What is a large language model (LLM)?

A large language model (LLM) is an AI model trained on very large volumes of text that can understand, summarize, and generate language. In companies, LLMs can make information from reports and data available more quickly, for example. For accurate answers, an LLM needs a reliable data foundation. Bissantz connects LLMs with data from the BI system so that controllers can use natural-language questions to drill down from a variance to its underlying cause.

LLM stands for Large Language Model and roughly translates into German as “large language model.” It describes a neural network that statistically models language and can therefore understand and generate text.

“Large” refers to two dimensions: the amount of training data and the number of parameters, which for current models is in the billions.

LLMs are a form of generative AI and are based on deep learning. They are the technology behind well-known AI assistants and many AI features in business software.

An LLM predicts step by step which piece of text is most likely to come next. This simple task gives rise to capabilities such as summarization, translation, and question answering when combined with enormous amounts of data.

Technically, almost all current LLMs are based on the transformer architecture. At their core is the attention mechanism: For each word, the model determines which other words in the text are important for its meaning.

Important for business use: An LLM does not store a fact database; it stores probabilities. As a result, it can generate convincing but incorrect statements. These so-called hallucinations can only be mitigated through verified data sources and human oversight.

LLMs differ in terms of input type, size, and deployment. For companies, it is particularly relevant whether a model can process images and whether it can be operated in their own data center.

A multimodal large language model (MLLM), for example, processes multiple types of data, such as text, images, audio, or video. It can describe a chart, extract information from a scanned invoice, or answer a question about a screenshot, for example.

The best-known large language models include the GPT, Claude, Gemini, Llama, and Mistral model families. Providers release new versions at short intervals, so the table below lists only the model families.

For selecting an LLM for business use, the model name matters less than where the data is processed, which contracts apply, and how well the model fits into existing systems.

Many companies use LLMs to make knowledge from data and documents available more quickly. In controlling and business intelligence, LLMs primarily shorten the path from a number to an explanation.

Typical tasks in controlling include:

• Commenting on reports: The LLM turns identified variances into an understandable management comment.
• Asking questions in natural language: Business users can ask, “Why is revenue in the South below plan?” instead of setting filters.
• Unlocking unstructured data: The LLM analyzes meeting transcripts, support tickets, or production reports and uses them to supplement the KPIs.
• Deriving recommendations for action: Based on the analysis, the LLM suggests possible actions that a human reviews and decides on.

An LLM does not replace a data warehouse or BI software. It sits above them as a language layer and is only as good as the data quality underneath.

Large Language Model Optimization refers to two different things: the technical adaptation of a model to a specific task and, in marketing, the visibility of content in AI-generated answers.

In the technical sense, optimization initially refers to a mathematical process. During training, a loss function is minimized until the model produces predictions that are as accurate as possible. For users, the more relevant question is how to adapt a finished model for their own use:

The right solution depends on which data the LLM is supposed to process and how reliable the result needs to be. For KPIs, the rule is: The higher the accuracy requirements, the more closely the LLM needs to be tied to a verified data model.

The most common mistakes occur when companies assign an LLM tasks it was not designed for or when the underlying data is missing.

Bissantz uses LLMs in DeltaMaster and DeltaApp as a language layer above the BI data model.

The roles are clearly defined. DeltaMaster provides the KPIs from the verified data model and automatically identifies variances and anomalies. The LLM translates the results into executive summaries, answers follow-up questions via chat or voice control, and guides users to the underlying cause through drill-downs.

Through AI-powered data integration, the analysis also incorporates unstructured sources, such as sales meeting transcripts or production reports. Bissantz controls access through MCP servers. Companies determine themselves which systems are connected and which information is included in the AI analyses.

A manufacturing company identifies the following in its monthly report: The contribution margin in the South region is 8% below plan. Previously, the controller would have set filters, reconfigured reports, and written a management summary manually.

With DeltaMaster, the sales manager asks directly: “Why is the contribution margin in the South region below plan?” The AI analysis performs a drill-down across product groups and customers. It identifies the cause: Two major customers received higher discounts, while sales volume is on plan.

The LLM adds context from the connected sales meeting transcripts. One of them documents a competitor’s offer that explains the discounts. The AI turns this information into an assessment and suggests measures, such as reviewing the discount structure. After review by the controller, the DeltaMaster Publisher sends the summary to the CFO.

The key is the division of labor: The numbers come from the BI data model, while the LLM explains them. This ensures that every statement can be traced back to a KPI.

A large language model is a language model that understands and generates text by calculating the most likely next word. In controlling, an LLM delivers value when it works on a verified data foundation and explains numbers rather than calculating them itself. Bissantz combines both in DeltaMaster: KPIs come from the BI model, while explanations and recommendations are generated by the LLM.

artificial intelligence (AI) generative AI machine learning decision intelligence agentic AI business intelligence MCP server neural networks human-in-the-loop