What You Can Actually Do With AI
Most useful products combine a small set of patterns: assist, retrieve, act, code, generate media, and structure information.
AI product announcements become easier to decode when you identify the job rather than the brand. Most current systems combine several recurring patterns.
1 · Assist. Explain a concept, draft an email, rewrite for an audience, translate, summarize, or propose options. Assistants are fastest when a human remains in the loop and the output is easy to inspect.
2 · Retrieve. Retrieval-augmented generation selects relevant documents or passages and puts them into context before generation. It is useful for manuals, policies, research, and current information. Retrieval can reduce unsupported answers only when it finds good sources and the response stays faithful to them.
3 · Act. An agent chooses tools, observes results, and continues across steps. It might search suppliers, update a ticket, or prepare a report. Each additional tool and step introduces another place for permissions, data, or reasoning to go wrong.
4 · Code. Coding agents can search a repository, edit files, run tests, and prepare a diff. Code is unusually suitable because compilers, linters, tests, and version control provide machine-readable feedback. Those checks are incomplete, so review still matters.
5 · Generate media. Models can create and edit images, video, music, speech, and sound. Diffusion is one important family, alongside autoregressive, flow-matching, and hybrid systems. Product capabilities and provenance controls vary; “AI-generated” does not name one engine.
6 · Structure. Extract fields from invoices, classify tickets, match records, detect anomalies, or turn documents into a schema. This less theatrical category often has the clearest acceptance criteria: compare extracted fields with labeled examples and measure error.
Choose the smallest pattern that solves the job. Autonomy is a cost and risk, not a badge.
Before building, define what a good result looks like, what evidence is available, and what happens when the system is uncertain. A one-step extraction with human review may create more value than a fully autonomous workflow whose errors are expensive and hard to notice.