One button, not a chat window — 4 UI patterns for adding AI to an existing service

When a team decides to add AI to an existing service, the first idea is usually a chat window in the corner of the screen. But a chat window hands users one more chore: figuring out what to ask. Users don't want to talk to an AI — they want to finish what they were already doing, faster.

So when you plan an AI feature, it helps to start by asking, "Does this really need to be a conversation?" Often a single button or a pre-filled input field is enough. This post covers four UI patterns commonly used to add AI to an existing service, along with the review and edit screens and failure handling each one needs. We based it on Microsoft's "18 Guidelines for Human-AI Interaction" and Google PAIR's "People + AI Guidebook."

Pattern 1: Summary button

Put a "Summarize" button next to a long post or a long list. The AI only runs when someone clicks, so call costs are easy to estimate, and users see results only when they want them.

• Good fit: long posts, collections of product reviews, customer inquiry history
• Review screen: add "View original" under the summary so users can compare right away
• Failure handling: if summarizing fails, show the original as is, with a retry button

Pattern 2: Draft fill

The AI fills an empty input field with a draft first, and the user edits and saves it. For example, an online store's admin screen could have a "Generate draft" button next to the product description field (example).

The key is that the AI never saves on its own. Label the draft "AI draft" to separate it from human-written text, and let a person press the save button. This is exactly what Microsoft's guideline "Support efficient correction" is about.

Pattern 3: Auto-classification and tag suggestions

The AI pre-assigns categories and tags to incoming inquiries or posts. Users don't have to click anything, so it feels so natural they may not even notice a feature was added.

• Good fit: classifying customer inquiries, tagging community posts, tidying attachment file names (example)
• Review screen: a small "AI suggested" label next to the category, plus a dropdown to change it in one step
• Failure handling: when confidence is low, leave it blank and mark it "Needs classification"

That last point matters. Instead of forcing a choice when things are unclear, hand it to a person — the same idea as the guideline "Scope services when in doubt."

Pattern 4: Suggested text

While someone is writing, show the next sentence or a reply suggestion in faint text. Press Tab to accept, or just keep typing.

It's convenient, but it can easily get in the way. So make ignoring it cost nothing, and let users turn it off in one place in settings. In guideline terms, that's "Support efficient dismissal" and "Provide global controls."

The four patterns at a glance

Three things to check for any pattern

• Have you marked which parts the AI created? (separate from human-written content)
• Can users fix or undo it in one step?
• When it fails, does the original feature keep working instead of showing a blank screen?

Google PAIR's guidebook likewise defines errors from the user's point of view and recommends always leaving users a path to what they can do next. Rather than trying to eliminate errors entirely, it's more realistic to make sure work doesn't stop when an error happens.

Summary

Adding AI doesn't have to mean adding a chat window. Fit one of these patterns — summary button, draft fill, auto-classification, or suggested text — into the flow of what users already do, and take care of labeling, editing, and failure handling. Even first-time users will be able to use it without friction.

References

• Microsoft HAX Toolkit, Guidelines for Human-AI Interaction: https://www.microsoft.com/en-us/haxtoolkit/library/pattern/
• Microsoft Research, Guidelines for Human-AI Interaction (CHI 2019): https://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/
• Google PAIR, People + AI Guidebook — Errors + Graceful Failure: https://pair.withgoogle.com/chapter/errors-failing/

SOL LABS AI

We build web apps, mobile apps and business systems, then add and run the AI they need. We build and operate our own meeting notes, contract and company analysis AI.

See our projects