AI Training

Training examples and the phrase library that shape how replies sound.

The AI Training page lets you teach the AI how your company communicates, so drafted replies sound like they came from you — not from a generic template. Open Settings > Training to manage training examples and your phrase library.

Training examples

Training examples are past response letters that the AI references when drafting new replies. The more examples you provide, the more the AI sounds like your team.

Three sources of examples

Uploaded — paste the full text of a response letter you’ve written in the past. Include the greeting, body, and signoff. Optionally describe the situation — the system auto-extracts a summary if you skip it. Processing takes a few seconds.
From sent replies — every reply you send through Callback HQ is automatically saved as a training example. No extra work needed.
Gold (canonical) — flag any example as “canonical” to give it extra weight in retrieval. Use this for replies you consider ideal. The AI leans on canonical examples most heavily when drafting.

Automatic classification

Each example is automatically classified by claim type — water/plumbing, roof/exterior, HVAC, and so on — so the AI can pull the most relevant examples when drafting a reply for a specific category of complaint.

Managing examples

Expand any example to view the full response text and situation summary.
Delete examples you no longer want influencing drafts.
Toggle the canonical flag on or off for any example at any time.

Start with 5 – 10 strong examples from different claim types. This gives the AI enough variety to match your tone across different scenarios.

Phrase library

The phrase library has two lists that shape the AI’s word choices. Both are loaded into the draft prompt context, so the AI knows your company’s voice boundaries before it writes a single word.

Preferred phrases

Language you want the AI to use. Examples:

“We appreciate you reaching out”
“Our inspector noted…”
“Thank you for bringing this to our attention”

You can add a note explaining why a phrase is preferred, which helps the AI understand the intent behind it.

Phrases to avoid

Language you never want in a reply. The AI actively avoids these. Examples:

“This is not our problem”
“We are not liable”
“You should have…”

How training data is used

When the AI drafts a reply on the claim review page, it pulls the most relevant training examples by claim type and applies your phrase library rules. The result is a draft that matches your company’s tone and avoids language you’ve flagged.

Training data also works alongside your response policy, which controls the substance of replies (what to offer). Training controls the style (how it sounds).