Use tone and empathy responsibly

Prompts such as “be friendly” or “sound empathetic” are vague. Translate tone into behaviors:

• use the customer’s preferred name only when verified and appropriate;
• acknowledge the concrete inconvenience in one sentence;
• use plain language and short paragraphs;
• lead with the answer or next action;
• avoid blame, defensiveness, marketing language, and forced cheerfulness;
• match formality without mimicking insults, slang, or distress;
• state uncertainty directly;
• close with a realistic path forward.

These choices are reviewable. They also work across cultures and channels better than asking a model to imitate a personality.

Tone should fit the situation. A cheerful emoji may work for a low-stakes product tip and feel dismissive after a financial loss. A highly formal template may be suitable for a regulated notice and frustrating during simple troubleshooting.

A concise support response can pair recognition with action:

1. Recognize the event: “Your order shows no movement for six days.”
2. Recognize the impact: “That leaves you without the item you expected.”
3. State the action: “I’ve opened a carrier trace and will update this case by Tuesday.”
4. Set the boundary: “I cannot confirm a delivery date until the carrier responds.”

Not every reply needs all four sentences, but the pattern prevents apology from replacing service. Customers usually need a correct answer, reduced effort, and clear ownership more than a long emotional preface.

When the organization caused a verified failure, take accurate responsibility: “We sent the wrong size.” When cause is uncertain, do not invent blame: “You received a different size from the one shown on the order” acknowledges the fact while investigation continues.

AI can produce phrases that sound caring but cross boundaries:

• “I know exactly how you feel.”
• “Don’t worry.”
• “Everything will be fine.”
• “You’re part of our family.”
• “I personally guarantee this.”

These statements claim knowledge, certainty, relationship, or authority the system does not have. Replace them with specific recognition and supported expectations.

Do not infer protected traits, health status, financial condition, or emotional state from sparse language. If a customer explicitly discloses hardship, use the minimum detail needed and follow the appropriate policy. Avoid repeating sensitive disclosures unnecessarily in summaries or replies.

Customers may write in all caps, repeat themselves, or use hostile language. The model should not mirror aggression or lecture them about tone. Extract the request, answer what can be answered, and apply conduct policy consistently.

Neutral phrasing matters:

• Prefer “The email address does not match the account record” over “You entered the wrong email.”
• Prefer “The payment did not complete” over “Your card failed.”
• Prefer “To protect the account, we need to verify ownership” over “You have not proved this is yours.”

Neutral does not mean evasive. If a request cannot be granted, say so plainly, give the relevant reason, and offer permitted alternatives. Excessive cushioning can make a denial harder to understand.

Abusive or threatening messages may require an escalation or boundary response. Empathy does not require agents to tolerate harassment, and a model should not improvise safety policy.

Useful adaptation comes from explicit preferences and context: language choice, accessibility needs, channel constraints, prior troubleshooting already completed, or requested level of detail. It should not come from guesses about age, education, culture, disability, or technical ability.

For translation, preserve identifiers, policy conditions, amounts, dates, and uncertainty. Have a qualified reviewer check high-impact or nuanced messages. A fluent translation can still change legal meaning, politeness, or the strength of a promise.

Accessibility is part of tone. Use descriptive link text, explain acronyms, keep instructions sequential, and avoid relying only on color or screenshots. Offer an alternate channel when the approved process supports it.

A tone pass must not rewrite facts. Models often improve warmth by adding a cause, apology, concession, or promise. Use a preservation contract:

Improve clarity and respectful tone only.
Preserve every fact, condition, uncertainty marker, permitted remedy, owner,
and date. Do not add blame, emotional claims, compensation, guarantees,
or familiarity. Return a change log for meaning-sensitive edits.

Then review the reply with two lenses. Substance: is it correct, grounded, complete, and authorized? Experience: is it clear, proportionate, respectful, and easy to act on? Neither compensates for the other.

Collect feedback on recurring wording. If customers interpret “soon” as a promise, replace it with an approved time range. If agents routinely delete elaborate apologies, tighten the template. Style guidance should evolve from observed outcomes, not an abstract brand adjective.