Gains:
- Ability to design a natural and reliable speech with the principles of clarity, brevity, repair, consistency and progression.
- Ability to establish intent architecture in a small, clear and discrete manner and define a gentle repair and handover path for incomprehensible input.
- Ability to stabilize the bot's voice with a tone guide and deliver bad news with empathy while keeping concrete information tied to the source
It is not enough for a bot to work technically; It must give the customer a good experience. Consider two bots giving the same information: one saying "INVALID LOGIN. TRY AGAIN." says the other, "I think I didn't get your order number exactly, can you write your 10-digit number?" says. They both do the same job, but one annoys the customer, the other reassures them. The discipline that makes this difference is called conversation design: it is the art of planning what a bot will say, how it will say it, how it will behave in case of an error, and how it will direct the customer to the solution. Artificial intelligence makes this design both easier (generating dialogue texts) and harder (free production can get out of control).
In this unit, we will see how to make virtual assistants speak naturally and reliably, how to establish intent architecture, and the principles of conversation design. First a term: virtual assistant is an advanced assistant that is broader than a simple bot that does a single task, can multitask, remembers context, and works across multiple channels (phone, chat, app).
Basic principles of speech design
Good speech design is based on several principles. Clarity: the customer should know what to do at each step; Ambiguous questions ("cargo, return, or any other issue" when necessary instead of "how can I help you") increase the speed of solution. Brevity: the bot should write short, single-minded sentences, not long paragraphs; In a voice bot, this is even more critical because the customer does not listen to the long announcement. Repair: When the customer is not understood or makes a mistake, the bot should gently redirect without blaming. Consistency: the tone of the bot should remain the same throughout; It should not be formal in one sentence and sincere in the other. Sense of progress: the client should feel like they are getting closer to the solution with each step, not going around in circles.
The skeleton of a speech design is established with these steps:
- List user goals: What 10-15 tasks do customers come to this assistant for most? (These are intentions.)
- For each intention, write down the “happy path”: the shortest dialogue in which everything goes well. The customer comes, the bot understands, collects the information and solves it.
- Write down detours: Customer gives false information, goes off topic, gets angry, says "representative." One repair sentence for each deviation.
- Mark the handover points: At what moments does the bot stop and connect to the human?
- Tone and personality guide: 5-6 example sentences that match the brand voice; Sentences that the bot will never say.
Tip: The quickest way to test a dialogue design is to read it out loud. If it sounds robotic, formal, or rude, it will sound that way to the customer. Good bot text should flow like a good agent's speech.
Intention architecture: few and clear intentions, many examples
In the classical approach, dozens of "example sentences" (utterances) are collected for each intent: "where is my cargo", "when will my order arrive", "track my package" are all connected to the "cargo_tracking" intent. This job has become easier with generative AI; The model largely understands these variations on its own. But there are two traps. The first is intent conflict: if two intents are too similar (e.g. “return” and “exchange”), the bot will confuse; Separate intentions clearly. The second is scope creep: when the model cannot recognize the intent, it tries to generate a made-up intent instead of saying "other"; Therefore, there must be a "did not understand → clarify or delegate" way.
The following table compares good and poor intent design:
Subject
poor design
good design
number of intentions
80 scattered, conflicting intentions
12-15 clear, discrete intentions
Incomprehensible input
Produces a made-up answer
"I don't quite understand" + option/transfer
missing information
gives error
Asks nicely (slot-filling)
out of scope
continues talking
Politely sets limits and delegates
tone
variable, robotic
Consistent, brand voice
Four copyable templates
1) Speech design guide production:
Your role: conversation designer. I'm designing a virtual assistant for [brand]. For the following 12 intents, generate each: (a) 1-sentence intent, (b) first question the bot will ask, (c) assets to collect if missing, (d) solution sentence, (e) repair sentence if not understood. Tone: polite, short, "you" address, not robotic. Intentions: <<list>>
2) Production of repair (error) sentences:
Write polite repair sentences that do not blame the customer for the following situations:1) The order number is in the wrong format,2) The bot did not understand the customer at all (2nd time),3) The customer is angry ("I'm fed up"),4) Authentication failed. No more than 2 sentences each; Let it direct you to the solution and, if necessary, delegate it to people.
3) Tone/personality guide:
Produce tone guide for [brand] virtual assistant:- Brand values: <<trustworthy, warm, homely>>- 5 sample sentences of "We talk like this",- 5 sample sentences of "We don't talk like that" (too formal/slang/robotic),- Emoji and capitalization rule,- Tone sample when delivering bad news (out of stock, return refusal).
