worked well within their list but failed on anything new, and adding a new intent meant gathering fresh training data.
Modern NLU is often built on large language models, a form of generative AI. These models have learned from enormous amounts of text, so they cope much better with unusual phrasing, long rambling explanations and several requests in one message. They can also be instructed in plain language, for example “capture the customer’s address and the reason for contact”, rather than trained from scratch. The trade-off is that they need careful guardrails and testing, because their behaviour is less predictable than a fixed list.
Where NLU is used in a contact centre
- Self-service. Understanding callers and chat users well enough to resolve their request or route them correctly.
- Email triage. Reading incoming messages, extracting details into fields and routing by content.
- Agent assist. Following a live conversation and surfacing the right knowledge article.
- Analytics. Categorising every interaction by reason for contact in speech analytics.
- Summaries. Identifying what mattered in a conversation for AI call summarisation.
Getting NLU right
NLU accuracy has a direct effect on customer experience. Misunderstanding an intent sends a customer down the wrong path, and misreading an address can send a crew to the wrong street. Practical steps include:
- Test on real conversations, not just examples written by the project team. Real customers are less tidy.
- Include local language. NZ and Australian slang, te reo Māori place names and your own product names all need checking.
- Use confidence thresholds. When the system isn’t sure, it should confirm or clarify rather than guess.
- Confirm critical details back to the customer before acting on them.
- Review misunderstandings regularly and adjust the configuration as patterns emerge.