Classifying Inquiry Intent and Switching Response Patterns

Introduction

"Inquiry" is one word, but its contents vary widely. "I want to return this" and "will this product suit me?" call for different information and a different tone entirely.

Try to process every inquiry with the same prompt and the draft comes out wishy-washy. So the reply-draft system first classifies the inquiry's intent and switches the response format accordingly.

This article explains the thinking behind intent classification and the mechanism for switching response patterns by intent.

Why Classify Intent

The "types" of inquiry are limited

Look at enough inquiries and the contents actually converge into a few types. Returns and exchanges, shipping-status checks, stock inquiries, product-selection consultations. For each type, the information to check and the response structure are set.

If you identify the type first, you can give the AI a precise instruction: "this is a shipping-status check, so answer around the order status." Process it without identifying the type, and you get an incoherent draft — stock guidance mixed into a return conversation.

Required information changes by intent

Once intent is known, the context to inject changes too. Shipping status needs order status and delivery destination; returns need purchase date and return policy; product consultations need stock and product specs. The information to prioritize differs by intent.

Doing classification first lets you optimize the downstream context gathering and response generation for that intent.

Classifying Intent and Switching Responses

Intent detection with the LLM

Intent classification is left to the LLM. It's handed the inquiry text and asked which of the predefined intent categories it fits. Unlike rule-based keyword matching, it can catch diverse phrasings — "hasn't arrived," "still hasn't come," "has it shipped?" — all as the same "shipping status" intent.

A single inquiry sometimes contains multiple intents (e.g., "I want to return this, but first I'd like to check stock"). Even then, the design centers on the primary intent while treating secondary intents as supplementary.

Applying intent-specific templates

Based on the detected intent, the system switches the response template. "Template" here isn't a fill-in-the-blank sentence but a set of guidelines telling the AI "for this intent, in what order and in what tone should you write."

For returns, for instance, it passes a structural template like "first acknowledge receipt, then show the steps numbered, and finally point to the contact for any questions." This yields drafts of consistent quality per intent.

Product consultations get an added service knowledge base

Product consultations get special treatment. Simply returning product info isn't enough; they call for a service-style dialogue that draws out "what's the use case," "the budget," "what have you used until now" while proposing.

So when an inquiry is classified as a product consultation, the system additionally injects a service knowledge base capturing product-diagnosis conversation patterns into the prompt. By teaching how a veteran salesperson layers questions and leads to a proposal, the draft shifts from "manual handling" to "genuinely consultative handling."

What I Keep in Mind When Designing Classification

Allow "unclassifiable"

Not every inquiry fits neatly into a type. Complex complaints or inquiries tangling several topics get worse if you force them into a type. So the design anticipates cases where "I can't confidently classify this into any type," and in that case it either keeps the draft generic or defers to the operator's judgment. Not forcing automation is itself a way to protect quality.

Templates are something you grow

Intent-specific templates aren't made once and done. Watching actual drafts and operators' edits reveals improvements like "for this intent, it's better to add this one line." Refining templates while operating gradually reduces the editing effort.

Conclusion

Classifying inquiry intent and switching response formats by intent significantly raises draft precision.

• Identify the type first, then optimize the injected information and response structure
• LLM classification's strength is catching diverse phrasings as the same intent
• Product consultations get an added service knowledge base, elevating them to consultative handling
• Allow unclassifiable cases, and grow templates through operation

How humans review and send the intent-optimized drafts comes next, in Designing the Approval Flow and Operator Experience. For gathering context, see Injecting Customer Context; for the big picture, the hub article.