What’s wrong with AI Products (not LLM’s)
The core issue with AI products today is simple: they barely exist. What most companies are calling “AI products” are little more than wrappers around large language models. Instead of building dedicated, thoughtfully designed tools that unlock the potential of this technology, they’ve taken a shortcut—drop a chatbot into a box, slap on a label that says “What can I help you with?”, and call it a product.
This is a far cry from how software used to be made. Traditional digital products are built with intent. They have structure. They provide menus, views, workflows, and feedback loops that are tailored to what the user is trying to do. A video editing app gives you a timeline. A project management tool gives you a calendar, kanban board, and task hierarchy. A note-taking app gives you folders, tags, and version history. These interfaces exist for a reason—they reflect the underlying logic of the task at hand.
AI products, by contrast, are largely interface-less. Or more accurately, they’re interface-minimalist to a fault. They rely almost entirely on natural language input and thread-based output. For a few narrow use cases—like asking a trivia question, summarizing an article, or drafting a quick email—this can be incredibly powerful. But for more ambitious tasks, it quickly falls apart.
Let’s look at a few real examples:
Brainstorming: AI can help generate ideas, but all the output is dumped into a linear thread. There’s no way to cluster, group, or visually navigate your ideas. Want to compare options side by side? Too bad—you’ll be scrolling.
Project management: Some models can suggest tasks or timelines, but there’s no actual structure. No deadlines, dependencies, boards, or task hierarchies. You’re expected to track everything manually in one long, unstructured thread.
Research: LLMs are great at summarizing and analyzing content, but there’s no dedicated interface to store sources, organize findings, or highlight contradictions over time. It’s up to the user to keep things coherent across sessions.
Writing and editing: Drafting is easy. Refining is not. You can’t compare tone options side by side. You can’t manage multiple drafts. There’s no way to define a consistent style across sessions or versions.
Thread management: Long conversations become a mess. You can’t branch, tag, summarize, or version a conversation. There’s no notion of scope, only an endless scroll of loosely related messages.
It doesn’t have to be this way.
Take Notion, for example. At its core, it’s just a digital notebook. But Notion doesn’t stop there. It offers structure. Formatting options. Templates. Embeds. Plugins. Databases. Views. Page linking. In short, it gives you the tools to build with your notes, to shape information in ways that reflect how you think and work. You’re not just left with a blank page—you’re given a set of flexible building blocks. That’s what a good product does: it turns potential into power. It gives you leverage.
By comparison, most AI products hand you the equivalent of a blank notepad and expect you to remember the right incantations to make it useful. And if it doesn’t work the way you want? You’re told to keep prompting it until it does. That’s not a product—that’s a sandbox with a model inside.
Worse, these products can’t even get the basics right. I’ve asked ChatGPT over 50 times to use sentence case—that is, to capitalize the first letter of each sentence, as has been standard in written English for centuries. Despite this, the system frequently defaults to lowercase. When I reached out to OpenAI support, their advice was to start each new prompt by restating my formatting preference. Imagine opening the New York Times app and seeing every headline, subhead, and article in all lowercase. You write to support and they tell you to type “Please use proper formatting” every time you launch the app. That’s the level of absurdity users are being asked to accept.
And if that weren’t enough, AI products often test changes directly on paying users—without warning or opt-in. You might open the product and suddenly be part of an A/B test you never asked for. The interface changes. Or instead of a single response, you get three different replies with a note saying “We’re experimenting with multiple perspectives.” But you weren’t asking for multiple perspectives. You were asking for a quick answer. Now you’re stuck parsing experimental results instead of getting what you came for.
This isn’t innovation. It’s friction disguised as progress. It turns users into test subjects without consent and erodes trust in the experience.
The irony is that LLMs themselves are capable of nuance, depth, and flexibility. But most AI products aren’t. Because most of them aren’t actually products. They’re thin interfaces sitting on top of models, with almost no thought given to experience design, workflow, or structure.
Until AI companies start treating product design as a first-class priority—not a cosmetic layer—we’ll continue to have powerful technology trapped inside frustrating, inconsistent experiences. The next wave of breakthroughs won’t come from better models alone. They’ll come from better products—designed with care, built around real user needs, and capable of supporting real work.