Types of AI convos
There are three types of AI conversations:
Curated conversations are designed by a human using an NLU model. The intents are defined and the model is trained to direct certain utterances to that intent. At Intuit, these types of models are used for legally regulated topics or when the GenAI model is down.
2. Generative AI (GenAI)
These conversations are generated by an LLM where the text and persona may vary, depending on temperature and prompt boundaries. GenAI models like Intuit Assist typically take raw data (facts) and can turn them into insights or suggestions.
Data (fact): Descriptive, now, “you are here.”
Example: “You have 2 overdue invoices.”
Insight: Predictive, happened recently or will happen soon—a peek into a possible trajectory. Gives the "so what" or the “why” of raw data. See also: Insights response pattern
Example: “Your labor expenses are 20% lower than similar businesses.”
Suggestion: Prescriptive, near future—suggests other possible trajectories.
Example: “You could reach out to former customers, or offer discounts to new customers.”
Hybrid conversations use a curated intent structure—including training the intent—but may have part of the response generated by an LLM. It may also use the LLM to do sentiment analysis or other work behind the scenes without the customer seeing any genAI text.
Confused on some of these terms? Check out the conversation design terms to know
No matter what type of AI conversation you’re writing for, keep the following guidance in mind.
Conversational maxims for AI
In addition to following the conversational basics, Intuit Assist and other generative AI conversational experiences must follow customer expectations and conversational norms to be successful. Conversations that don’t follow Gricean maxims are annoying and frustrating.
Good conversation is:
- Cooperative. Systems should actively support the user and require less effort to interact with.
- Goal-oriented. User goals and needs should be explored via user research as part of the design process
- Context-aware. The more a system can respond to contextual cues, the better it'll be at having a natural conversation. Intuit Assist must understand the difference between a new question and a follow-up question. If the customer asks about profitability last quarter, then a follow-up question about expenses is likely also about last quarter, even if the customer doesn’t specify it.
- Quick and clear. Save users time and mental exertion by being succinct and unambiguous. In normal human-to-human conversation, we expect and tolerate pauses. When using software, customers have higher expectations for speed, expecting immediate answers to what they consider simple questions. They don’t have patience for long latency—and customer research has validated this.
- Turn-based. Functional conversations should avoid long monologues on the part of the system and make it clear whose turn it is at every moment. See also: Context and listening
- Truthful. There should be a strong correlation between what the user expects and what the system offers.
- Polite. Design interactions that respect a user’s time and don't impose.
- Error-tolerant. Texting language is notoriously imprecise and error-prone. Generative AI experiences like Intuit Assist must be able to forgive customer errors in typing, spelling, and grammar, as well as understand made-up words like business names or non-English customer names.
See also: Conversation basics