What is an LLM and why does it matter?

Everyone is talking about ChatGPT, Claude, and Gemini. But what is actually under the hood? A Large Language Model (LLM) is not magic and not science fiction. It is a statistical model that has become very good at predicting the next word, and from that simple task an entire generation of new software has emerged. This article explains what an LLM is, how it works, what it can and cannot do, and where it concretely makes a difference for SMBs in the Netherlands.

What is an LLM in one paragraph

A Large Language Model is a neural network trained on enormous amounts of text. "Large" refers to the number of parameters (the adjustable knobs in the model, often hundreds of billions) and to the volume of text it was trained on. The model learns patterns in language: which words follow which words, which sentences fit which contexts, which answers go with which questions. The result is a system that can write text, summarize, translate, generate code, and reason step by step, all by repeatedly predicting the most likely next piece of text.

If you prefer to start with general AI fundamentals: read what is AI: an explanation for SMBs first.

How an LLM works in 60 seconds

Three concepts capture it: tokens, training, and prediction.

Tokens. An LLM does not see letters or words, but tokens. A token is a piece of text, often 3 to 4 characters or half a syllable. The word "datadream", for example, is split into two or three tokens. Modern models are trained on hundreds of billions to trillions of tokens. Llama 3 was trained on 15 trillion tokens, GPT-4 on an estimated much higher amount.

Training. During training, the model sees billions of examples and has to predict the next token each time. Given "The capital of the Netherlands is", the model has to learn that "Amsterdam" is a more likely next token than "Berlin". This happens trillions of times, and parameters get adjusted until the model becomes good at predicting. Then comes a second phase: fine-tuning with human feedback (RLHF), where humans teach the model what counts as a useful or unsafe response.

Prediction. When you ask a question, the model predicts the most likely answer token by token. It does not think, it does not know anything in the real sense. It has become very good at pattern-matching what a good answer should look like. That is both its strength and its biggest limitation.

Why 2022 was the tipping point

LLMs had existed for a while, but 2022 changed everything. Three things came together.

In 2017, a Google team published the paper Attention is All You Need. It introduced the transformer architecture, the computational model that lets you process text in parallel rather than word by word. That made training at massive scale possible.

In 2020, OpenAI launched GPT-3, with 175 billion parameters. Big enough to do astonishing things, but not yet accessible to the general public.

In November 2022, OpenAI launched ChatGPT, based on GPT-3.5. Within five days it had a million users, within two months a hundred million. For the first time, anyone, not just researchers, could talk to an LLM in plain language. That was the tipping point: AI moved from the lab to the laptop.

Since then the pace has been wild. Anthropic, Google, Meta, and Mistral have all released powerful models, context windows have grown from 4,000 tokens to 1 million plus, and cost per million tokens has dropped by more than 90%.

What an LLM is good at

An LLM is strong in everything related to language and pattern recognition.

Writing and rewriting. Blog posts, emails, product descriptions, summaries, social posts. With a good prompt and your brand voice as input, an LLM produces a first draft in seconds. Not always finished, but a solid start. More on content at scale at /en/ai-content.

Summarizing and structuring. Give an LLM a 50-page report and ask for the three key conclusions. Or have rough notes turned into a clean document. An LLM beats a human comfortably here, including on time.

Translation. LLMs often translate better than Google Translate, especially for longer pieces where context matters. Anthropic published a comparison showing that Claude 3 outperformed earlier models on multilingual benchmarks.

Code writing and debugging. According to GitHub's Developer Survey 2023, 92% of professional developers were already using AI tools then, and the Stack Overflow Developer Survey 2025 confirms it has settled at around 84%. LLMs not only write code, they explain existing code, find bugs, and suggest refactors.

Step-by-step reasoning. Modern models like Claude Opus 4.7 and GPT-5 can solve complex problems by writing out intermediate steps explicitly (chain-of-thought). That works well for analyses, planning, and math.

Where an LLM can go wrong

Equally important: know what an LLM cannot do, because that is where most implementations fail.

Hallucinations. An LLM predicts the most likely answer, not the most correct one. If it does not know something, it will confidently produce a fabricated fact, quote, or citation. Research from OpenAI shows that even the latest models hallucinate between 1% and 30% of the time, depending on the question type. For publications or legal work: always fact-check.

Knowledge cutoff. A model is trained up to a certain date (the "knowledge cutoff"). Ask it about this week and it knows nothing, unless it has access to live web search. Claude, ChatGPT, and Gemini all have web grounding built in now, but you have to enable it explicitly.

No real understanding. An LLM does not "get" what it writes, it predicts patterns. With math, logic, or complex planning, it sometimes makes mistakes a 10-year-old would catch. For these tasks, always combine an LLM with external tools (a calculator, a calendar, a database).

Cost. API costs have dropped a lot, but at large volume they add up. An LLM call costs between $0.50 and $30 per million tokens, depending on model and provider. For a chatbot handling 100,000 conversations per month, that quickly becomes a few thousand euros monthly. Good architecture (caching, smaller models for simple tasks, RAG) can cut costs by 70%.

Privacy. If you put sensitive company data into an LLM, that is something to think about. More on this below.

The major LLMs in 2026

By 2026 the landscape is fairly settled. Four major players and a growing open source movement.

