A chatbot can answer common questions reliably yet stumble when a user asks something outside its planned topics. Traditional conversational AI works by recognizing a request as a predefined intent, then following the answer or dialogue flow assigned to it. If no intent fits, the user may get a generic “I didn’t understand” response. section 6.1
From request to action
“I want to book a flight” can match a booking intent and prompt the bot to collect a destination. “I need a one-way ticket to Costa Rica” can match the same intent while also supplying a destination entity, allowing the bot to ask for departure details instead. The entity refines the next step without requiring a separate intent for every phrasing. section 4.1.3
Table 4.1 Example utterances may be handled differently based on the presence or absence of entities.
|
Utterance
|
Intent
|
Entity
|
Possible next step
|
|---|---|---|---|
|
“How many bags can I check?”
|
Bag_Allowance
|
Display bag check policy
|
|
|
“I want to book a flight”
|
Book_Flight
|
Collect destination
|
|
|
“I need a one-way ticket to Costa Rica”
|
Book_Flight
|
Costa Rica
|
Collect departure details
|
|
“I’d like to upgrade my seat to first class”
|
Flight_Upgrade
|
first class
|
Initiate upgrade process
|
Compare how example utterances map to intents and how a detected entity can change the next step. section 4.1.3
An intent groups paraphrases of the same goal. A classifier learns from example requests grouped under those intents; conversational logic then chooses an action, such as answering, asking a follow-up question, or starting a transaction. This suits FAQ bots, topic routing, and some self-service tasks because the expected topics and responses can be planned in advance. section 4.1.3 section 2.4
The planning is also the constraint. A cake-ordering bot might classify familiar questions and guide orders through rules, but users asking for many different dinner recipes create a difficult new range of requests. Adding an intent for every possible topic is costly and can blur distinctions between intents. Prioritize topics that matter to users rather than treating unlimited coverage as the goal. section 2.3 section 4.1.3
When answers miss the mark, improve the classifier methodically:
- Inspect the intent list and training examples for overlapping topics, uneven example counts, and gaps between the bot’s purpose and what users ask. section 4.3.1
- Test against representative real user utterances with known correct intents; establish a baseline, then repeat tests as training changes. section 4.6
- Start with high-volume intents that perform poorly. Where requests remain ambiguous, consider clarifying questions, disambiguation, or entity detection. section 5.5
Traditional AI need not handle every request alone. Search can serve less common questions; retrieval-augmented generation can turn retrieved passages into a more specific answer. An existing recipe intent could also route requests to a large language model. These extensions call for safeguards: generated answers may require fallbacks or handover to a human agent. Keep measuring whether users receive the right answer or make progress toward their goal. section 2.3 section 6.6 section 4.6
