What shapes how your AI agent sounds
The default instructions already set a tone: “excellent, friendly and efficient replies”, ending “with a positive note”. Replace that line with your own description. See Instructions.
Write a tone section
Vague words like “professional” or “fun” leave the AI agent guessing. Describe behavior it can check itself against:- Who it is: a name and role, if you want one. “You are Sky, the assistant for Acme Airlines.”
- Register: formal or casual, first names or not, “you” or a formal form in languages that have one.
- Length: “Keep answers under four sentences unless the customer asks for detail.”
- Format: lists or paragraphs, bold text, emoji or none.
- Words: phrases to use, and phrases to avoid, such as jargon or apologies that repeat.
- Hard moments: how to sound when the customer is angry, or when the answer is no.
Examples by brand
Each example is a tone section to add to your General instructions. Replace the company details with your own.- Friendly retail
- Formal financial
- Technical SaaS
- Playful consumer
Example
Model and temperature
Temperature changes wording, not knowledge. At 0, the default, the AI agent gives much the same answer each time. Closer to 1, the wording varies between conversations. Keep it low for support, billing, and policy questions where every customer should get the same answer. Raise it if replies feel repetitive in conversations where variety helps, such as product recommendations. See How temperature changes answers. Models differ in how they sound and how closely they follow tone rules. “Does its tone suit your customers and your brand?” is one of the questions in How to choose a model. Compare a few on your own questions before switching.Keep the tone consistent across channels
Your global instructions apply to every channel by default, so one tone section covers the Chat Bubble, WhatsApp, email, and the rest. A few places can drift:- Channel-specific instructions. Turning off Sync with global instructions on a channel’s Overview tab replaces the global instructions for that channel. Copy your tone section into it, then add the channel’s differences, such as shorter replies on WhatsApp.
- Phone. Turning off Sync voice instructions with base in the Phone channel’s Phone AI settings gives calls their own instructions. Keep the tone section there too.
- Voice delivery. In Live voice, Speaking style (up to 500 characters) describes how the voice sounds: pace, warmth, formality. In Classic voice, voices that support it take Voice instructions. Describe the sound there and keep rules about what to say in your instructions. See Live Voice.
- Q&As. A matching Q&A is sent word for word, so write it in the same voice as your instructions.
- Names. The Chat Bubble’s Display name and Agent picture also appear in Slack. Email uses its own Email Display Name. See Email Display Name.
- Your human team. In the Helpdesk, AI Compose can make a teammate’s draft More friendly or More formal before sending. See AI Compose.
Test the tone
1
Compare versions side by side
On Build > Instructions, click Compare. Give two chats different tone sections or temperatures, then send the same customer message to both. See Compare area.
2
Try the hard cases in the Playground
Send an angry message, a request you have to refuse, and an off-topic question. Tone rules break most often in these.
3
Add a Test
Create a Test such as: simulated user “A customer whose order #1042 arrived broken. Angry, uses capital letters.” Success criterion: “The agent stays calm and polite, doesn’t use emoji or jokes, and offers a replacement or refund path.”
4
Review real conversations
After you publish, read conversations in Activity and tighten any rule the AI agent isn’t following.
Troubleshooting
The AI agent ignores my tone rules
The AI agent ignores my tone rules
Make the rule specific and checkable: “No more than one emoji per reply” works better than “don’t overuse emoji”. Move rules that must always hold to the Guardrails section, remove instructions that contradict each other, and compare a different model in Compare.
Replies sound different on one channel
Replies sound different on one channel
Check that channel’s Overview tab. If Sync with global instructions is off, the channel uses its own instructions and needs its own copy of your tone section.
The AI agent sounds too robotic or repetitive
The AI agent sounds too robotic or repetitive
Raise the temperature a little, give one or two sample phrasings in your instructions, and check that Q&A answers aren’t written in a stiffer voice than the rest.
Frequently asked questions
Is there a setting for my AI agent's tone?
Is there a setting for my AI agent's tone?
No. Tone comes from what you write in Build > Instructions, helped by the model and temperature you choose. Voice conversations add Speaking style or Voice instructions for how the voice sounds.
Can my AI agent have a name and persona?
Can my AI agent have a name and persona?
Yes. Give it a name and role in your instructions, and set the same Display name and Agent picture on the Chat Bubble’s Display tab. To keep users from talking it into a different persona, add a rule such as “Don’t adopt any other persona” to the Guardrails section.
Can each channel have a different tone?
Can each channel have a different tone?
Yes. Turn off Sync with global instructions on a channel’s Overview tab and write instructions for that channel. They replace the global instructions there, so include everything that channel needs.
Does temperature change the AI agent's personality?
Does temperature change the AI agent's personality?
It changes how varied the wording is, not what the AI agent knows or the rules it follows. Personality comes from your instructions. Keep temperature low when answers must be consistent.
How do I stop the AI agent from using emoji?
How do I stop the AI agent from using emoji?
Add “Don’t use emoji” to the Guardrails section of Build > Instructions, then test it in the Playground.
Related
Instructions
The full reference on instructions, Guardrails, model, and temperature.
Best Practices
Worked examples of good and bad instructions.
Multilingual support
Control which language your AI agent replies in.
AI disclosure
Tell customers they’re talking to an AI agent.