What AI Actually Is

A quick note before we start

If you've seen older study material referencing "AI-900," here's the update: Microsoft retired AI-900 on June 30, 2026 and replaced it with AI-901, same certification name, Azure AI Fundamentals, genuinely different exam underneath. The old version tested whether you could describe AI concepts. AI-901 tests that too, but roughly 55 to 60 percent of it is now about actually building something using Microsoft Foundry, Azure's unified platform for AI development. This course is built for the exam that exists today.

The return desk that couldn't decide

Aditi managed customer support for an online retailer. For years, returns worked off a fixed rulebook: a program checked a handful of conditions, item within 30 days, receipt attached, category eligible, and spat out approve or deny. It was software in the traditional sense: given the same input, it always produced the same output, because a person had written out every rule in advance.

Then Aditi's team added a system that read the customer's actual return message, "the blender stopped working after two uses, box is a bit damaged from shipping," and made a judgment call closer to what an experienced human agent would make, weighing context the fixed rulebook never could have anticipated. That's the shift from traditional software to AI.

What actually makes something AI

Traditional software follows rules a person wrote explicitly: if this, then that. AI systems are different in one specific way: instead of being told the rules directly, they're trained on examples, and they learn patterns from that data well enough to handle situations nobody explicitly programmed for.

That's the whole distinction, underneath all the terminology. A traditional program can only do what its rules anticipated. An AI system can respond reasonably to a return message it's never seen worded quite that way before, because it learned the pattern of "reasonable return" from many examples, not from one rigid rulebook.

The shapes AI work comes in

AI-901 organizes AI workloads into a handful of categories, worth knowing by name since they structure the rest of this course:

• Generative AI — creates new content: text, images, code, based on a prompt. This is what most people picture first now.
• Agentic AI — goes a step further than generating a response, an agent can take actions, call tools, and carry out a multi-step task with some autonomy.
• Computer vision — understanding and generating images: detecting objects, reading text in a photo, generating an image from a description.
• Language (NLP) — understanding and generating text: sentiment, translation, summarization, exactly what handled Aditi's return message.
• Speech — converting speech to text and text to speech.
• Information extraction — pulling structured data out of unstructured documents, invoices, forms, contracts.

Later modules in this section go through each in more depth; the module after this one covers how machine learning actually produces a working model in the first place.

Why "it works" isn't the only bar

Aditi's new system makes judgment calls a fixed rulebook never could, which is exactly why it needs guardrails a fixed rulebook never needed either. This is Responsible AI, and it's tested throughout AI-901, not as a side topic: fairness (not systematically favoring or disadvantaging any group), privacy and security (protecting the data the system was trained on and the data it processes), transparency (people should be able to understand why a decision was made), and accountability (a human, not the system itself, is ultimately responsible for the outcome).

A return-approval AI that quietly denies claims from one demographic more often than another is a fairness failure, even if its overall accuracy looks fine on a dashboard. Responsible AI is what catches that before it becomes a real harm.

Check your understanding

Five questions, all from this module. Get 4 out of 5 to mark it complete.