An agent shopping for AI in 2026 runs into a genuine problem: the products describe themselves in nearly identical language, and the language doesn’t map to what the products do. Assistant, copilot, agentic, intelligent platform — the words have drifted loose from their meanings.
Underneath the vocabulary there are really only four categories. Each solves a different problem, each has an honest limitation, and most confusion comes from comparing a product in one category against a product in another.
1. General-purpose chat
The browser-based models. You open a tab, you type, you get output.
What it’s genuinely good at: writing tasks where you supply the specifics. Listing descriptions, a rewrite of an awkward email, a social caption, talking through a negotiation. It’s capable, it’s cheap or free, and it’s the right first step for anyone who hasn’t used AI at all.
The honest limitation: it knows nothing about your business. Every task starts from zero. The output quality tracks almost perfectly with how much context you paste in, which means the tool’s real cost isn’t the subscription — it’s the gathering. We wrote about that math in detail in the hidden cost of AI prompts, and it’s the reason most agents use prompts enthusiastically for two weeks and then stop.
If you want to start here, our library of 25 prompts for agents is free and includes exactly what you have to paste in for each one.
2. Point tools
Single-job products. A listing-description generator. A video or virtual-staging tool. A transcription service. A social scheduler with AI captions.
Good at: their one job, usually better than a general model, because the workflow is built around it.
The limitation: they don’t compound. Five point tools is five logins, five bills, and five systems that don’t know about each other. The listing tool doesn’t know the buyer tool exists. Useful, but they stay useful at a fixed level forever — using one for a year doesn’t make it better at knowing you.
3. AI features inside a platform
The big real estate platforms have all shipped AI, and this is where the shopping gets hardest, because the category spans an enormous quality range under one description.
At the light end: a chat box added to an existing database, or auto-generated content that sounds like everyone else’s. At the heavier end: lead scoring, automatic routing, response drafting, and behavior alerts that actually watch the pipeline.
The question that separates them is the one we keep coming back to when decoding “AI-powered”: what does it know about you specifically, and what does it do when you’re not looking at it? A feature that requires you to open it and ask is a point tool that happens to live inside your database. A feature that notices something and tells you is a different animal.
The limitation that applies across the whole range: it’s an addition to a system designed for storage. The database came first and the intelligence was fitted to it, which caps how deeply the two can be connected.
4. Systems built around the AI
The smallest and newest category: software where the model isn’t a feature but the organizing principle. The system is given your material — your writing, your history, your contacts, your documents — and everything else is arranged around what that makes possible.
Good at: the work that requires knowing your specifics. Drafting in your voice without being told how you write. Handing you a call list with reasons attached. Reading a document and flagging the problem. Getting more useful the longer it runs, because the material it’s working from grows.
The honest limitations, and they’re real ones. It requires actual setup — someone has to put your material in and build your rules, and if that setup doesn’t happen, you own an expensive empty shell. It’s a newer category, which means fewer integrations and shorter track records than the incumbents. And the trust question is heavier: you’re handing over more, so where the data lives and who owns it stops being a footnote.
This is the category Theia Vault is in, and we’d rather state the limitations than have you discover them in month two.
How to use the taxonomy
Figure out which category you’re actually looking at, then judge it on that category’s terms.
A point tool that does one thing beautifully isn’t failing because it doesn’t know your database — that was never the deal. A platform feature isn’t a private AI no matter what the homepage says. And a system built around the model should be held to a much higher standard on setup, ownership, and what it does unprompted, because those are the things it’s claiming.
Most bad software decisions in this space aren’t a matter of picking a bad product. They’re a matter of buying category three while expecting category four, or paying category-four prices for category one with a login screen.