One business, five very different problems
Picture a company on a busy Monday morning.
A customer sends a message asking why an order is late and whether it can still arrive before the end of the week. A supplier sends a six-page invoice as a PDF, and the information inside it needs to enter the business process without somebody typing it all in manually. Finance is looking at thousands of open invoices and wants to know which customers are most likely to pay late. An employee asks whether unused vacation days can be carried over into next year, and the answer depends on the company’s own HR policy. Meanwhile, another team is working on a highly specialized business problem for which none of the existing models quite fits.
These situations all involve AI, but the problems are fundamentally different. One is about understanding and generating language, another begins with a business document. Another asks for a prediction from structured data, the other depends on knowledge that belongs specifically to the company, while the last may require something built for a very particular purpose.
Trying to solve all five with exactly the same AI would be a little like asking one tool to do every job in a workshop. The more useful question is: What kind of AI does this problem actually need?
When the problem begins with language
Let’s start with the customer: "Why is my order late, and can I still receive it this week?"
Understanding a request like this, summarizing information, generating an explanation, or maintaining a natural-language interaction are tasks for which large language models can be useful. But even here, there does not have to be a single model behind every experience.
Through the generative AI hub, developers can access a range of SAP and third-party models and explore which model is best suited to a particular use case. The architecture is designed so that organizations can move between available models rather than tying every solution permanently to a single provider. That gives teams room to think about what the use case actually needs. One model may be better suited to a particular task, while another may offer a different balance of performance or other requirements.
So the starting question does not have to be: "Which model do we use for everything?", it can be: "Which model fits this particular job?".
Sometimes the model knows the language but not your business
Let’s move down the corridor to HR. An employee asks: "Can I carry my unused vacation days into next year?". A language model can understand every word in that sentence and still not know the right answer for this organization.
Perhaps the company allows carryover only under certain circumstances, perhaps the policy changed this year. The information might sit in an internal policy document that was never part of the model’s original training.
That does not necessarily mean the organization needs to retrain the model. With grounding, relevant information from trusted business sources can be brought into the AI interaction as context. Document grounding, for example, can retrieve relevant content from enterprise documents so that the model can respond using information connected to the organization’s own knowledge.
Other forms of business context can go further, SAP Knowledge Graph can make business objects and the relationships between them explicit, helping AI work with the meaning and connections behind enterprise data.
The model still provides the language capability, but now it is working with information that belongs to the actual business. That changes the question from: "What does the model know?" to: "What does the model need to know to answer this correctly?".
When the business information arrives as a document
Across the company, Accounts Payable has a different problem: a supplier invoice has arrived. It contains the supplier, invoice number, dates, line items, totals, and other information the business needs, but those details are sitting inside a document rather than neatly arranged in the fields of a business application. Before anyone can approve, match, analyze, or act on that information, the document has to be understood.
This is the kind of problem SAP Document AI is designed to address. SAP Document AI supports end-to-end processing of business documents. It can ingest documents, extract text and information, classify and process content using a combination of AI technologies, and pass the resulting information into downstream business processes. It also provides preconfigured content for common business documents such as supplier invoices, purchase orders, and delivery notes.
The important point is that document processing is more than asking a language model to read some text, a document can have layout, fields, tables, images, and business-specific structures that all matter. SAP Document AI brings together technologies such as OCR, transformer-based models, generative AI, and vision capabilities to handle that kind of problem.
Once again, the business outcome comes first. The company does not really want "AI for PDFs."
It wants the information trapped inside those documents to become usable business data.

When the answer is hidden in rows and columns
Finance is asking a completely different kind of question: "Which of these customers is most likely to pay late?". The answer is not primarily sitting in a paragraph, an email, or a PDF.
It may be hidden in patterns across invoice amounts, payment terms, customer history, previous payments, and other structured business records.
We already saw in Unit 2 how SAP-RPT can work with structured business data for predictive tasks such as classification and regression. Instead of asking the AI to generate an explanation of why customers sometimes pay late, the business can ask a predictive question about the records in front of it. For example, Finance might use that prediction to identify invoices with a higher likelihood of late payment and decide where earlier follow-up would be useful. That distinction is easy to miss because both experiences may ultimately appear through an AI interface.
But underneath them, the jobs are different, one AI capability may help explain a situation in natural language. Another can identify patterns in structured data and estimate what is likely to happen next.
The best experience may use both.
Sometimes you need something more specialized
Consider the final team. Perhaps it is working with highly specialized operational data. Perhaps the business has developed a proprietary method that gives it a competitive advantage. Or perhaps the task is simply too specific for the models already available. In those cases, the answer does not have to be to force the problem into an existing model.
SAP also supports custom model training and model serving, allowing organizations to build business-specific models and deploy or serve their own models as part of an AI solution. The generative AI hub also supports an open architecture in which organizations can access SAP and third-party models or bring their own models. This creates a much broader set of choices than simply selecting one general-purpose model and trying to make every problem fit it.
The interesting part is what happens when they work together
Now let’s return to our company on Monday morning. The customer’s message may be interpreted using a language model, with the information needed to answer the question grounded in company and business context. A supplier invoice may enter the process through SAP Document AI, while SAP-RPT may identify a payment risk from structured records. A specialized model might contribute to a scenario unique to that organization, and an agent could potentially bring several of those capabilities together as part of a larger business process.
The user does not need to experience this as five disconnected technologies, from their perspective, they asked for an outcome. Behind that experience, however, different AI capabilities can each contribute what they do best.
This is where the generative AI hub becomes particularly important, it provides access to a range of models and supports model exploration and switching, while also providing orchestration capabilities such as grounding, data masking, content filtering, monitoring, and governance.

The aim is therefore not to discover the mythical one model that does everything. It is to create the right combination for the business problem in front of you.
Start with the problem, then choose the AI
It can be tempting to start an AI initiative by saying:" We want to use an LLM."
But that already assumes part of the answer. A better conversation starts a little earlier.
What is the business trying to accomplish? Is the challenge understanding language, extracting information from documents, predicting something from structured data, using company-specific knowledge, or solving a highly specialized problem?
Once that is clear, the right AI approach becomes easier to identify. And sometimes the answer will not be one technology at all, it will be several specialized capabilities working together, each contributing to the same outcome.
The real opportunity is not to put the same AI everywhere, it is to bring the right intelligence to each problem and connect those pieces when the business needs them to work as one.
You can also explore SAP Document AI to learn more about using AI to extract and process information from business documents.
Lesson Summary
- Different business problems can require very different forms of AI.
- Large language models can support language-based experiences, grounding can connect those models with trusted enterprise knowledge, SAP Document AI can turn business documents into usable information, and SAP-RPT can support predictions based on structured business data.
- When existing models do not fit the need, organizations can also use custom or externally provided models.
- The generative AI hub provides access to multiple models and supports capabilities such as model exploration, grounding, orchestration, security, and governance.
- The goal is not to force every business problem through one model, but to use and combine the AI capabilities best suited to the outcome the organization wants to achieve.