Skip to content
DT Nova

NLP applications in business: what natural language processing does well

NLP applications are software that reads, sorts, understands or writes human language: routing emails, pulling fields out of documents, searching a company's own knowledge, summarising, translating, transcribing speech and answering questions. This guide explains the main NLP applications in business, shows what each needs to work, how to test accuracy, and which EU rules apply.

The short version

  • Natural language processing is the field; large language models are today's most capable tool in it, not the only one.
  • The applications that pay back first handle high volumes of similar text: tickets, emails, invoices, forms, feedback.
  • Accuracy is a number you measure on your own examples before launch, not a promise from a vendor.
  • Where mistakes are costly, the system proposes and a person confirms the uncertain cases.
  • Text usually contains personal data, so the GDPR applies, and the EU AI Act adds transparency duties.

What NLP is

Natural language processing covers any software that works with human language as people write or speak it. For decades that meant hand-written rules, keyword lists and statistical models trained for one task each, such as spam filtering. Since the arrival of large language models, one general model can classify, extract, summarise, translate and draft text from instructions and a few examples. That has made many NLP applications practical for ordinary companies, because a task that once needed thousands of labelled examples can now start with a few dozen and a clear instruction. The older methods remain useful when a task is narrow, must run fast and cheaply at huge volume, or must behave identically every time.

How widely companies use it

Eurostat's survey of enterprises with 10 or more employees, published on 11 December 2025, found that 20.0% of EU enterprises used AI technologies in 2025, up from 13.5% in 2024. The most common use was analysing written language, at 11.8% of enterprises, followed by generating pictures, video or sound (9.5%), generating written or spoken language (8.8%) and converting speech into machine-readable form (7.2%) (Eurostat, 20% of EU enterprises use AI technologies, read 11 November 2026). In other words, three of the four most used AI technologies in EU companies are NLP applications.

NLP applications by task

ApplicationWhat it doesTypical business use
Classification and routingPuts text into categoriesSorting tickets, emails and feedback to the right team
Information extractionPulls named fields out of textInvoices, orders, contracts, forms, CVs
Search and question answeringFinds and quotes the right passageInternal knowledge bases, product documentation
SummarisationShortens long text to its pointsMeeting notes, case histories, reports
TranslationMoves text between languagesProduct content, support replies, documents
Speech to textTranscribes audioMeetings, voice notes, contact centre recordings
Sentiment and topic analysisMeasures tone and themes at scaleReviews, surveys, social mentions
Conversational assistantsAnswers and acts in dialogueCustomer service, internal help desks
DraftingWrites first versionsReplies, descriptions, reports for a person to finish

Classification and routing

The simplest NLP application is often the most valuable. Every incoming email, ticket or form is given a category, a priority and a language, and sent to the right queue. A support team that spends the first minutes of each ticket deciding who should handle it gets those minutes back, and urgent cases stop waiting behind routine ones. Classification also feeds reporting: once every message has a reason code, the company can see which product issues generate the most contact. These systems are easy to measure, because each decision is right or wrong.

Information extraction from documents

Extraction turns documents into data: supplier, amount, due date and order number from an invoice; parties, dates and notice periods from a contract; products and quantities from an order email. Structured e-invoices need no NLP at all, so extraction is for everything else: PDFs, scans, emails and attachments. The result is posted to the accounting, ERP or CRM system through an integration, with fields below a confidence threshold shown to a person. Our document workflow automation page describes how extraction fits with approvals, archiving and signatures, and the RPA use cases guide shows where a software robot finishes the job.

Search and question answering over your own content

Many companies hold the answer to most questions somewhere in manuals, policies, past tickets and product pages, and nobody can find it. Retrieval-based systems index that content, find the passages relevant to a question and have a language model answer from them, with a link to the source. Done well, the answer can be checked in one click, and questions the content does not cover are passed to a person rather than guessed. This is the core of most useful internal assistants and customer service bots; see our chatbot development page. For how such a system is built with permissions and deletion, see our RAG development services.

Summaries, translation and speech

Summarisation saves reading time on long case histories, meeting transcripts and reports, as long as the summary is used as a guide to the source, not a replacement for it in decisions that matter. Machine translation is now good enough for support replies and first drafts of product content in the major European languages, with a native speaker reviewing anything published. Speech to text turns meetings and voice notes into searchable text and is the first step for any analysis of spoken conversations. Recording and transcribing people needs a legal basis and clear notice, and in some countries consent from everyone in the conversation.

