AI-Assisted Support

A Review Checklist for AI Replies That Reference Your UI

A practical checklist for reviewing AI-drafted support replies that mention buttons, menus, and settings, so customers get labels, paths, and platform details that match what they actually see.

SupportMe••9 min read

When an AI-drafted support reply says "Go to Settings > Billing and select Cancel plan," every part of that sentence is a claim about your product. Before you send it, check four things. Does each element exist? Does each label match the screen word for word? Is the path complete and in the right order? Does the customer's platform, plan, and role actually show that screen?

The rest of this post turns those checks into a checklist you can run in a minute or two. It also explains why UI references are a common place for AI drafts to go wrong.

Why UI references need their own review step

Wrong UI references come from a few documented causes:

  • Language models guess when they lack specifics. OpenAI's research on hallucinations says that standard training and evaluation "reward guessing over acknowledging uncertainty," and that "arbitrary low-frequency facts... cannot be predicted from patterns alone" (OpenAI). The exact label on your third settings tab is that kind of fact. If the model doesn't have it, it may write something plausible, like Account settings when your app says Preferences.
  • Your UI changes faster than your docs. Small teams rename buttons, move settings, and ship redesigns without updating every help article. A draft built from older replies or docs can describe a screen that no longer exists.
  • The same product looks different to different users. Even Google's own Android help warns readers on custom Android versions to check with their device manufacturer if the steps don't work (Google Android Help). Your product probably differs too, by platform, plan, permission level, or feature flag.

The analysis that follows is ours. A wrong label rarely looks like a mistake when you read the draft, because it sounds right. The customer finds it when they can't find the button, and then they write back.

The checklist

1. Every named element exists in the current version

  • Open the screen the reply describes, or check a current screenshot or label list.
  • Confirm that each button, tab, menu, toggle, and field in the draft is really there.
  • Watch for invented "obvious" features, like an Export button on a screen that has none.

If you can't verify an element quickly, rewrite the step around the task ("open your billing settings") instead of a label you're unsure of.

2. Labels match the screen exactly

Style guides agree that UI references should match what users see. Google's developer style guide says to reproduce the capitalization shown on screen, with one exception: all-caps labels go in sentence case. It also says to drop trailing ellipses, so you write "Click Browse," not "Browse..." (Google developer documentation style guide). Microsoft's guide says the same about ellipses and trailing colons (Microsoft Writing Style Guide).

When reviewing, check:

  • Spelling and wording: Sign in vs. Log in, Workspace vs. Team
  • Singular vs. plural: Integration vs. Integrations
  • Product jargon the AI may have swapped for a generic word

If your product ships in more than one language, the label the customer sees may not be the English one. The multilingual AI reply checklist covers that case.

3. The navigation path is complete and in order

  • Walk through the path yourself, one step at a time.
  • Look for skipped steps, like a profile menu you have to open before Settings appears.
  • Use the > shorthand only for simple menu sequences. Google's guide limits angle brackets to menu items, not chains of mixed elements (Google). Microsoft's guide says to use them only when the path is clear and every step uses the same action (Microsoft).

If the path has more than three or four steps, a numbered list is usually clearer than one long line. The setup steps review checklist goes deeper on multi-step instructions.

4. The reply fits this customer's platform, plan, role, and version

This is the check most likely to need context from outside the draft:

  • Platform: Is the customer on web, iOS, Android, or desktop? Do the steps match that platform?
  • Plan: Is the feature limited to a paid tier?
  • Role: Can this user see the setting, or only an admin or account owner?
  • Version: Does the customer's app version have this screen? If the reply names a version, see the version number review checklist.
  • Feature flags and betas: Is the element only shown to some users?

If you don't know the customer's platform or version, say which one the steps are for, or ask.

5. Verbs fit the device

Microsoft recommends input-neutral verbs like "select," "open," and "go to" instead of "click" or "swipe," because people use mouse, touch, keyboard, and voice (Microsoft). Google's guide lists both "click" and "tap" as acceptable (Google).

