---
title: "Designing AI experiences: a designer's guide to chat, personalization, and LLM-powered UI"
description: "A practical guide for designers working on AI-powered experiences: when to use chat, how to design LLM-driven personalization, handling uncertainty and errors, and prototyping with real models."
url: "https://mikestilling.com/guides/designing-ai-experiences/"
date: "2026-11-22"
author: "Mike Stilling"
---

> AI-generated from Mike Stilling's work, notes, and writing.

# Designing AI experiences

Mike Stilling's work at Stripe has grown to include AI-powered experiences, from live chat on stripe.com to [LLM-driven personalization](https://stripe.com/personalize). Mike also runs small AI experiments on this site, like a [personal agent](https://mikestilling.com/experiment/personal-agent/) that answers questions about Mike's work. This guide covers what designers should think about when the interface is partly written by a model.

## The short answer

**Good AI experiences are designed around what people need, not around the model.** The fundamentals still apply: clear hierarchy, honest copy, and less friction. What's new is that the output varies, can be wrong, and takes time to arrive. Designing for those three facts is most of the job.

## Chat is not always the answer

One of Mike's hot takes is that design is, in part, responsible for the rise of chat interfaces: web pages got so heavy with decoration that a plain text box started to feel like relief. Chat has real strengths, but it isn't the right answer for everything.

| | Chat works well when... | A designed interface works better when... |
|---|---|---|
| **The question** | Is open-ended or very specific to the person | Is common and predictable |
| **The user** | Knows what to ask | Doesn't know what's possible yet |
| **The answer** | Needs explanation or back-and-forth | Is best shown visually, compared, or scanned |
| **The goal** | Is to get unstuck | Is to complete a known task quickly |

A strong pattern is to use both: a designed interface for the common path, with AI available for the questions it can't anticipate. Another is to use AI invisibly, to choose, rank, or summarize content inside a normal page.

## Designing personalization

LLM-driven personalization lets a page respond to who's looking at it. Instead of one generic story for everyone, a visitor can see what matters for their business. Some things to get right:

- **Ask for little, give back a lot.** If you ask people for input, make the payoff obvious and immediate.
- **Constrain the output.** Let the model choose and adapt content within a designed structure, rather than generating an entire page from scratch. Structure keeps quality consistent and the brand intact.
- **Ground it in real content.** Personalized output should be built from accurate, approved information about the product, not the model's general knowledge.
- **Keep the default good.** Many people won't engage with personalization at all. The non-personalized experience still needs to work well.
- **Be clear about what's happening.** People should understand why they're seeing what they're seeing.

## Designing for uncertainty

Models can be slow, inconsistent, and wrong. Design for that directly:

- **Loading states.** Stream output when you can, and show progress that's honest. A blank screen for several seconds feels broken.
- **Errors and refusals.** Design what happens when the model fails, times out, or can't answer. Always offer a next step.
- **Correction.** Make it easy to edit, retry, or rephrase.
- **Sources.** Link to where information came from when you can, so people can verify it.
- **A way out.** Keep a clear path to a human, to documentation, or to the regular site.
- **Limits.** Budget tokens and requests sensibly. The personal agent on this site has a per-session limit, and the interface explains it rather than failing quietly.

## Prototype with real models

Static mocks are especially misleading for AI experiences. In a mock, every answer is the perfect length, perfectly formatted, and instant. Real output isn't.

Build working prototypes with real models and real content as early as possible. They reveal problems with length, tone, accuracy, and speed that no mock will show you. They're also far more convincing: stakeholders can try their own questions instead of trusting a happy-path screenshot. There's a practical walkthrough in [prototyping in code for designers](https://mikestilling.com/guides/prototyping-in-code-for-designers/).

**AI is a multiplier, and a thousand times nothing is still nothing.** Someone still has to judge the output and refuse to ship slop. That someone is often the designer.

## Measuring AI experiences

The usual growth metrics still apply, but AI features need a few more:

- **Task success:** did people get what they came for?
- **Quality:** are answers accurate and on-brand? Review real transcripts and outputs regularly.
- **Fallbacks:** how often do people abandon the AI path, retry, or ask for a human?
- **Downstream impact:** do people who use the AI experience go on to convert or activate at a healthy rate?
- **Cost and speed:** latency and model costs are part of the user experience and the business case.

## Frequently asked questions

### Should every AI feature be a chat interface?

No. Chat is great when people have open-ended questions, but it puts the work of knowing what to ask on the user. Often a better design uses AI behind the scenes to choose, summarize, or personalize content inside a normal interface.

### What is LLM-driven personalization?

It's using a large language model to tailor what someone sees based on what they tell you or what you know about them. For example, generating an overview of a product that focuses on the visitor's industry and goals, instead of showing everyone the same page.

### How do you design for AI that can be wrong?

Set honest expectations, show where information comes from when you can, make it easy to correct or retry, keep a clear path to a human or to non-AI content, and design the loading, empty, and error states as carefully as the happy path.

### How should designers prototype AI experiences?

With real models and real data wherever possible. Static mocks can't show how varied, slow, or surprising model output is. A working prototype reveals those problems early and is far more convincing to stakeholders.