Sentiment and feedback analysis

Reading thousands of reviews or survey answers by hand is slow, and star ratings say little about why. NLP can tag each comment with topics, such as delivery, price, quality or staff, and with tone, then show which topics drive negative feedback and how that changes month by month. The output is a ranked list of problems in customers' own words, which is more useful than a single sentiment score.

Testing accuracy before you rely on it

Every NLP application should be measured on your own text. The method is simple: collect a few hundred real examples, have someone who knows the work write the correct answer for each, and run the system against them. For classification the result is the share of correct decisions per category; for extraction, the share of correct fields; for question answering, whether the answer was right, supported by the quoted source, or correctly declined. Then decide the threshold: above it the system acts, below it a person reviews. Keep sampling after launch, because incoming text changes and so does accuracy. The types of measurement used are explained in our guide to the types of analytics, and the full build and test sequence in how to make an artificial intelligence program.

Data protection and security

Emails, tickets, transcripts and documents almost always contain personal data, so the GDPR applies to every NLP application that processes them. In practice: a named purpose and legal basis; a data processing agreement with every AI service the text passes through; EU processing where it is offered; no use of your data to train a provider's models without a decision to allow it; and removal of names and identifiers where the task does not need them. Systems that read external text, such as emails and web pages, must also be protected against prompt injection, where the text contains instructions meant to make the model misbehave; an assistant that can take actions should only be able to take the ones its task needs.

The EU AI Act and NLP

The AI Act's general application date was 2 August 2026. From that date, Article 50(1) requires AI systems that interact directly with people, such as chatbots, to be designed so that people know they are dealing with an AI unless that is obvious. Article 50(2) requires providers of systems that generate text, audio, images or video to mark the output in a machine-readable way; systems already on the market before 2 August 2026 have until 2 December 2026. NLP used for purposes listed in Annex III, such as screening or ranking job applicants or assessing access to essential services, is high-risk, with obligations from 2 December 2027. Most business NLP applications, such as routing tickets or extracting invoice fields, are not high-risk, but the classification should be checked and written down for each system.

Where to start

  1. Pick one high-volume text flow where people do the same reading every day, such as support email or supplier invoices.
  2. Write down what a correct result looks like and collect real examples.
  3. Check whether the data can go to an AI service under your rules, and which one.
  4. Build a small version, measure it on the examples and set the review threshold.
  5. Run it with a person reviewing the uncertain cases, and widen its scope only when the numbers hold.

How DT Nova helps

We map the text flows with the people who handle them, test candidate NLP applications on your own examples and recommend the simplest tool that meets the accuracy needed, sometimes a rule, often a language model with retrieval, always with a person on the uncertain cases. Then we build it into the systems your team already uses, with EU processing as standard, logs of every decision and the code handed over. For the wider picture of where AI fits in a company, see AI consulting.

Questions people ask

What are the most common NLP applications?

Sorting and routing text such as emails and tickets, extracting fields from documents, searching a company's own content, summarising long text, translating, turning speech into text, measuring sentiment in feedback, and chatbots and assistants that answer questions. Eurostat's 2025 survey found that analysing written language was the most common AI use among EU enterprises.

What is the difference between NLP and a large language model?

NLP, natural language processing, is the field: any software that reads, understands or produces human language. Large language models are one family of tools within it, and now the most capable for many tasks. Older NLP methods such as keyword rules and small trained classifiers are still useful where speed, cost, predictability or a very narrow task matter.

How accurate are NLP applications?

It depends on the task and the text, so it has to be measured, not assumed. We build a test set of a few hundred real examples with the correct answers, measure the system against it before launch, and keep measuring on samples afterwards. For tasks where a mistake is costly, low-confidence results go to a person.

Does NLP work in Finnish and other smaller languages?

Current large language models handle Finnish, Swedish and the major European languages well for classification, extraction, summarising and drafting. Accuracy is lower for dialects, heavy jargon and very short text, so every language is tested with real examples before launch.

Which EU rules apply to NLP applications?

The GDPR whenever the text contains personal data, which emails, tickets and transcripts usually do. The EU AI Act adds transparency duties: Article 50(1) requires people to be told when they interact with an AI system unless it is obvious, applying since 2 August 2026, and Article 50(2) requires machine-readable marking of AI-generated text and media, with a transition to 2 December 2026 for systems already on the market. NLP used for high-risk purposes listed in Annex III, such as screening job applicants, has further duties from 2 December 2027.

Sources