AI, Machine Learning & Generative AI: The Map

Picture the meeting. A steering committee, a slide with “AI Strategy” in 60-point font, and a client sponsor turning to you — the SAP person in the room — and asking: “So what’s actually the difference between AI, machine learning, and this generative AI everyone keeps talking about?”

If your honest answer is “they’re all kind of the same thing,” you’re about to lose the room. And worse: you’re about to design a Joule rollout, a RAG pipeline, or an “AI agent” project without knowing which of these three very different technologies you’re actually building with.

This is lesson one of SAP AI Foundations, and it’s deliberately short. Its only job is to hand you a map that’s accurate enough to survive a client’s follow-up question — because a wrong map here compounds into a wrong architecture twelve lessons from now.

Four words, four different things

Here’s the thesis of this entire lesson, in one line: these four terms are not synonyms, and conflating them is the single most common credibility mistake SAP consultants make when they start talking about AI.

• Artificial Intelligence (AI) is the umbrella term for machines doing things that look intelligent. That’s it. Nothing here requires learning from data — a hand-coded tax-calculation engine or an SAP validation rule counts as AI under this definition, and always has.
• Machine Learning (ML) is the slice of AI that learns patterns from data instead of being explicitly programmed rule-by-rule. A churn-prediction model that gets smarter as it sees more customer data — that’s ML.
• Deep Learning is ML built on large neural networks. It’s the engine that made the last decade of AI breakthroughs possible.
• Generative AI is deep learning that creates new content — text, code, images — instead of just classifying or predicting a number. The text-generating members of this family are Large Language Models (LLMs): the models behind ChatGPT, SAP-ABAP-1, and everything inside Joule.

Resist the urge to draw this as four neat circles, one inside the next. It isn’t that clean — plenty of AI, like rule-based systems, sits entirely outside the machine-learning circle. The loose version is close enough: generative AI is a kind of deep learning, which is a kind of machine learning, which is one branch of AI — with rule-based logic sitting quietly off to the side, still very much AI, never having learned anything from data at all.

Get this straight once, and you’ll never again nod along vaguely when someone says “the AI” as if it’s a single thing.

The idea that will actually get you in trouble

Here’s where it gets personal. Everything you’ve built your SAP career on — ABAP, configuration, testing — rests on one assumption: determinism. Same input, same output, every single time. That assumption is why you can write a unit test, why a transport either works or it doesn’t, why “it worked in QA” is supposed to mean something in production.

Generative AI breaks that assumption on purpose. An LLM predicts the next most likely token — a word-fragment — over and over. Ask it the exact same question twice, and you can, quite legitimately, get two different answers.

This is not a bug SAP or OpenAI or anyone else is going to patch. It’s the mechanism. And if you walk into a client demo expecting Joule or any generative AI feature to behave like a BAPI — same call, same result — you will eventually be standing in front of a steering committee explaining why the answer changed since last week’s rehearsal.

Once this lands, three things fall out of it immediately, and they’re the spine of the rest of this series: how you prompt a model to behave reliably (Lesson 3), how you ground it in real data so it stops guessing (Lesson 5), and why a human still has to approve anything consequential it does (Lesson 13). Everything else in SAP AI Foundations is really just working out the consequences of this one sentence.

Where each layer actually shows up in your SAP landscape

You are not choosing one of these technologies for your career — you already work with all of them, probably in the same week:

• Rule-based AI runs the business processes you configure every day — pricing procedures, validation checks, output determination.
• Predictive ML is quietly behind things like cash-flow forecasting and demand planning.
• Generative AI is what powers Joule, and increasingly, the developer tooling you use to write and review ABAP.

The professional skill isn’t “know AI.” It’s knowing, in any given conversation, which of these three you’re actually talking about — because a forecast model earns your trust statistically, over many predictions, while a generative answer earns nothing automatically. You verify it, every time.

Try it yourself — five minutes, and you’ll feel the difference

Theory is cheap. Here’s how to make non-determinism and hallucination real, not just a phrase you nod at:

1. Open any public LLM — ChatGPT, Claude, Gemini, doesn’t matter which.
2. Ask it, in three separate fresh chats, the exact same SAP question: “What does the function module BAPI_SALESORDER_CREATEFROMDAT2 do, and list its key parameters?”
3. Compare the three answers word for word. The wording will differ — that’s non-determinism, live, in front of you.
4. Now hunt for one detail that’s plausible but wrong — a parameter that doesn’t actually exist, a description that’s subtly off. Write it down.

That fabricated, entirely confident detail is a hallucination, and you just produced one on demand. Nobody had to explain the concept to you — you watched it happen. That’s the entire reason “verify before you trust” isn’t a slogan on this site, it’s rule number one, and it’s why every lesson from here has a > Where this shows up in SAP callout instead of a bare claim.

Key takeaways

• AI ⊃ ML ⊃ deep learning ⊃ generative AI ⊃ LLMs — loosely, and rule-based AI sits outside ML entirely.
• Generative AI is probabilistic, not deterministic. Same input does not guarantee the same output — plan your demos and your architecture accordingly.
• LLMs hallucinate: confident, fluent, and sometimes wrong. Structural, not a bug, not going away.
• SAP uses all three layers today. Your job is knowing which one you’re standing on when someone asks you a question in a meeting.

Next lesson, we open the LLM up and look inside: tokens, context windows, and the answer to a question every architect eventually has to face — what does this actually cost per call?