DEV Community

john parker
john parker

Posted on Fully Autonomous

If Your AI Feature Loses Internet Access, What Can Users Still Do?

You’re editing a document with an AI writing tool. You highlight a paragraph, ask for a rewrite, and keep typing while it works. Then the connection drops. When it comes back, the original paragraph is still there, but the status just says Something went wrong.

Now you have three questions. Were your latest edits saved? Is the rewrite still coming? Can you keep working without it?

Those are product questions. The answers depend on what your app does when it loses its connection to the model.

Keep the user’s work available

In a writing app, the document and the AI suggestion have different jobs. The document holds the user’s work. The suggestion is a proposed change they can accept or ignore.

Design the failure around that difference. If the rewrite stops, leave the original paragraph and any newer edits intact. Keep manual editing available wherever the app supports local editing. If the app saved the draft on the device but hasn’t synced it, say “Saved on this device. Sync pending.” Each part tells the user something they can act on.

Decide what happens to the unfinished request

Suppose the user asks the AI to rewrite a paragraph, goes offline, and then edits that same paragraph by hand. If the original request runs automatically when the connection returns, it may produce a suggestion for text that no longer exists.

The right recovery depends on the task.

What happened What the app can offer
A quick rewrite failed and the selected text is unchanged. Show a retry button beside the selection.
The user edited the passage while waiting. Let them request a rewrite of the current text. If an older result arrives, label it as based on the previous version.
The app already accepted a longer job before the connection dropped. Check its status on reconnect. Show the result if it finished, or let the user decide whether to start over.
The model is unavailable, but the editor still works. Keep editing available and show which AI action is temporarily unavailable.

That last row can make the biggest difference. A user trying to finish a document may prefer to write the sentence manually and move on.

Run the interruption test before you ship

Take one workflow and disconnect at three points: before the request reaches the server, while a response is streaming, and after the server finishes but before the browser receives the result. In each case, change the document before reconnecting.

Then try to answer these questions from the screen alone:

  • Can I tell which edits were saved and where?
  • Can I continue the task manually?
  • Will retry use the text I’m looking at now?
  • If an old result arrives, can it overwrite anything I’ve changed?

A useful fallback lets people finish the work they came to do and tells them what happened to the AI request.

What’s one thing your AI feature should still let users do when the model is unreachable?

By the International AI Certification Board (IAICB).

Top comments (0)