Brand voice has become a standard AI feature. Jasper can infer a voice from uploaded writing samples and apply it to agents and chat, while Typeface describes brand kits that enforce language and style rules across generated content. The promise is consistency at scale. The risk is scaling the same polite, polished voice every other company uses.
A July critique of AI brand-voice tools identified the weak input behind that sameness: lists such as “professional, trustworthy, innovative, and customer-focused” describe nearly every company. A support chatbot needs a more executable document. Use this six-part voice contract before editing the prompt:
| Contract field | Question it must answer |
|---|---|
| Belief | What does the business believe that shapes this answer? |
| Decision | What real policy or tradeoff proves that belief? |
| Language | Which sentence patterns, terms, and level of detail fit? |
| Avoid | Which phrases, claims, and habits make the voice generic? |
| Pressure behavior | How should the voice change when a visitor is confused or angry? |
| Example pair | What does an approved and rejected answer look like? |
The contract is short enough to maintain and specific enough to test. It also keeps brand voice tied to customer help instead of turning the chatbot into a mascot.
Start With Decisions, Not Adjectives
“Friendly and professional” gives a model a mood, but no point of view. Thousands of prompts use the same words. The model fills the gap with its statistical idea of business-friendly writing: a warm opening, several soft qualifiers, and a cheerful closing question.
A decision gives the voice a reason to sound different. Consider an outdoor-equipment company that repairs products before recommending replacements. That operating decision can shape customer replies:
- Explain the repair path before presenting a new product.
- State what can be fixed at home and what needs inspection.
- Avoid artificial urgency and upgrade language.
- Recommend the less expensive option when it solves the problem.
Those rules are recognizable because a competitor might genuinely disagree with them. They also help the customer complete a task. “Adventurous, authentic, and bold” does neither.
Collect three to five decisions from support, sales, product, and leadership. Ask what the company repeatedly chooses when speed, revenue, simplicity, and customer interest conflict. Good inputs include a generous return rule, a refusal to hide limitations, a preference for repair, or a commitment to show prices before collecting contact details.
Do not invent beliefs for the prompt. If the company says “we value transparency” while hiding required fees until checkout, the chatbot will expose the contradiction faster than a style guide can conceal it.
Give the Voice a Job
A website chatbot is not writing a campaign. It is helping a person who may be hurried, uncertain, comparing vendors, or already frustrated. Its voice must serve that situation.
Define the job in observable behaviors:
Answer order. Decide whether the bot gives the direct answer, the evidence, or the next step first. For most support questions, direct answer first is the useful default.
Sentence shape. Specify short or long sentences, contractions, first-person language, and how much explanation follows a simple answer. “Keep simple answers under 60 words” is testable; “be concise” is open to interpretation.
Vocabulary. List the terms customers use, the terms the business uses, and any jargon the bot should translate. This is where “subscription,” “membership,” and “plan” stop drifting between replies.
Confidence. State when the bot can speak directly, when it should name a source, and when it must say that information is missing. Voice should never make an unsupported answer sound more certain.
Next step. Decide how the bot offers help without adding “Is there anything else?” to every message. A useful next step should follow from the question: open the setup guide, compare two plans, collect a support request, or hand off the conversation.
The broader chatbot prompt-engineering guide covers role, scope, formatting, knowledge boundaries, and escalation. The voice contract sits inside that structure. It should not replace the rules that keep the chatbot accurate and useful.
Write a Voice Contract the Model Can Execute
Here is a complete contract for the fictional outdoor brand Northline. It is intentionally compact enough to paste into a custom instruction field.
VOICE CONTRACT — NORTHLINE SUPPORT Belief: Durable gear should be maintained and repaired before it is replaced. Decisions that prove it: - Explain cleaning, repair, and warranty options before suggesting a purchase. - Recommend the lowest-cost option that solves the customer's stated need. - State product limitations plainly; do not turn them into vague benefits. Language: - Give the direct answer in the first sentence. - Use short paragraphs, concrete verbs, and the customer's product name. - Use "we" only for an action or policy Northline actually owns. - Add one relevant next step, not a generic offer of more help. Avoid: - "Great question," "Absolutely," "premium," "perfect for everyone," artificial urgency, and repeated apologies. - Adventure metaphors when the customer is troubleshooting or upset. Under pressure: Name the specific problem, state what can happen next, and do not become more cheerful as the customer's frustration rises. Approved pattern: Direct answer → evidence or limit → one useful next step.
Notice what is absent: a fictional personality biography, a dozen adjectives, and instructions to sound human. The contract describes choices a reviewer can see in the final reply.
Keep facts outside the voice contract. Prices, warranty periods, product specifications, and operating hours belong in maintained sources or controlled Q&A pairs. Otherwise a style edit can accidentally change business truth.
