How to Keep AI Drafts From Repeating Fixes Customers Tried

When an AI draft tells a customer to "try restarting the app" after they wrote that they already restarted it twice, the customer hears one thing: you didn't read their message.

To stop this, treat "what the customer already tried" as its own input. Pull it out of the thread before the AI writes anything. Give it to the model as a separate, clearly labeled list with a direct rule against repeating those steps. Then check the draft against that list before you send it. The rest of this post covers how to do each step.

Why this matters to customers

Having to repeat yourself is one of the most common complaints in support research.

• In Zendesk's CX Trends 2026 report, 74% of consumers said it is frustrating to tell their story over and over to different agents.
• Gartner's customer effort research names repeating information and repeat contacts as drivers of high effort. A Gartner survey of 1,492 B2B and B2C customers found that 62% of customer service channel transitions were "high effort", as CX Today reported in July 2023.

Neither study looks specifically at AI drafts suggesting steps the customer already tried. The problem is closely related, though. A reply that repeats a failed fix makes the customer explain again and adds another round of back-and-forth.

Why AI drafts repeat fixes

The points below are analysis, not documented research findings. They're common patterns in how drafting systems are built.

• Tried steps are buried in the message. Customers often mention them in passing ("I already logged out and back in, cleared the cache, nothing") between the problem description and an apology. A model told to "answer this question" may give them little weight.
• Knowledge base articles start with the basics. Most troubleshooting docs begin with restart, update, and reinstall. If the model pulls in that article, those steps come along with it.
• The model only sees part of the thread. If the system sends just the newest message, anything tried in earlier emails is gone.
• Customers name steps differently than your docs. "Force-quit" and "restart the app" may be the same step to you but not to a model matching text.
• "I tried everything" is unclear. Without specifics, the model falls back to the standard list.

Step 1: Extract the "already tried" list before drafting

Before the model writes the reply, run a separate step that reads the full conversation and returns a short structured list:

• Steps the customer says they tried
• What happened each time (no change, error, partial fix)
• Environment details they gave (OS, app version, device, plan)
• Vague claims such as "tried everything" that need follow-up

Keeping extraction separate from drafting has two benefits. The list is easy to check by eye, and the drafting step gets a clean list instead of having to find the steps itself.

For small teams this can be one extra prompt call, or a quick note you write yourself on tricky tickets.

Step 2: Put the list in the prompt as its own labeled block

Model providers recommend structuring prompts so that different kinds of content are clearly separated. Anthropic's prompting guidance, for example, suggests using XML tags to separate instructions, context, and documents. A drafting prompt might look like this (illustrative template):

<already_tried>
- Restarted the app (no change)
- Reinstalled from the App Store (no change)
- Logged out and back in (same error: "Sync failed 403")
</already_tried>

<instructions>
Do not suggest any step listed in <already_tried>, including reworded
versions of the same step. Briefly acknowledge what the customer tried.
Start troubleshooting from the next step after those.
If a listed step truly must be repeated, explain the specific reason.
If nothing new is left to suggest, say so and ask for the details
needed to investigate further.
</instructions>

Some details make the rule work better:

• Cover rewordings. Telling the model that reworded versions count helps with the naming problem.
• Ask for a short acknowledgment. One line such as "Thanks for already trying a reinstall and a fresh login" shows the customer you read their message. It also makes it easy to see during review whether the model used the list.
• Give a fallback. If the model has no allowed step to suggest, it needs another option, such as asking for logs or escalating. Otherwise it may suggest a forbidden step anyway.

Step 3: Send the whole thread, not just the latest message

If a customer says "still broken" in their third email, the useful context is in the first two. Give the drafting step the full conversation, or at least a running summary that includes the "already tried" list. Update that list each time the customer reports a new attempt.

Step 4: Write troubleshooting docs as numbered levels

Analysis and recommendation: AI drafts improve when your knowledge base is easy to match against.

• Give each fix a short, stable name and list common customer wordings next to it ("Restart app — also: force-quit, close and reopen, kill the app").
• Order fixes from basic to advanced so the model can "start from the next step" after what's been tried.
• Write down when a step is worth repeating. For example, a reinstall may only help after a specific update. That gives the model a real reason when repetition is justified.

Step 5: Ask specific questions when the customer says "I tried everything"

Don't let the draft guess. Tell the model that when tried steps are unclear, it should ask a short, specific question instead of listing basics. For example: "To avoid sending steps you've already done, could you tell me whether you've tried signing out on all devices, or only this one?" Keep it to one or two questions so the customer doesn't feel quizzed.

Step 6: Explain any repeat that is truly needed

Sometimes a step really does need to be repeated, such as a restart after a config change you just made on the server. That's fine if the draft says why. "I know you already restarted. We've just reset your sync token on our side, so one more restart should pick up the change" reads very differently from a bare "Please restart the app."

Step 7: Check every draft against the list before sending

Even with a good prompt, a human check catches what the model misses. A quick checklist:

1. Does the draft suggest anything on the "already tried" list, in any wording?
2. Does it acknowledge what the customer tried?
3. Does the first new step actually come after what they tried?
4. If something is repeated, is there a stated reason?
5. Does the draft ask for missing details instead of guessing?

This check takes seconds, and it's one reason a human-in-the-loop setup works well for support. SupportMe, the AI support assistant behind this blog, is built this way. It drafts replies, and nothing is sent without your approval.

Step 8: Feed your corrections back into the system

When you delete a repeated step from a draft, that edit tells you something. Either the prompt, the extraction step, or the knowledge base needs fixing. Look for patterns. If the same fix keeps slipping through, add its other names to your docs or make the rule more specific.

Some tools do part of this for you. According to SupportMe's product description, it compares its draft with your final edited reply and uses the difference to update your writing style profile and knowledge base. The tool is in pre-launch. Whatever tool you use, check what a tool actually learns from your edits rather than assuming it will stop a specific mistake.

A hypothetical before-and-after

This exchange is illustrative, not a real customer case.

Customer: "Exports keep failing with 'timeout.' I've updated to the latest version, restarted my Mac, and tried exporting a smaller file. Same error every time."

Weak draft: "Sorry about that! Please make sure you're on the latest version and try restarting your computer. If that doesn't work, try a smaller file."

Better draft: "Thanks for already updating, restarting, and testing a smaller file. That rules out the usual causes. Next, could you check whether exports work with your VPN turned off? Some VPNs cut long uploads. If it still fails, please send the log from Help → Export Logs and I'll look at it directly."

The better draft names what was tried, moves to the next step, and asks for something specific.

Conclusion

AI drafts repeat fixes mainly because the steps the customer already tried aren't treated as a separate, required input. Pull that list out of the full thread, label it in the prompt with a clear rule and a fallback, write troubleshooting docs as numbered levels, and check each draft against the list before sending. Research from Zendesk and Gartner shows customers dislike repeating themselves, and these steps help your replies avoid making them do it.

References

• Zendesk, CX Trends 2026
• CX Today, 62% of Customer Service Channel Shifts Are "High Effort", Finds Gartner (July 2023)
• Anthropic, Prompting best practices
• SupportMe, product website