4) Dialogue review prompt:
Review the bot dialogue below from a conversation designer's perspective:<<dialogue>>Flag the following: (1) overly long/robotic sentences,(2) vague questions, (3) risk of looping/repetition,(4) missing cycle point, (5) tonal inconsistency.Suggest corrected sentences for each problem. Making up new information/policies.
Weak prompt / Strong prompt
Weak prompt:
Write a virtual assistant dialogue.
Vague: which brand, which subject, which tone, which channel, no turnover rules. The output becomes generic and unusable.
Powerful prompt:
Write the "policy renewal" intent dialog for the [Brand: health insurance] phone voicebot. Tone: calm, reassuring, short sentences (to be read aloud). Obtain policy number and date of birth (for final verification) from the customer. Assume that the renewal amount comes from the LIVE system, do not make it up; If the amount does not arrive, say "I could not withdraw the information, I am binding you." When it comes to the payment step, never scan/save the card information as required by PCI-DSS, direct it to the secure payment step.
Difference: channel (voice), intent, tone, data limit, fabrication ban and security rule clear.
three mini cases
Case 1 — The power of repair. A bank's voicebot misunderstood the customer 18% of the time due to ASR (speech recognition) error. In the past, in this case, it would say "Cannot understand, returning to the main menu" and the customer would go crazy. When repair phrases were renewed ("I couldn't hear you clearly, can you just say 'card', 'credit' or 'account'?"), retry success increased by 41% and handover requests to a human decreased by 23%. Same error rate, different design, very different experience.
Case 2 — Tone disaster. An e-commerce bot will send you bad news, such as a return rejection, saying "YOUR REQUEST IS NOT ELIGIBLE." he was giving. Complaints increased. When the tone guide was added and the same information was given as "Unfortunately, this product falls outside our return policy, but I can offer you the following options...", the CSAT score of customers who received the same rejection increased from 2.1 to 3.6 (out of 5). The information is the same, the style has changed everything.
Case 3 — Conflict of intent. Since the intentions of "invoice objection" and "bill payment" were trained with very similar example sentences in a telecom assistant, the bot confused the two and put the customer who wanted to pay into the objection flow. By separating the intentions and adding a clear confirmatory question to each (“Do you want to pay the bill or dispute the bill?”), misdirection was reduced by 90%.
Common mistakes
- Long and robotic sentences. Especially in a voice bot, the customer does not listen to long announcements; Use short, single-minded sentences.
- Vague open-ended questions. “How can I help” is good sometimes, but for most flows, providing a clear option will resolve things faster.
- Forgetting the repair path. Error and misunderstanding are inevitable; each step should have a "retry/clarify/delegate" path.
- Letting the tone slide. A sincere and official boot in one place does not inspire trust; tone guide is a must.
- Breaking bad news coldly. Giving news such as rejection, delay, out of stock with empathy and alternatives saves the experience.
Attention: Generative AI writes dialogues very fluently, but this fluency is not a guarantee of accuracy. Every concrete information (amount, date, condition) produced by the bot must also depend on the source; Design beauty does not justify fabrication.
In summary
Conversational design is the discipline that moves a bot from being a “worker” to a “good experience giver.” Embrace the principles of clarity, brevity, repair, consistency, and a sense of progression; For each intention, write down both the happy path and the deviation paths; keep intentions small and clear; Be sure to define a polite exit for unintelligible input and out of scope. Stabilize the bot's voice with the tone guide and deliver bad news with empathy. Remember: generative AI writes fluently, but concrete information must still be sourced and verified.
Application task
For an industry of your choice (banking, e-commerce, insurance, telecom) write a full conversation design of a single intent (“initiate a return” or “make an appointment”): happy path dialogue, at least 3 detours (misinformation, out of scope, angry customer), and a repair sentence for each detour. Then read the text out loud and shorten any sentences you find robotic/long. Finally, have your own dialogue audited by the AI with the "4) Dialogue auditing" template.
checklist
- [ ] For every intention I have designed both the happy path and the paths of deviation.
- [ ] I have polite repair and handover phrases for incomprehensible input and out of scope.
- [ ] My intentions are few, clear and distinct; no conflict.
- [ ] I wrote a tone/personality guide and the bot speaks consistently throughout.
- [ ] I prepared sentences that give bad news with empathy and alternatives.
- [ ] I verified that the concrete information in the dialogue depends on the source.