3.3.3. Information Extraction and Agentic Workloads
💡 First Principle: Two workloads dominate the AI-901's implementation half. Information extraction pulls structured fields out of unstructured content (documents, images, audio, video) — the job of Azure Content Understanding. Agentic AI goes a step beyond generation: an agent reasons about a goal and takes autonomous actions (calling tools, querying data) to achieve it, rather than just returning text.
Information extraction is more than OCR: OCR gets you raw text, but extraction identifies meaningful fields — the invoice total, the contract date, the customer name — and returns them in structured form ready for an application. Agentic AI is the difference between a model that answers "what's the weather?" and an agent that checks a weather tool and books you an umbrella delivery. These two workloads anchor Phases 6 and 4 respectively, so recognizing them now sets up the implementation work.
⚠️ Exam Trap: Don't equate an agent with a chatbot. A chatbot generates a text reply; an agent can decide to call tools and take actions to fulfill a request. The presence of autonomous action toward a goal is the signal for "agentic," and it's heavily tested.
Reflection Question: Why is "extract the total amount due from this scanned invoice" an information-extraction task rather than just an OCR task? What does extraction add?
