What is a large language model? LLMs explained for fashion professionals
Large language models power chat assistants, product copy tools and AI shopping advisers. A plain explanation of how they work, what they do well in fashion, and where they fail.
KEY TAKEAWAYS Summary by the editors
- A large language model (LLM) is a deep learning model, usually built on the transformer architecture, that is trained on very large amounts of text to predict the next token in a sequence.
- Because LLMs generate statistically plausible text rather than looking up verified facts, they can hallucinate, which means confident answers that are false.
- In The State of Fashion 2026, McKinsey reports that more than 35 percent of fashion executives already use generative AI in areas such as online customer service.
- Zalando announced in October 2024 that its AI-powered fashion assistant, built on its own models and OpenAI's large language models, was available in 25 markets in local languages.
- Under the EU AI Act, obligations for providers of general-purpose AI models have applied since 2 August 2025, so fashion companies should know which model sits behind each tool they use.
A large language model (LLM) is an AI system trained on very large amounts of text so that it can understand and generate language: answering questions, summarising documents, drafting product descriptions or translating copy. It works by predicting the most likely next piece of text, which makes it fluent and flexible but not inherently accurate. For fashion professionals, the practical question is less how clever the model is and more which tasks, data and controls it is given.
What is a large language model?
IBM describes LLMs as deep learning models trained on immense amounts of data that can understand and generate natural language and other content across a wide range of tasks. Most are built on the transformer neural network architecture. The best known products, such as ChatGPT, Gemini or Claude, are chat interfaces on top of such models, but the same models also run quietly inside search boxes, customer service tools, translation services and content systems.
The word "large" refers both to the volume of training text and to the number of internal parameters (weights) the model adjusts during training. A general-purpose LLM is not trained on any one company's data: it knows about fashion in general, from public text, but it knows nothing about your assortment, your prices or your order book unless that information is supplied to it.
How does a large language model actually work?
The process can be explained in four steps, following IBM's overview:
- Tokenisation: text is split into small units called tokens, which may be whole words, parts of words or characters.
- Embeddings and attention: tokens are turned into numerical vectors, and a self-attention mechanism weighs how each token relates to the others, including distant ones, so the model can use context.
- Pretraining: the model learns from billions or trillions of words by repeatedly predicting the next token and adjusting its weights to reduce errors. No manually labelled data is needed for this stage.
- Fine-tuning and inference: the pretrained model is further trained to follow instructions and behave helpfully, then generates answers one token at a time when it receives a prompt.
The key consequence for business users is in step three. An LLM is, in IBM's words, a giant statistical prediction machine. It produces text that is likely given its training and the prompt, which is not the same as text that has been checked against a source of truth.
What can LLMs do for fashion teams today?
LLMs are strongest where the input and output are language and a human can review the result. The table below summarises common fashion tasks and how well suited they are.
| Task | Example | Fit for an LLM | Human review needed |
|---|---|---|---|
| Product copy | Drafting descriptions from attribute data | Good, if attributes are complete | Yes, for accuracy and brand voice |
| Translation and localisation | Adapting copy for several markets | Good | Yes, for sizing terms and legal claims |
| Customer service | Answering order and returns questions | Good with access to policies and order data | Escalation path required |
| Internal knowledge | Summarising supplier manuals or line sheets | Good with retrieval of source documents | Spot checks |
| Styling advice | Conversational outfit suggestions | Moderate, depends on catalogue data | Monitoring of outputs |
| Numbers and forecasts | Predicting sell-through or buy quantities | Weak on its own | Use dedicated forecasting models |
The last row matters. LLMs can explain a forecast or write a summary of sales figures, but they are not forecasting engines. Demand planning, pricing and allocation still rely on statistical and machine learning models trained on a company's own transactional data.
Where are fashion companies already using LLMs?
Adoption is no longer experimental. In The State of Fashion 2026, published in November 2025, McKinsey and The Business of Fashion report that more than 35 percent of executives already use generative AI in areas such as online customer service, with image creation, copywriting, consumer search and product discovery among other uses. The same report notes that shoppers increasingly use LLMs to search for and compare products, making chatbot answers, in its words, the new SEO.
A published example is Zalando. In October 2024 the company announced that its AI-powered assistant, which gives signed-in customers personalised fashion advice in local languages and takes context such as weather and occasion into account, was available in 25 markets. Zalando stated that the assistant is powered by its own models together with OpenAI's large language models. That combination, a general LLM plus company-specific data and models, is the typical pattern for useful deployments.
What are the limits and risks of LLMs?
- Hallucinations: the model can produce false information that sounds plausible, such as an invented fibre composition or a wrong delivery date.
- Bias: models can reflect and amplify biases in their training data, which matters for imagery prompts, sizing language and customer interactions.
- Stale knowledge: a model only knows what was in its training data unless current information is retrieved for it.
- Data leakage: pasting confidential line plans, cost prices or customer data into public tools can expose them.
- Cost and energy: IBM notes that training and running LLMs requires large amounts of computing power and energy.
What does regulation mean for fashion companies using LLMs?
The EU AI Act entered into force on 1 August 2024. Its obligations for providers of general-purpose AI models, the category that covers large language models, have applied since 2 August 2025. Most fashion companies are deployers rather than providers of such models, but they still need to know which model sits behind each tool, what data flows into it and how outputs are labelled and reviewed. Transparency towards customers who are talking to a chatbot, and care with personal data under data protection law, remain the company's own responsibility.
How should a fashion business start with LLMs?
The companies that get value from LLMs usually start narrow. They pick one language-heavy process with a clear owner, such as product copy for one category or answers to frequent wholesale customer questions, and they connect the model to clean, current company data rather than relying on its general knowledge. They define what the model may and may not do, keep a human in the loop for anything customer-facing or financial, and measure quality before and after. Clean product master data, consistent attributes and documented policies are the real prerequisites; the model itself is increasingly a commodity.
Frequently asked questions
What is the difference between an LLM and ChatGPT?
An LLM is the underlying model that predicts and generates text. ChatGPT is a product, a chat interface built on top of OpenAI's large language models. The same distinction applies to other assistants, which can run different models behind the same interface.
Can an LLM write product descriptions for a fashion brand?
Yes, LLMs are well suited to drafting product copy, provided they receive complete and correct attribute data such as material, fit and care. Without that data they may invent details. A human should check accuracy, legal claims and brand voice before publication.
Why do large language models make mistakes?
LLMs generate the statistically most likely next words based on their training and the prompt; they do not verify facts against a database by default. This can lead to hallucinations, confident but false statements. Connecting them to verified company data and requiring sources reduces, but does not remove, the risk.
Are LLMs regulated in the EU?
Yes. Under the EU AI Act, which entered into force on 1 August 2024, obligations for providers of general-purpose AI models apply from 2 August 2025. Companies that use these models in their own tools must also comply with transparency, data protection and any rules that apply to their specific use case.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
