how to train an AI agent's tone of voice.
heysteff is an AI platform for customer support and sales across WhatsApp, Instagram, Messenger, SMS, Gmail and Shopify; Steff is the AI agent that runs it. This guide is for whoever manages the agent and wants it to sound like the brand — not generic, not robotic — before it ever talks to a real customer.
◆ prerequisites
Admin access to your heysteff workspace, brand material on hand (FAQs, policies, style guides if you have them), and, if you want to start from real history, your WhatsApp connected by QR with recent conversations between your team and customers.
Step 1: Write your brand material into context files
When you create the workspace, heysteff builds the agent's four base files from your business description: who it is, how it talks, how it works and when it escalates to a person. On top of those you add your own context files, written or pasted in the panel: shipping policy, exchanges and returns, size guide, brand rules, ways of saying things and words to avoid. Each file loads in every reply, or only when the conversation mentions certain keywords (or matches a regular expression). That way replies lean on real information about your business instead of generalities that could belong to any brand. The richer these files, the less the agent has to improvise.
Step 2: Choose the register and set creativity
In Agent → Profile you choose how the agent talks to your customers: formal, friendly or casual, each with an example of how it would reply. Changing the register rewrites the who-it-is, how-it-talks and how-it-works files, so anything you edited by hand in them is replaced; the panel warns you before it does, and it's best to choose it before editing them. Then you set creativity, from 0 to 100%: it controls how much the wording varies from one reply to the next, not how formal it is. Lower gives steadier replies; higher, more varied ones. There's no single correct combination — it depends on how your brand talks today, and the starting point is to try and adjust, not to guess it all at once.
Step 3: Start from your history and correct with "Teach the agent"
If your team already handled WhatsApp and you connect it by QR, heysteff reads recent conversations and proposes the agent's who-it-is, how-it-talks and how-it-works files, written in your team's voice. You review the proposal before using it. This only works with the QR connection: it doesn't apply to Instagram or Gmail. After that, when a reply doesn't sound right, tell the agent in "Teach the agent": "this reply was wrong, it should have been this", "always address clients formally", "never promise delivery dates". The agent proposes the changes to its context files and you decide which ones to apply. A concrete example of your brand talking carries more weight than an abstract description of personality.
Step 4: Set hard rules the agent never breaks
Beyond tone, define explicit limits: things the agent must never say or promise, such as exact delivery dates you can't guarantee, unauthorized discounts, or medical or legal claims out of place for your category. Write them into a context file that loads in every reply, and add the keywords for topics that should go to a person to the escalation file. These hard rules are different from tone — tone is how it sounds, rules are what it won't negotiate, no matter how much the customer pushes.
Step 5: Iterate in the test playground
Before the agent talks to real customers, test it in the "Test the agent" playground by simulating your typical conversations: a catalog question, a complaint, a return-policy query. Fix what doesn't sound right, test again, and repeat. Tone training isn't a one-off step — it's worth revisiting periodically as you see real conversations land in the inbox.
Tone isn't static
Once the agent is live, the tone work continues: review real conversations in the inbox, teach the agent that way of answering when the human team nails something, and adjust hard rules if a case you hadn't anticipated shows up. A brand's tone also shifts over time — a campaign, a season, a repositioning — and the agent should be able to keep up without being retrained from scratch every time. Treat tone the way you'd treat a style guide for your human team: a living document that gets tighter every time you spot a reply that could have sounded more like you.
FAQ
How long until the agent sounds right?
It varies by business. With complete brand material and a couple of rounds in the playground, most workspaces reach an acceptable tone before going live, then keep refining it with real conversations.
Can hard rules be bypassed if the customer insists?
No. Hard rules exist precisely for that — the agent doesn't break them even under pressure, and instead can escalate to a human when appropriate.
Do I need to code anything to train the tone?
No. It's all configuration from the panel: register, creativity and context files written in plain language. No code required.
◆ next step
Train Steff with your brand's tone before it talks to your customers.