Build a small, varied conversation set.
Words such as “friendly”, “professional” and “warm” are open to interpretation. Real messages show sentence length, punctuation, vocabulary, humour and how directly your team answers a question. Start with approved conversations that cover different kinds of work:
- a routine question with a short factual answer;
- a sales enquiry where the customer is comparing options;
- a complaint that was acknowledged without becoming defensive;
- a firm boundary, such as a request for an unapproved discount; and
- a complicated question that needed a human decision.
Remove personal details before using the examples, but keep the order and rhythm of the exchange. A polished marketing paragraph is not a substitute for the way the business actually speaks in WhatsApp or email.
Annotate the reason behind each reply.
Beside every example, add a short note explaining the decision. “One sentence because the customer asked a yes-or-no question” is useful. “Sounds like us” is not. Mark the greeting, level of formality, length, preferred words and the point at which the conversation should move to a person.
Also include negative examples. Show wording that feels pushy, vague or too casual, then state the correction. This gives the agent a clear boundary instead of asking it to guess what “too much” means.
Keep voice separate from business rules.
Tone changes how a fact is expressed; it must not change the fact. Approved prices, links, refund limits and promises need separate, testable controls. Configure the agent to stop when a price, link or promise cannot be verified, and send the case to a named person. Do not treat a prompt instruction as proof that the control works.
Define a clear order of priority: accuracy and safety first, the customer’s immediate need second, and brand style third. That order prevents a tone instruction such as “always be helpful” from turning into an invented promise. Kai currently works through WhatsApp and email, replies in the customer’s language, and identifies itself as AI when asked; the same voice guide should remain recognisable in both channels.
Test the voice on conversations it has not seen.
Use a separate set of past conversations for testing. Check whether the agent answers the real question, preserves approved facts, keeps the expected level of warmth and hands off sensitive cases at the right moment. Include short requests, voice-note transcripts, unclear questions and customers who change topic halfway through.
Keep a fixed, versioned set of realistic test cases. Record each failure, adjust one rule at a time, and repeat the same set after every change. Do not approve a voice because one hand-picked example sounds good, and do not claim a replay gate unless the deployed system has been verified to run it.
Maintain the guide as the business changes.
Give the guide an owner and review it when products, policies or the team’s writing style change. Add a newly approved example when it teaches a genuinely new pattern; remove old examples when their facts are no longer current. The result should stay short enough that a person can review it.
Next, define customer escalation rules and safe price-list rules. See Kai’s current capabilities, operating profile and available roles for the controls behind the customer-facing voice.