The practical rule is simple. Don't tell someone on a phone to "right-click," and don't tell a desktop user to "swipe." If you don't know the device, use "select."

6. No directions based only on position or looks

Google's guide advises against "above," "below," and "right-hand side," because they hurt accessibility and localization. It gives "In the left-side panel, click the button with three lines" as an example to avoid (Google). Layouts also change on smaller screens, so "top right" on desktop may be a menu on mobile.

Replace position hints with the element's name, the section it's in, or its tooltip text.

7. Icon-only buttons are named by their tooltip

For icons without visible text, Google suggests using the tooltip name. If there's no tooltip, check the element's aria-label or title attribute (Google). AI drafts often describe icons loosely ("the gear icon," "the hamburger menu"). Check that the description matches your actual icon and that the name matches its tooltip.

8. Formatting works in the channel

Docs often use bold for UI labels. Microsoft notes that in general content outside docs, you can use bold, quotation marks, or wording that sets the label apart. Whichever you pick, use it the same way throughout (Microsoft).

For support email, check how the reply will render. Markdown asterisks can show up as literal ** in plain-text clients, and app store review responses generally don't support rich formatting. Pick one style that survives the channel and use it for every label in the reply.

9. The reply says what should happen next

A good UI step tells the customer how to know it worked: "You'll see a confirmation banner," or "The status changes to Active." Check that the expected result is accurate, too. AI drafts sometimes describe a confirmation dialog or email that your product doesn't send.

10. Links and screenshots match the text

If the draft links to a help article or settings page, open it and confirm it shows the same labels the reply uses. The link review checklist for AI support replies covers broken and mismatched links in more detail. If you attach a screenshot, make sure it comes from the current UI and the right platform.

11. UI detail is only there when it helps

Both Google and Microsoft suggest describing the task first and naming UI elements only when the reader needs them (Google, Microsoft). "Refresh the page" is better than a step-by-step guide to the refresh button. Fewer UI claims means fewer things that can be wrong.

A hypothetical before-and-after

This exchange is made up to show the checklist in use.

A customer on the iOS app asks how to turn off weekly summary emails.

AI draft:

Click the gear icon in the top right, go to Account Settings > Notifications, and uncheck "Weekly digest."

Problems a reviewer would catch:

  • "Click" and "top right" assume a desktop layout. On iOS, settings sit behind the profile tab.
  • The screen is called Settings, not Account Settings.
  • The toggle is labeled Weekly summary, not "Weekly digest."
  • It's a toggle, not a checkbox.

Reviewed reply:

On iOS, open the Profile tab, then go to Settings > Notifications and turn off Weekly summary. You'll stop getting the email starting next week.

The fix took a few seconds once someone checked the real screen. Without that check, the customer would have looked for a gear icon that isn't there.

Make the checklist faster over time

Some habits cut review time. These are recommendations, not documented results:

  • Keep a short label list. A plain document listing your main screens, their exact labels, and which plans or platforms show them. It gives you, and any AI tool that uses your knowledge base, one source of truth.
  • Update it when you ship UI changes. Renames are the most common source of outdated drafts.
  • Fix the source, not just the draft. If the same wrong label keeps appearing, the old wording is probably still in a help article or saved reply somewhere.

If you use SupportMe, your corrections feed back automatically. It compares its draft with the version you send and updates your style profile and knowledge base from the difference. Nothing goes out without your approval, so the review step stays with you. If you want to turn this list into your own version, building a support reply checklist in 10 minutes walks through the process.

Conclusion

A UI reference is a factual claim about your product. Language models fill gaps with plausible guesses, and your interface keeps changing, so those claims need a quick check before you send. Confirm that each element exists, match labels exactly, walk the path, and check the customer's platform, plan, role, and version. Then tidy up the verbs, position hints, formatting, and expected results. Most of these checks take seconds once you have the real screen in front of you.

References

Tags

AI support repliesUI referencesreview checklistbutton labelsmenu pathscustomer supportindie developerssupport reply accuracy

Related posts