Generative AI is the branch of artificial intelligence that creates new content, such as text, images, audio or code, from the patterns it learned from a huge number of examples. One verb answers “what is generative AI?”: create. It doesn’t choose among options that already exist. It makes one that didn’t.
North Light Studio, in Portland, Maine, shoots portraits and product photos. It has used AI for years without calling it that: the camera detects faces and focuses, and the inbox sets the junk mail aside. That AI recognizes and classifies. Generative AI does something else: you ask it to “describe the portrait session for the website” and it writes a paragraph that didn’t exist anywhere.
During training it sees an enormous number of examples and learns what tends to go with what. When you ask for something, it composes a new result that resembles what it learned. For text, the engine is a large language model, or LLM, which writes one word, or piece of a word, at a time; those pieces are called tokens. What you type to ask for it is called a prompt.
The older kind answers with a label or a number: “this is junk mail,” “this customer will come back.” Generative AI answers with a whole piece of work. That’s where the usual confusion comes from: not all AI generates, and generative AI isn’t “the good AI.” It’s a different job.
Generative AI produces something when you ask: a draft, an image, a summary. “Agentic AI” is the name for systems that put a generative model to work on a goal: they plan steps, use tools such as a search or a calendar, and act with less hand-holding. So agentic AI is a layer on top, not a rival: an AI agent uses generative AI to decide what to do next and to write its output. The more such a system does without asking, the more its limits and its review step matter.
Writing a first draft, summarizing a long document, answering questions, suggesting ideas, changing the tone of a text. At the studio: the write-up for each session, the reply to someone asking for a quote, the caption for a photo. To start in an orderly way, there’s the guide to implementing AI in a small business; the rest is in the guides.
It makes things up: what it produces sounds right even when it’s false, and that goes by the name AI hallucinations. It doesn’t know your company unless someone has told it about it. And what it creates has to be reviewed, because nobody has checked it before: it didn’t exist.
Four things generative AI produces, with an example of each. Switch to “Classify” to see the older kind of AI next to it.
Describe the portrait session for the studio’s website
One hour in the studio, with natural light and a plain backdrop. No stiff poses: we talk, we try things, we shoot. You take home twenty edited photos, picked together with you.
Flipa: yes, that’s its job It writes from what it has read about your company, and someone reviews it.
A product still life on marble, lit from the side
Flipa: works with text It does read photos of documents and scanned pages.
A twenty-second voicemail greeting
Flipa: doesn’t read audio Its job is text. It doesn’t read video either.
A function that works out the price of a session
function sessionPrice(hours, rate) {
return hours * rate;
}Flipa: drafts text Its job is your company’s writing.
An email comes in→A label comes out: “junk”
It writes nothing. It decides which folder the email goes to.
Your purchases go in→A ranking comes out: “you might like”
It doesn’t invent products. It sorts the ones there are.
These examples were written by hand, not taken from an AI: they show what kind of thing each one produces.
| Where | The older kind of AI | Generative AI |
|---|---|---|
| In email | Decides whether a message is junk | Writes you a draft of the reply |
| In photos | Recognizes a face and focuses | Creates an image from a description |
| In an online store | Recommends products based on what you bought | Writes a product listing |
Flipa uses generative AI mainly for one thing: writing. It answers and drafts text from what it has read about your company, shows the sources it checked and says when it doesn’t know something. It reads photos of documents and scanned pages; it doesn’t read audio or video. What it writes is a draft that someone reviews. The app’s interface is in Spanish today. The step-by-step is in how it works.
It’s the part of artificial intelligence that produces content (text, images, audio, video or code) instead of only analyzing what’s already there. What it produces is new, even if it resembles what it learned.
In two phases. First it’s trained on an enormous number of examples until it picks up their regularities. Then, when a request comes in, it generates the result step by step: for text, one word or piece of a word at a time. It doesn’t look answers up in a file of ready-made ones, which is why it can be wrong and still sound sure.
AI is the whole field: it includes systems that classify, predict or recommend. Generative AI is the part that produces content. A spam filter is AI; an assistant that drafts your reply is generative AI.
Generative AI creates content when you ask for it. Agentic AI describes systems that use a generative model to pursue a goal over several steps, calling tools and making decisions along the way. One writes the draft; the other also decides what to do next, so it needs clearer limits and a review step.
Mostly for getting past the blank page: drafts of emails, product listings, posts and summaries. It does better when it knows the company, because it stops writing generalities. And always with someone reviewing before anything goes out.
How to start using AI at a company, step by step and without jargon.
The engine of generative AI when what it generates is text.
Why it makes facts up, and how to keep one from reaching you.
Where to start, who to involve and what not to do in the first month.
Flipa writes from what it has read about yours, and shows you where it got it.