Voice and tone
Teach the AI how your team writes. Sample replies beat adjectives, every time.
Voice and tone
The AI writes in your team's voice when you give it your team's writing. Adjectives like "friendly" and "professional" do almost nothing on their own. Twenty real, edited replies do almost everything.
How voice is set
Three inputs, in order of impact:
- Sample replies. Paste 20 to 50 real replies you would be proud of. This is the single biggest lever.
- A short style note. A few sentences in plain English about how you sound.
- Tone preset. A starting point: warm, neutral, terse, playful.
Configure all three on AI → Behavior.
Sample replies are the lever
The AI imitates patterns in the samples: greeting style, sign-off, sentence length, when to use bullet points, when to apologize. If your samples never start with "Hi there!", the AI will not either. If your samples sign off with first name only, the AI does that.
A few rules for picking samples:
- Use real replies from your inbox, not idealized prose.
- Edit out anything customer-specific (names, account IDs).
- Cover a range: short FAQ answers, longer explanations, a refund decline, a how-to walkthrough.
- Skip outliers. The angry-Monday reply you regret is not a sample.
20 is the floor. 50 is great. Past that, returns diminish.
The style note
A short, direct paragraph. Not a brand bible.
Example that works:
> We are casual but not cute. No exclamation points unless something is genuinely good news. We say "yes" or "no" before we explain. We apologize once, not three times.
Example that does not work:
> We delight customers with thoughtful, personalized communication that reflects our brand values.
The first is concrete. The second is wallpaper.
Tone presets
Pick one as a starting frame:
- Warm. Friendly, contractions, "happy to help".
- Neutral. Direct, no extra ceremony, the default.
- Terse. Few words, no filler. Best for power-user channels.
- Playful. Light humor, occasional informality. Use carefully.
Presets are weak compared to samples. Treat them as a tilt, not a voice.
Per-channel voice
You can override voice per channel. Many teams use:
- Email: warm.
- Chat widget: neutral.
- Slack Connect: terse.
Per-channel sample sets are supported. Replies on the chat widget will copy chat samples, not email samples.
Forbidden phrases
Set a small list of words and phrases the AI will not use. Common entries:
- "I understand your frustration."
- "Per my last email."
- "Unfortunately,".
- Em-dashes.
- Marketing language about your own product.
Keep this list short. Long lists make drafts read like they are dodging.
Required phrases
Use sparingly. A signature line, a legal disclaimer on certain topics, a specific link to a status page. The AI will include them when relevant.
How edits feed voice
When an agent edits a draft before sending, the edit is captured. Patterns across many edits feed back into Quality assurance review and the Learned Q&A library. If your team consistently rewrites "Hello," to "Hey," the AI learns it.
You do not have to manually re-train. Voice updates over time.
Testing voice
Before turning anything on, the Playground lets you paste any past customer message and see what the AI would draft with your current voice settings. Iterate until five sample tickets all read like your team.
What voice does not control
- Length. See Reply length.
- Whether the AI replies at all. See Guardrails and bypass labels.
- Facts. Voice changes how things are said, not what is said. Facts come from The brain: KB graph.
Recommended setup
- Tone preset: neutral.
- 30 sample replies pulled from your last month of email.
- Three-sentence style note.
- Five forbidden phrases your team hates.
- No required phrases.
Update samples every quarter. Voice drifts as your team grows.
Related
Was this article helpful?