Claude (Anthropic). The top model is now Claude Opus 4.7, with a 1 million token context window. Strong at long documents, code, research, and agentic tasks. Pricing: about $15 per million input tokens, $75 per million output tokens. Claude has a strong focus on constitutional AI and safety. My personal preference for serious work.

GPT (OpenAI). GPT-5 is widely available, GPT-4o remains popular for speed and multimodality (image, voice, video). Pricing sits around $5 to $20 per million tokens, depending on model. ChatGPT itself is the first AI contact point for many people. See the OpenAI documentation.

Gemini (Google). Gemini 2.5 Pro with native multimodality and very large context windows (1M+ tokens). Strong at research thanks to direct Google Search grounding. Pricing is competitive, often slightly cheaper than OpenAI or Anthropic. Google's Gemini documentation.

Mistral (French/European). Mistral Large is the European alternative, strong on multilingual work and often interesting for companies that want European data residency. Priced lower than the US top tier, with quality that holds up for most business tasks.

Llama (Meta). Open source. Llama 4 is free to download and self-host, or available through providers like Groq and Together AI. For companies wanting full control over their infrastructure, or with strict privacy requirements, Llama is the default choice. Llama models are also commercially usable under the Meta license.

For most SMB use cases, picking between Claude and GPT works fine. If you want European hosting or self-hosted: Mistral or Llama.

How LLMs make a difference for your business

So much for the technology. What does it concretely mean for an SMB? Five applications where you see results within weeks.

1. Customer service that never sleeps. A chatbot on top of your knowledge base and FAQs answers 60 to 80% of standard questions automatically. The rest gets handed to a human, with the conversation history already filled in. Result: faster response times, lower cost per ticket, employees focused on real problems. Important: this only works with good RAG (see below), otherwise the bot will hallucinate about your products.

2. Content at scale without flat output. With a solid prompt library and your brand voice as input, you generate product copy, blog posts, and emails 5x faster than by hand. An editor reviews and publishes. For a webshop with 200 SKUs, that saves 3 to 5 days of work per month.

3. Research and analysis from hours to minutes. Market research, competitive analyses, summaries of reports or contracts: an LLM does in 10 minutes what a junior consultant takes a day to do. Combine that with live web search and you have a research assistant that includes new information too.

4. Personal communication at scale. Emails, proposals, and presentations actually tailored to the recipient. Not generic mail merge, but content that genuinely fits each customer or segment. Open rates and response go up by tens of percent when this is done well.

5. Agents that take actions, not just answer. The difference between a chatbot and a real AI agent is that the second can use tools: send emails, fetch data, fill forms, schedule appointments. Read more in AI agent vs chatbot: the difference and at /en/ai-agents.

McKinsey estimates that generative AI adds $2.6 to $4.4 trillion in annual value globally. For SMBs the upside is less spectacular but more concrete: 20 to 40% productivity gains on specific processes, when implemented well.

Open source vs closed source: which to pick

A question that comes up regularly: should I run my AI on a commercial API, or self-host with open source?

Closed source (Claude, GPT, Gemini) is the default for most companies. You pay per use, you get the best model of the moment, and you do not manage infrastructure. Downsides: your data goes to the provider (though you can lock this down with enterprise contracts), and you depend on their pricing and availability.

Open source (Llama, Mistral, Qwen) is interesting if you have strict privacy requirements, your own GPUs, or if you want to fine-tune a specific model on your data. Downside: you pay much more upfront (infrastructure, expertise) and the models tend to lag half a generation behind the top closed source models.

For 90% of SMBs: start with closed source via API. Open source becomes interesting once your volume grows or compliance demands it.

Privacy and data: what you need to know

Two things to remember before putting sensitive data into an LLM.

1. Data processing agreement and data residency. If you use a regular consumer chatbot (free ChatGPT, Claude.ai), you have no data processing agreement and your data may be used for training. For business use, always a paid team or API account, with DPA and the right opt-outs. Anthropic, OpenAI, and Google all offer this. For strict EU data residency: Mistral, or an EU region with Azure OpenAI.

2. AI Act and GDPR. The EU AI Act has been in force since 2024 and rolls out in phases. For most SMB use cases (chatbots, content generation, productivity), you fall in the "limited risk" category: you must make clear that a user is talking to AI. For applications around hiring, credit, or biometric data, much heavier requirements apply. GDPR still applies in full: do not just throw personal data into an LLM without a legal basis and proper security.

The pattern that became standard in 2026: Retrieval-Augmented Generation (RAG). Instead of letting the model know everything, you give it the relevant pieces of your own documents at the moment of the question. Benefits: up-to-date knowledge, fewer hallucinations, control over what the model "knows". For business chatbots and internal knowledge tools, RAG is now the norm.

Conclusion: LLMs are a new category of software

An LLM is not Google that happens to talk, and not a human that happens to live in a computer. It is a new category of software: a statistical language model that recognizes and continues patterns. That is less romantic than the marketing sometimes suggests, and at the same time genuinely revolutionary, because so much work in a company consists of processing language and information.

The difference an LLM makes for your company depends on two things: pick the right applications where the technology actually adds value, and have the discipline to integrate it neatly with your data, your brand voice, and your processes. Without that discipline, it is an expensive toy stack. With that discipline, it is the biggest productivity step in 20 years.

Want to know where AI can already make a difference in your business? Take the free AI scan or schedule a conversation about a concrete AI strategy. No consultancy circus, just an honest inventory of where the first 20% gain sits.