There’s a pattern I keep seeing across every SaaS product that shipped an AI feature in the last 18 months.
A text box. A blinking cursor. A placeholder that says “Ask me anything.”
And then nothing. No rails. No suggestion. No signal that the user is doing it right. Just a void, and the implicit assumption that your user knows exactly what to type to get value from your product.
They don’t. And that empty box is costing you more churn than your pricing page.
We accidentally imported the ChatGPT interface into everything.
OpenAI built a blank canvas because they were launching a general-purpose tool. They had no choice. The use cases were infinite. That design decision made sense for them.
You are not building a general-purpose tool. You have a specific user, with a specific job, inside a specific workflow. And yet you put the same empty box in your product because it looked like AI.
This is cargo cult UX. Copying the aesthetic of intelligence without doing the work of making your product actually usable.
The result? Users stare at the box, type something too vague, get a mediocre response, and conclude that your AI is bad. It’s not bad. It’s just unguided. You handed someone a Formula 1 car and forgot to tell them it needs a specific key.
The best AI UX I’ve seen doesn’t feel like AI at all.
Linear doesn’t ask you to “chat with your issues.” It surfaces the right issue at the right moment, learns your workflow, and gets out of the way. Superhuman doesn’t prompt you to “ask the AI to write your email.” It inserts triage logic invisibly into a flow you already have.
The pattern: AI as acceleration, not AI as interface.
The failure pattern: AI as a feature bolted onto a product that wasn’t redesigned to hold it.
When AI is the interface, your UX depends entirely on your user’s ability to prompt. That’s a massive skill gap you’re handing them as homework. When AI is embedded in the workflow, your UX carries the user. Prompting becomes optional. Outcomes become consistent.
One model builds habits. The other builds confusion.
Three questions to audit your AI UX right now.
1. What does a first-time user type first? If you don’t know, go watch five session recordings. If they’re typing things like “hello,” “test,” or “what can you do?” you have a discovery problem, not an AI problem.
2. What’s the worst possible input your AI can receive? Design for that. Show an example output before they type anything. Give them three starter prompts that demonstrate the range. Constrain the surface area until they understand the space.
3. Does the AI output look different from everything else in your UI? It should. Prose dumped into a product interface is jarring. Think about how the response is formatted, chunked, and actionable. A wall of text is not a UX decision. It’s the absence of one.
The silent killer: AI features don’t get support tickets.
When a user is confused by a form, they submit a ticket. When they’re confused by an AI feature, they just stop using it. Then they churn quietly and tell your retention dashboard it was about price.
This is the most dangerous thing about AI UX debt: it’s invisible until it’s terminal.
The fix isn’t more powerful models. It isn’t a better system prompt. It’s the same thing that’s always fixed bad UX: sitting with real users, watching them fail in real time, and designing around the failure instead of pretending it doesn’t exist.
The products that will win the next three years aren’t the ones with the most capable AI.
They’re the ones that make users feel capable.
That gap? That’s still a design problem. And it’s yours to solve.