SAP Certified - SAP Generative AI Developer

Generative AI and LLM Foundations

Generative AI work begins with a disciplined model of what a large language model can do, what it cannot know on its own, and what an enterprise solution must add around it. This lesson connects the basic vocabulary to SAP use cases without pretending that a fluent response is automatically a reliable business result.

Start with the Problem, Not the Model

A large language model predicts and generates language from patterns in its training. That ability can support summarization, classification, drafting, extraction, question answering, and assisted workflows. It is not the same as a guaranteed source of current company facts, an approval authority, or a replacement for process ownership. Before discussing a model, write the job in plain language: who needs help, what decision or task is being improved, what information is allowed, and what an unacceptable result looks like.

Consider an employee who needs an explanation of a policy. The useful system outcome is not “produce text.” It is “provide a concise explanation that is based on an approved policy, identifies uncertainty, and lets the employee find the underlying source.” This framing reveals why a model alone is insufficient. The solution needs a trusted information boundary, a user experience that shows source context, and a way to evaluate whether the answer helped.

Use LLM Vocabulary as a Decision Tool

A prompt is the instruction and context presented to a model. A response is generated output. Context can include a role, rules, examples, user input, or retrieved information. Tokens are units used when text is processed. These words matter because they describe design levers. If output is too broad, improve the task, constraints, examples, or context before assuming the model is defective. If an answer needs company-specific facts, supply controlled evidence rather than asking the model to guess.

Separate deterministic rules from generative judgment. A rule such as “only managers may approve a request” belongs in the surrounding application and authorization design. Generative AI can explain the request, summarize evidence, or help the manager compare options. It should not silently invent authority. In an exam scenario, a correct choice often preserves this separation: business controls remain explicit, while the model adds language understanding or generation where it creates value.

Understand Data as a Product Responsibility

Data quality is not only a technical cleanup exercise. It determines whether a generated answer can be relevant, traceable, and safe to use. Ask whether a source is current, authoritative, complete enough for the question, and permitted for the intended audience. A document that is accurate for finance staff may be inappropriate for a broad employee assistant. A dataset may be technically reachable while still lacking a valid business reason for use.

Use a simple source card for every information collection: owner, purpose, audience, update rhythm, access boundary, and expected failure mode. The failure mode might be an outdated procedure, duplicated policy, ambiguous wording, or sensitive content. This is a practical mental model for grounding and evaluation later in the course. It also prevents the common mistake of treating all available data as equally suitable context.

SAP's current foundation learning includes enterprise LLM risks and SAP's relevant, reliable, and responsible AI approach. Apply those ideas as a set of questions: is the output relevant to the actual task, can its supporting evidence and behavior be examined, and are people, data, and affected processes treated responsibly? Prompt injection is one example of an enterprise risk. A user-supplied instruction must not be allowed to override the system's access boundary, approved task, or safety constraints. The practical response is layered: constrain the application action, separate trusted instructions from untrusted input, limit tools and data to what the user may use, and evaluate attack-like requests before release.

Map a SAP Scenario to an AI Outcome

SAP's current generative AI hub learning journey focuses on solving business problems using SAP's generative AI capabilities. In a scenario, map the request through four layers. First, name the business outcome, such as faster support response or more consistent document review. Second, identify the model behavior needed, such as extraction, summarization, or grounded question answering. Third, identify the SAP environment and governed data or services involved. Fourth, define the evidence that shows the feature works for the intended users.

This mapping avoids product-first thinking. For example, a support assistant may require an orchestration flow that identifies intent, retrieves controlled knowledge, calls a permitted service, and returns a response with clear limits. A proposal that only says “use an LLM” has not yet described a usable solution. Ask what must happen before the prompt, during the model interaction, and after the response is returned.

Practice a Failure Analysis

Take a generated answer that sounds plausible but lacks source support. Diagnose it in order. Was the task clear? Was relevant information supplied? Was the supplied information authoritative and current? Did the prompt ask the model to distinguish evidence from inference? Was there an evaluation that would catch the unsupported claim? Each question points to a different improvement. Rewording a prompt may help, but it cannot repair missing or unauthorized enterprise evidence.

When comparing multiple answers, avoid selecting solely by style. Score each one against a purpose statement: factual support, usefulness, completeness, tone, safety, and handoff behavior when it cannot answer. A shorter answer with clear sources and an honest limitation can be better than a confident answer that obscures uncertainty.

Readiness Checkpoint

• I can distinguish model generation from an enterprise system's rules, authorizations, and source ownership.
• I can state the business outcome, data boundary, and failure condition before choosing an AI approach.
• I can explain why data access is not the same as data suitability.
• I can diagnose an unsupported response without blaming the model alone.

Official Scope and Verification

Verification date: 2026-08-01. Sources: SAP generative AI hub learning journey, SAP course on LLM fundamentals and techniques, current SAP Generative AI Developer certification, and SAP artificial intelligence learning hub.