Worked Rewrite: The Same Policy in a Specific Voice
Suppose a customer asks:
My two-year-old rain jacket is leaking at the shoulders. Do I need a new one?
A generic “friendly” chatbot might answer:
Absolutely! I’m sorry to hear about your jacket. Northline offers a wide range of premium waterproof jackets that could be perfect for your next adventure. Would you like help finding a replacement?
The reply is upbeat, but it ignores the company’s repair-first decision and pushes a purchase before diagnosing the problem.
Using the contract, the answer becomes:
You may not need a new jacket. Leaking at the shoulders often comes from worn water-repellent treatment or seam damage. Clean the jacket with technical wash and renew the treatment first; if the inside seam tape is lifting, send us two photos for a repair assessment. We’ll recommend replacement only if the shell cannot be repaired.
The second answer has a recognizable voice because it makes a recognizable choice. It answers first, uses concrete language, reflects an operating policy, and ends with one relevant action. Removing the logo would still leave clues about the company behind it.
Now change the emotional context:
I already followed your care guide twice. It still leaks, and I’m tired of being sent in circles.
The voice should adapt without disappearing:
You already tried the step we would normally recommend, so repeating it would waste your time. Send the order number and two photos of the shoulder seams, and we’ll route the jacket for a repair assessment. If it cannot be repaired, the team will explain the warranty options before suggesting a replacement.
The bot becomes more direct and less decorative. It does not abandon the repair-first belief, perform exaggerated empathy, or make a warranty promise before inspection.
Test Recognition, Usefulness, and Drift
Previewing one polished answer is not enough. Jasper’s current Brand Voice workflow compares output with and without the selected voice, which is a useful start. A customer-facing chatbot also needs tests for pressure and multi-turn drift.
Run these six cases whenever the instructions, model, or major support content changes:
| Test | Prompt | Passing signal |
|---|---|---|
| Blind recognition | Mix five bot replies with five competitor replies | A reviewer can identify yours from decisions and language, not slogans |
| Simple answer | Ask a yes-or-no pricing or policy question | The direct answer arrives before personality |
| Frustration | Say that two earlier steps failed | Cheerfulness drops; specificity and ownership rise |
| Sales pressure | Ask for the “best” option with a small stated need | The bot recommends fit, including a cheaper option when appropriate |
| Missing fact | Ask about an undocumented feature | Voice remains steady while the bot admits the evidence gap |
| Long conversation | Revisit the same issue after 15 turns | Beliefs, terminology, and boundaries remain consistent |
Score each reply on four dimensions from 0 to 2: recognizable, useful, grounded, and appropriate to the visitor’s emotional state. A reply that sounds distinctive but buries the answer should not pass. Neither should a correct reply that violates a core business decision.
Use at least two reviewers from different functions. Marketing may catch generic language, while support notices when a charming sentence creates another customer turn. Disagreement is valuable: it shows which contract rule is too vague to enforce.
Keep Voice Separate From Safety
Voice rules describe how an allowed answer should sound. Safety and policy rules decide whether the chatbot may provide the answer or take an action. Keeping those layers separate makes both easier to test.
For example, “be candid about limitations” does not authorize the bot to reveal internal incidents. “Use the customer’s own terminology” does not permit repeating abuse or discriminatory language. “Recommend the least expensive fit” does not allow the bot to invent discounts.
Run the voice suite beside the chatbot brand-safety tests, not instead of them. If the voice contract conflicts with a knowledge, privacy, authorization, or escalation rule, the safer rule wins and the chatbot should preserve as much of the voice as the safe answer allows.
Let Real Conversations Refine the Contract
A voice contract should change when conversation evidence reveals a pattern, not whenever one stakeholder prefers a different adjective.
Review a sample of short answers, failed answers, escalations, and frustrated conversations every week. Mark exact moments where the bot became generic, overly cheerful, evasive, too verbose, or inconsistent with a company decision. Then edit one rule or example and rerun the saved tests.
The AI chatbot QA process turns production failures into regression cases. Apply the same loop to voice. If “Absolutely!” keeps returning after model changes, add it to the test. If the bot sounds right in English but stiff in another language, create a language-specific example rather than translating an adjective list word for word.
Measure outcomes alongside reviewer scores. Extra turns after a simple answer, repeated questions, abandonment after refusal, and handoff quality reveal whether the voice helps customers move forward. Brand consistency that increases effort is not a successful support voice.
Make the Business Recognizable Through Its Choices
A chatbot sounds like a business when its replies carry the same decisions the business makes: what it recommends first, which tradeoffs it explains, where it admits limits, and how it treats a frustrated customer. Sentence style supports that identity, but it cannot create one from generic adjectives.
In Agentkit, custom instructions can hold the voice contract, while conversation logs provide real replies for the review loop.
No credit card required.
