Natural language generation (NLG) is the branch of artificial intelligence that produces human language, written or spoken, from data or from a prompt. Classic NLG systems turn structured data, such as sales figures or sensor readings, into sentences using rules and templates. Large language models (LLMs) are now widely used for free-form NLG, such as summaries, answers and draft replies, while rules and templates remain common where constrained, repeatable output is required. NLG is the output side of natural language processing (NLP): understanding language goes in, generating it comes out.
At a glance
- NLG turns data or prompts into readable text or speech: reports, summaries, answers, replies and descriptions.
- Two broad approaches exist: rules and templates, which are predictable and limited, and language models, which are flexible and can be wrong.
- It shows up in BI tools, reporting, contact centers, chatbots, virtual agents and writing assistants.
- Output quality depends on the source data, the method and the review process; fluent text is not the same as accurate text.
What problem it solves
Organizations produce far more data than people have time to read. A dashboard full of charts still needs someone to explain what changed and why it matters, and many routine documents, such as weekly performance updates, portfolio reports or product descriptions, follow the same pattern every time. NLG turns that data into plain-language narrative automatically and at scale, so people who don’t read charts can still act on the numbers and staff spend less time writing repetitive text.
In customer-facing work, NLG writes the responses of chatbots and virtual agents, drafts replies for human agents and summarizes calls after they end, reducing after-call work.
How it works
Template and rules-based NLG. The classic approach follows a pipeline. The system decides which facts to include (for example, revenue, margin and the biggest regional change), orders them, groups related facts into sentences, chooses wording (“rose” versus “jumped”) and applies grammar to produce the final text. Everything it says comes from the data and the rules, which makes it accurate as long as the data is right, but it can sound repetitive and can only say what its designers anticipated.
Language-model-based NLG. Generative AI systems built on LLMs predict text word by word based on patterns learned from large amounts of training data and the prompt they’re given. They can summarize, explain, rephrase and answer in natural, varied language and handle requests nobody scripted. They can also produce plausible statements that the data doesn’t support (hallucinations), and the same input can produce different outputs.
Hybrid approaches. Many business products combine the two. They calculate figures with conventional code, retrieve relevant documents or records, a technique called retrieval-augmented generation (RAG), and then use a language model to write the narrative around them, sometimes with checks that numbers in the text match the source.
Delivery. Generated text can appear in a dashboard, a report, an email, a chat window or an agent’s screen, or be converted to speech for voice assistants and phone systems.
When it matters for buyers
- When evaluating BI and reporting tools. Many analytics platforms now offer automatic narrative summaries; test them on your own data.
- When automating contact center work. Call summaries, suggested replies and agent assist features rely on NLG; measure accuracy and edit rates, not just time saved.
- When output reaches customers or regulators. Decide where human review is required and how errors are caught before publication.
- When data is sensitive. Ask where prompts and data go, whether they are used for training and how long they are kept.
Our artificial intelligence overview covers how to evaluate AI capabilities like these across vendors.
Questions to ask vendors
- Does the product use templates, a language model or both, and for which features?
- How do you make sure numbers and facts in generated text match our source data?
- Can we see the data or documents behind each generated statement?
- What review and approval steps can we add before text is sent or published?
- Which model do you use, where is our data processed, and is it used to train models?
- How do you measure and report accuracy, and what happens when output is wrong?
- What languages, tones and formats are supported?
How it differs from NLP
Natural language processing (NLP) is the broad field of AI that works with human language, and in product descriptions it often means the understanding side: interpreting text or speech to find meaning, intent, sentiment or key details. NLG is the generation side: producing language. Many products combine both. A virtual agent uses NLP to understand what a customer asks and NLG to answer, and an LLM does both within one model. When a vendor says a product “uses NLP”, ask whether it interprets language, generates it or both, because the risks differ: misunderstanding versus saying something wrong.