How generative AI works
Modern generative systems are based on so-called foundation models: very large neural networks, usually with a transformer architecture, pre-trained on broad datasets. A large language model (LLM) learns to predict the most likely next word fragment (token). With sufficient size, this seemingly simple task gives rise to the ability to write coherent text, answer questions, translate and write code.
The model is steered via the prompt: the input that describes the task, the context and the desired format. For business use, the base model is typically extended: retrieval-augmented generation (RAG) supplies it with up-to-date company knowledge, fine-tuning adapts its style or behaviour, and tool connections turn a pure text generator into an acting AI agent.
The most important modalities
- Text: answers, summaries, translations, marketing copy, structured analyses. This is the broadest field of use in everyday business.
- Image: illustrations, product visualisations, design drafts from text descriptions.
- Audio: speech synthesis for voicebots, transcription, translation of spoken language.
- Video: short generated clips and avatars. Technically the youngest modality, with rapid progress.
- Code: assisted programming, from autocompletion to independently implementing clearly defined tasks.
What generative AI does for SMEs
The benefit arises wherever content is created or processed every day. In practice, four fields contribute the most:
Content and marketing. Product descriptions, category texts, newsletter drafts and translations are produced in a fraction of the previous time. Human final review remains essential: generative AI delivers the rough draft, not the approved end product.
Customer service. Chatbots and voicebots answer recurring enquiries, summarise cases for staff and draft suggested replies. Combined with RAG, they draw on the company’s actual knowledge base instead of guessing.
Internal document work. Summarising contracts, structuring minutes, drafting quotes, prioritising emails: these unspectacular routine tasks add up to the biggest time savings.
Software development. Coding assistants speed up routine work, explain existing code and write tests. They do not replace engineering judgement, but they shift a significant share of time towards design and review.
An example from practice
A typical starter project in B2B trade: a retailer with several tens of thousands of items has to maintain product texts in several languages. A generative set-up creates descriptions from the structured product data (attributes, dimensions, target group) according to defined style guidelines, translates them and submits them for approval. The editorial team reviews and corrects instead of writing every text from scratch. The effort per item drops noticeably, the tone stays consistent, and because the texts are generated from the company’s own product data, their factual accuracy remains controllable. The same pattern of structured source data, generated draft and human approval works just as well for quotes, job adverts or support replies.
Limits and risks
Generative AI has systematic weaknesses that every deployment concept has to plan for.
Hallucinations. Language models produce statements that sound plausible but are wrong, especially for facts, figures and source references. Countermeasures are RAG (answers from verified sources), narrowly defined tasks and human review wherever mistakes are expensive.
Data protection. As soon as personal data flows into prompts or training data, the GDPR obligations apply: a data processing agreement with the provider, excluding your own inputs from model training, and clear internal guidelines. The business versions of the major providers include this as standard; free private accounts do not.
Copyright and labelling. The legal situation regarding AI-generated content is still evolving. Article 50 of the EU AI Act lays down transparency obligations, such as labelling interactions with AI and certain generated content; the general transparency obligations apply as scheduled from 2 August 2026 (as of July 2026). Anyone using generative AI in production should consider labelling and documentation questions early on.
Quality and skills. Results are only as good as the task definition and the review. That is why prompt engineering and building AI literacy in the team belong in every serious rollout.
Generative AI is also changing online visibility
Generative AI is not just a tool but also a new sales channel, or a competitor for attention. When ChatGPT, Perplexity and Google’s AI Overviews answer questions directly, visibility increasingly depends on whether your own content is cited in these generated answers. For companies, this means that classic search engine optimisation is joined by a second discipline, generative engine optimisation (GEO): structuring and backing up content so that language models recognise it as a source worth citing. Anyone introducing generative AI in their company should keep this external perspective in mind: the same technology that produces texts internally determines externally which providers still appear in AI answers at all.
In short, it pays to look at generative AI in both directions: as a productivity tool in your own processes and as a shift in the rules of marketing and sales. Both can be shaped, and both penalise waiting in the same way: the lead of those who learn systematically early on grows with every quarter.
Choosing and running a model: API, EU cloud or self-hosting
For business use, the question of where the model runs comes up early. Three operating models have become established, each with its own profile.
Provider API. Accessing models from OpenAI, Anthropic or Google via their interfaces is the fastest route to high quality: no infrastructure effort, immediate access to current model generations, usage-based billing. A prerequisite for serious use is the provider’s business terms: a data processing agreement, no training on your own inputs and, ideally, an EU data region.
Cloud platforms. The same or comparable models can be run via hyperscaler platforms in defined regions. This makes it easier to integrate them into existing cloud contracts, permission and audit structures, and it is often the pragmatic middle way for regulated industries.
Self-hosting open models. Open-weights models such as Llama or Mistral run on your own or rented hardware. Maximum data control and predictable costs under high load are offset by noticeable operating effort and, depending on model size, quality compromises compared with the leading models. This makes sense when data protection requirements or volumes justify it, and is rarely the first step.
In practice, the use case decides: non-critical text tasks can take the fastest route, while applications using sensitive data require the most controlled option. Many companies therefore run both in parallel.
How to get started: from pilot project to rollout
The companies in which generative AI has a measurable effect rarely start with a large-scale project. The recurring pattern of success has four steps:
- Prioritise use cases. Look for tasks that take up a lot of time, have a clear quality standard and little potential for damage: text drafts, summaries, internal search. Not fully automated customer communication as the first attempt.
- Pilot with metrics. One team, one process, four to eight weeks, with key figures defined in advance (processing time, correction effort, satisfaction). Without measurement, you can later neither prove success nor justify scaling.
- Guidelines and training. By now at the latest, you need rules: which data may go into which tools, how results are reviewed, who remains responsible. This is also the core of the AI literacy that the EU AI Act calls for.
- Scale what works. Successful pilots are transferred into regular operation, with clear responsibilities, cost control and regular reviews of the choice of model and provider. Failed pilots are ended with documentation; that is a result too.
Why projects fail: tool roll-outs without a link to processes (“we have licences now”), a lack of final review that destroys trust as soon as the first mistake becomes public, and the expectation that the model will replace professional responsibility. Generative AI amplifies good processes and exposes bad ones.
Foundation model: the broadly pre-trained base model on which specific applications are built; one model, many uses. Hallucination: fluently worded but factually incorrect model output; it does not disappear with better models, only with better safeguards. Token: the basic unit in which models process text and providers bill; roughly a part of a word. Context window: the amount of text the model can take into account at the same time; it limits how much material a single run can process. Multimodality: a model’s ability to understand or generate several types of media, for example interpreting a photo and writing text about it. And the most important distinction for projects: RAG brings up-to-date knowledge into the answer at runtime, while fine-tuning changes the behaviour of the model itself. Anyone trying to solve factual questions with fine-tuning almost always chooses the wrong tool.
Frequently asked questions about generative AI
Is generative AI the same as a large language model?
No. An LLM is currently the most important form of generative AI for text and code. Generative AI also covers image, audio and video models. Every LLM is generative AI, but not all generative AI is an LLM.
What data does generative AI need in a company?
Usually less than expected to get started: a clearly defined use case and access to the relevant documents or product data are enough for first pilots. Your own training or fine-tuning only pays off once standard models plus RAG have demonstrably proven insufficient.
Does generative AI replace employees?
In practice, it shifts job profiles: routine parts (first draft, summary, research) shrink, while reviewing, steering and subject-matter parts grow. The companies with the best results treat generative AI as a tool within the process, not as a substitute for responsibility.
Further definition: Generative AI on Wikipedia.
