Designing for AI Interfaces:
Patterns That Actually Work

Most AI interfaces fail not because the underlying model is weak, but because the interface forces users into a guessing game about what the model can do. A blinking cursor on a chat input is the most expensive UI element in modern software. It promises everything and shows nothing.

I built the chat widget on this site over the course of a few late-night sessions. It is a guided-flow conversational tool, not an open chat with a frontier model. That choice was deliberate. This piece is about why, and what that decision teaches about designing AI surfaces in 2026.

The blank-prompt problem

A blank chat input is the AI interface equivalent of a command line: powerful for experts, paralyzing for everyone else. Users do not know what the bot can do, what it cannot do, or how to phrase a question to get a useful answer. The result is a single-turn interaction or no interaction at all.

Bank of America rebuilt their Erica assistant in 2024 around this exact insight. They moved from a floating chat icon to a search-style interface. The reason was simple: older customers understood search, but felt uncomfortable with chatbot conventions. The redesign now serves 50 million users with 3 billion interactions, and 98 percent of users get answers within 44 seconds. Same underlying technology. Different mental model.

The lesson generalizes: match the interface to the user's existing mental models, not to the model's capabilities. The frontier of what AI can do is not the same as what users know how to ask for.

Five patterns that work

1. Communicate scope before the user types

Every successful AI interface I have studied or built starts the same way: the bot says what it can help with, in plain language, before asking the user anything. This is not a generic greeting. It is a scope contract. "I can help you check order status, return an item, or connect you to a human agent" sets a productive frame. "Hi! I am Maya, your assistant. How can I help?" does not.

The chat widget on Sirrona's site uses this pattern. The first message offers four routes (project inquiry, research, quick question, or direct contact) before asking anything else. Users self-select into the path that matches their need.

2. Structured options over open prompts

Quick-reply buttons, suggestion chips, and guided flows beat open text input for most use cases. Open input is for power users who already know what they want. Everyone else is faster, more accurate, and more satisfied with structured options.

This is unintuitive to designers who learned that "more flexibility equals better UX." For conversational AI, the inverse is true. Tighter scope produces better outcomes. The user spends less cognitive load formulating queries, and more attention on the answers.

3. Visible system state

When the model is thinking, show it thinking. When it is searching, show it searching. When it is uncertain, say so. Hidden processing is the source of most user frustration with AI interfaces. The user cannot tell the difference between "the bot is working" and "the bot is broken."

Typing indicators are the minimum. Better is showing what the bot is doing: "Searching your order history," "Looking that up in our docs," "Thinking through this with you." The honesty is calming. The visibility makes the interaction feel like a partnership rather than a black box.

4. Graceful failure with named exits

Every conversation will hit a moment where the bot cannot help. Plan for that moment. Do not return "I am sorry, I did not understand that." Return a clear next step: "I cannot answer that one. Would you like to talk to Aaron directly, or describe what you need a different way?"

The named exit (a specific person, channel, or alternative path) is the difference between failure that retains trust and failure that loses it. The bot that escalates well stays useful. The bot that loops on confusion gets closed.

5. Human escalation as a feature, not a fallback

The most adopted enterprise chatbots treat human handoff as a primary feature. Peloton's chatbot is built to escalate quickly. Bank of America's Erica explicitly tells users when she cannot help and routes them. Surfacing the human option does not reduce bot usage. It increases it, because users trust the bot more when they know they have an out.

For a small practice or single-operator business, this is not a 24/7 call center. It is a "leave your email and the principal will respond personally" path. The point is not the staffing model. The point is that the user always knows they can reach a person.

Anti-patterns that look smart but fail

The most expensive AI interface mistakes in 2026 are: anthropomorphizing the bot with a human name and avatar that pretends to be a person, hiding capabilities behind an open prompt, demanding the user phrase questions in specific ways, and treating typos or partial inputs as failures. Each one feels clever to the team that built it and frustrating to the users who encounter it.

A specific note on names and avatars: if your bot is going to use AI, make that visible. Calling it "Maya" with a stock photo headshot creates a false relational frame. Users feel deceived when they realize they have been talking to a model. The Sirrona chat widget identifies itself as "Sirrona Assistant" with the brand mark, not as "Aaron." That distinction matters.

The other one worth flagging is "personality" engineering. Do not write witty bot responses for entertainment value. Users are not there for entertainment. They are there to accomplish something. Personality should serve clarity, not distract from it.

When to use guided flows vs. open chat

This is the biggest design decision in any AI interface project, and most teams get it wrong. The default assumption is "we should let users ask anything." The right answer is almost always the opposite for non-expert users.

Use guided flows when your domain has clear, common user goals (order status, support tickets, lead qualification, appointment booking), most users are not experts in your service, and accuracy matters more than flexibility. This covers the majority of business use cases.

Use open chat when your users are technically sophisticated, the use cases are open-ended (research, content generation, code), and the cost of a wrong answer is low or the model can recover from it. This covers fewer use cases than the AI hype cycle suggests.

A useful hybrid is starting with guided flows and exposing open chat as an "ask something else" path for users who exhaust the structured options. This gets you the best of both: confidence-building structure for new users, and flexibility for the few who need it.

A practical checklist

If you are designing or auditing an AI interface for your business, work through these questions:

1. Does the first message communicate scope? Can a new user tell what the bot can and cannot help with within five seconds?
2. Are there structured options (buttons, chips) for the most common goals? Does the user always have a clickable next step?
3. Is system state visible? When the bot is thinking, processing, or searching, does the user know?
4. When the bot fails, does it offer a named exit? "Talk to a human," "Try rephrasing," "Email us" rather than just "Sorry, I did not understand?"
5. Is human escalation a feature, not a punishment? Is the option to reach a person always one click away?
6. Is the bot honest about being a bot? Does its identity match what it actually is?
7. When the conversation ends, what does the user get? An email follow-up? A booked call? A logged inquiry? Or just a closed window with no trace?

If you cannot answer all seven questions cleanly, the interface needs work. The model is rarely the problem in 2026. The interface around the model usually is.