When a company says it is an "AI company", what did it actually build?

When a company says it is an "AI company", what did it actually build?

It built one of five product shapes: a chatbot, an AI feature inside an existing product, an agent that does a task end to end, a workflow with a model in the middle, or a wrapper around someone else's model. "AI company" and "AI-powered" cover all five, and the five involve very different engineering.

The sorting matters because it tells you what the company's developers actually spent their days on. Two companies can use identical language on their websites while one team works with models every hour and the other touched a model for two weeks last spring. The shape of the product settles which one you are looking at, and this lesson is the field guide to the shapes.

Three of them, the chatbot, the agent, and the workflow, were covered in the plumbing lesson. The two new ones here are the AI feature and the wrapper, and the wrapper carries the most loaded word in the AI vocabulary.

What shapes does an AI product come in?

Five shapes cover nearly everything on the market.

A chatbot is the whole product as a conversation. The company sells the chat window itself, usually grounded in its own documents through RAG.

An agent is a product that completes a task end to end: it books the meeting, files the claim, or writes the code, taking its own steps along the way. Agent companies sit closest to the frontier, and their engineers spend the most time on model behavior.

A workflow product runs a fixed pipeline with a model doing one or two steps inside it. An invoice-processing tool that reads documents with a model and routes them with ordinary code is this shape. Most "AI automation" companies are workflow companies.

An AI feature lives inside a product that existed before AI. The product works without it, and the feature makes one part of it faster.

A wrapper is a product whose entire value comes from a model somebody else trained. The word gets its own sections below, because it is the shape developers argue about.

The shapes overlap in practice. A wrapper can be a chatbot, and an agent can live inside a feature. What the shapes sort cleanly is engineering time: how much of the team's work touched the model, and how much went to everything around it.

What is a wrapper?

A wrapper is a product built on a model someone else trained. The company sends prompts to a rented model over an API, shapes the answers, and sells the result. The name is literal: the product wraps a layer of software around a model at its core, the way a candy wrapper goes around the candy.

You already have the pieces to place this. The labs train foundation models, and nearly everyone else rents access to them. In the language of building versus using, a wrapper company sits fully on the using side: it pays for inference and builds a product around the output. By that definition, most of the AI industry is wrappers. The legal-research startup, the sales-email tool, and the medical-scribe product almost certainly all run on a model from one of the same three or four labs.

Is a wrapper an insult or a description?

Both, and knowing which is in play depends on who is talking. When a developer dismisses a product as "a wrapper," the charge is thinness: a chat box over an API, a product any competent developer could rebuild in a weekend, with no moat once the lab ships the same capability itself. The insult has real cases behind it. In 2023, hundreds of products launched that were a prompt, a payment page, and someone else's model.

Used plainly, though, the word describes almost every AI application company, including heavily funded, widely respected ones. The companies behind Cursor and many of the best-known AI products build on rented foundation models. The word tells you where the model came from. It does not settle whether the engineering on top is thin or deep, and that is the part worth finding out.

What makes a wrapper hard to build?

The model is the one part a wrapper company skipped. Everything else on the shelf stays, and it is most of the work.

The data layer decides what the model gets to see, which usually means a RAG system over the customer's documents, kept current and searchable. Permissions decide what each user's requests are allowed to touch, and getting that wrong in a legal or medical product is a company-ending event. The interface decides whether the model's raw text becomes something a professional trusts, which is ordinary product engineering of a high order. Cost per request needs managing, because every user action spends real money on inference, a pressure most software never faces. And evals decide whether the whole thing is actually good, measured over thousands of real cases rather than a demo.

Some wrapper teams also do fine-tuning, adjusting a rented model on their own data, which pulls them a step toward the building side. A developer from a strong wrapper company has usually worked several of these layers, and every one of them transfers.

What does an AI feature inside an existing product mean?

An AI feature is a model added to a product that already worked. The email tool gains a "draft this reply" button. The design tool gains image generation. The spreadsheet gains a formula that summarizes text. The product's core kept its old engine, and a model now powers one corner of it.

The common form has a name: a copilot is an assistant built into an existing product that drafts or suggests while a person works, with the person approving what it produces. GitHub Copilot, which does this inside a code editor, made the word generic, and now every product category has one.

For reading a team, the feature shape means the engineers spent most of their time on the product and some of it on the model. A developer who shipped a copilot feature at a big software company did real AI integration work: prompts, evals, and the plumbing to a rented model. The surrounding years of their experience were the product itself. That mix is genuinely useful, and it reads differently from a team whose entire product is the model interaction.

How do you read "AI-powered" in a company blurb?

Run one test: does the product stop working if the model disappears? A company whose product dies without a model is AI-native, whatever its shape, and its engineers work with models daily. A company whose product would keep running with one button grayed out has an AI feature, and its engineers mostly work on the rest.

The blurb usually gives the shape away. "AI assistant for X" is a chatbot or copilot. "Automates X end to end" is an agent or a workflow. "AI-powered X platform," where X existed as an industry before 2022, is usually a feature on an existing product. A launch date after 2022 plus a product that is entirely the model interaction is a wrapper, and the depth question stays open until you see what they built around it.

None of these shapes outranks the others as a place for talent to come from. They produce different experience. The anchor guide places all five on the full map, and the next lesson in this module gives the honest account of what the technology delivers today.