What we mean by AI in 2026
AI is software that imitates human behaviors and capabilities. In production, it shows up as a set of complementary workloads — each one solves a different class of problem and demands different data, infrastructure, and evaluation patterns.
The seven workloads worth knowing
- Machine learning: the foundation — train models that learn from data and generalize to new inputs.
- Computer vision: interpret images, video, and live cameras for inspection, search, and safety.
- Natural language processing: understand and generate written and spoken language.
- Document intelligence: extract structure and meaning from forms, contracts, and high-volume document flows.
- Knowledge mining: turn large volumes of unstructured content into searchable, governed knowledge stores.
- Generative AI: create original content — text, images, audio, code — under controllable constraints.
- Agentic AI: orchestrate models, tools, and policies to take multi-step actions on behalf of a user or system.
Production lens
Treat each workload as a system, not a model. Pair models with data pipelines, evaluation harnesses, observability, and clear human-in-the-loop policies. The most reliable AI systems are boring: small models for routing, larger models for reasoning, deterministic tools for actions, and strict guardrails for risk.
Modernization note
Refreshed the original Microsoft AI page into a structured learning module with categories, examples, and an applied production lens.
Originally published at /ai/microsoft-ai