By KYN AI Advisory Team — AI implementation specialists, Singapore
A sales agent sounds like your team when you give it a written voice sheet, a small set of message patterns and real examples, then test its output against your own sent messages. Conversation design is the work of deciding what the agent says, in what order, in what words, and when it stops. It is a writing job first and a technical job second. Do it before you touch a prompt.
What is conversation design for a sales agent?
It is the specification of a conversation: the goal of each message, the tone, the length, the questions, the fixed phrases and the exits. A model can write any text. Design narrows it to the text that is right for your business.
Without design, agents drift toward a generic voice: enthusiastic, long and vague. Buyers notice. The fix is not a better model. It is a tighter brief.
What goes in a voice sheet?
One page. Write it from your strongest real messages, not from adjectives.
| Item | Example entry | |---|---| | Greeting | "Hi Rina," never "Dear Sir or Madam" | | Length | Under 80 words in a first message, under 50 in a reply | | Sentences | Short. One idea each | | Sign-off | First name only | | Words we use | "quote", "site visit", "lead time" | | Words we avoid | "solutions", "synergy", "revolutionary" | | Emoji | None in email, one at most in WhatsApp | | Formality by market | Formal in Jakarta first contact, direct in Singapore | | Never say | Promises of dates, results or discounts |
Fill this from 20 to 30 messages your top salesperson sent that got a reply. Mark what they have in common. That is your voice.
Which message patterns does a sales agent need?
Most sales conversations use six patterns. Design each one once.
- First touch. One checkable fact about them, one line about what you do, one question.
- Reply to interest. Thanks, the next step, two time options.
- Reply to a question. The answer first, then a single follow-up question.
- Reply to "not now". Acknowledge, confirm a date to return, stop.
- Handover. Name the person, say what they will do, say when.
- Close of a thread. A short last message, leaving the door open without pressure.
Write each as a template with slots, plus two sample outputs. The agent fills the slots from the CRM and the thread. Fixed structure keeps replies consistent, and slots keep them specific.
Which phrases should be fixed and which free?
Fix anything with legal, financial or reputational weight. Free anything that is about fit and tone.
- Fixed text: the unsubscribe line, how you describe regulated services, the holding message for complaints, the wording on price validity, the description of what is and is not included.
- Guided text: the first line, the reason for the call, the answer to a product question. The model writes it within the voice sheet and from a cited source.
- Free text: the connective sentences. Keep them short.
Fixed text is not written by the model at all. It is inserted by code. That removes a class of errors.
How should the agent handle handover and exits?
Design the exits as carefully as the entries. A conversation ends in one of five ways: booked, handed over, deferred, declined or escalated. Each needs wording.
- Handover: "I am passing this to Arif, who handles marine insurance. He will message you today before 17:00."
- Deferred: "Understood. I will check back in March."
- Declined: "Thank you for letting me know. I will not contact you again about this."
- Escalated: a holding message, then a person.
Never let the agent disguise itself. If asked, it says it is an assistant working for your company. State this wherever local rules or platform terms require it. Check the rule for each market you sell in.
Avoid these tells of machine writing: openers that praise the reader, three-part lists in every message, summaries that repeat the question, and closing offers of further help in every reply.
How do you test whether it sounds like your team?
Run a blind test.
- Take 30 real replies from your team and 30 drafts from the agent for the same situations.
- Remove names and mix them.
- Ask three colleagues to mark each: team or agent.
- Ask a second question: would you send this as written?
If colleagues spot the agent more than half the time, your voice sheet is too thin. If they would not send more than a third unchanged, the patterns are wrong.
Then use the edit rate in production. Every edit a rep makes is data. Group the edits: tone, length, fact, structure. Update the voice sheet monthly. See how to measure an AI reply desk.
What about multiple languages?
Design the voice sheet once per language, not by translating the English one. Greetings, formality and the way a request is softened differ across English, Chinese and Bahasa. Have a native speaker from your team write the examples and review the drafts. Report test results by language.
What does a worked example look like?
Take an illustrative marine insurance broker. The voice sheet says: first name only, under 80 words, no emoji in email, "cover" not "solutions". The first-touch pattern has three slots: observation, offer line, question.
Hi Arif, your site now lists a second vessel under the Batam entity. Brokers who place hull cover for two-vessel fleets often find the renewal dates drift apart. We align them for a single review date. Is that on your list this year?
The observation comes from a cited source. The offer line is a fixed sentence written by the broker. The question is the only free text. A reviewer sees all three parts and the source link, and approves or edits.
Now compare the generic version a model writes without design: "I hope you are well! I came across your impressive company and would love to explore how our innovative solutions could help." It breaks four rules on the voice sheet and contains no checkable fact.
What does a message pattern library look like in practice?
Store patterns as records, not as paragraphs in a prompt.
| Field | Example | |---|---| | Pattern name | Reply to interest | | Goal | Get a 15-minute call booked | | Slots | First name, topic, two time options | | Fixed text | Sign-off, unsubscribe line | | Length limit | 50 words | | Banned content | Prices, dates not from the calendar | | Owner | Head of sales | | Last reviewed | Date |
A person edits the record. The agent reads it. Changes take effect without a deploy, and the log shows which version produced each message.
Where KYN fits
KYN designs the conversation with your team during the build: the voice sheet, the patterns and the fixed text, written from your own strongest messages. The agent runs in your own cloud account under a restricted account, and each message and edit is logged with the actor. Outbound messages wait for a named person. See sales systems and our note on why an agent must hold the line on discounts.
For this work, the 14-day review includes collecting your 20 to 30 strongest sent messages and sitting with your top seller to mark what makes them work. The full build ships in under 90 days. Rights in the custom code and configuration pass to you on full payment.
Frequently asked questions
Do I need a long prompt to make the agent sound like us?
No. A one-page voice sheet, six message patterns and 20 real examples do more than a long prompt. Keep the prompt short and put the detail in the sheet and the templates.
Should the agent say it is an AI?
It should not pretend to be a person. If a customer asks, it says it is an assistant working for your company. Check disclosure rules for your market and channel.
How often should the voice sheet change?
Review it monthly in the first quarter, then each quarter. Change it when the edit log shows the same correction repeated.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
