Every support taxonomy begins too clean.

A team lists the obvious reasons customers contact support: billing, refunds, shipping, account access, product questions, cancellations, technical issues. The categories look reasonable. They may even work for a while.

Then production happens.

Customers do not speak in taxonomy. They explain their lives. A refund question includes a policy complaint. A shipping issue includes an address correction. A billing concern includes a product misunderstanding. A customer starts in English, switches to Spanish, and asks a family member to clarify the account holder’s name. One conversation contains a stack of smaller needs.

Subintent discovery exists because broad intents are rarely enough.

Intent is the front door. Subintent is the room where the work happens.

An intent names the general purpose of a conversation. A subintent names the more specific job inside it.

Billing is an intent. Duplicate charge, failed refund, expired promotion, tax confusion, payment method update, and cancellation penalty are subintents. Delivery is an intent. Address verification, missing driver, late order, wrong drop-off, replacement request, and delivery photo dispute are subintents.

The distinction matters because AI support agents need operational specificity. A broad intent may route the conversation. A subintent determines what questions to ask, which policy to retrieve, which tool to call, and when to escalate.

A human agent can often infer the subintent from context. An AI agent needs the support system to represent it. That representation can come from manual design, production discovery, or both.

Subintent is where a support taxonomy becomes executable.

Why manual taxonomies decay

Manual taxonomies are necessary. They are also fragile.

Support leaders design categories based on known problems. Product teams launch new features. Policies change. Markets expand. Customers use unexpected language. Agents create workarounds. New failure modes appear inside old labels. Over time, the taxonomy begins to describe the organization’s memory more than the customer’s present reality.

This decay is easy to miss. A dashboard may still show “billing issue” or “delivery issue,” but the work inside that category may have changed completely. One broad tag can hide ten different operational problems.

Microsoft’s Customer Intent Agent is relevant because it uses generative AI to autonomously discover intents from historical conversations and create an intent library that supports both self-service and assisted-service scenarios. The underlying premise is important: customer intent should be learned from real interactions, not only defined in a planning room.

Subintent discovery pushes that idea deeper. It asks what smaller jobs are hiding inside the intent categories a company already uses.

Production conversations are the best taxonomy source

Synthetic examples are helpful for design. Production conversations are better for discovery.

Production data contains the strange phrasing, missing details, mixed intents, code-switching, emotional pressure, tool failures, and policy edge cases that clean examples remove. Customers rarely provide the exact slot values a workflow expects. They mention the wrong order number. They describe symptoms instead of causes. They ask for a refund when the real issue is eligibility. They ask for an agent because the system previously failed them.

Subintent discovery from production conversations treats those messy details as signal.

A system can summarize conversations, embed them, cluster them, label recurring groups, and propose new subintent categories. Those proposals are not automatically true. They are candidates for review. A support team can accept, reject, merge, split, rename, or monitor them.

A living taxonomy should be negotiated between AI discovery and human operational judgment.

How subintent discovery helps AI agents

AI support agents improve when their operating world becomes more explicit.

A broad intent may tell the agent that a customer has a billing issue. A subintent can tell the agent that the customer is disputing a promotional credit that expired because the payment method failed. Those are very different workflows. The first requires generic billing help. The second may require eligibility checks, payment history, policy lookup, and careful explanation.

Subintent discovery can improve several parts of the agent system:

It can create better routing. It can trigger better follow-up questions. It can attach more relevant policy context. It can limit tool exposure to what the scenario needs. It can generate more realistic eval cases. It can help Agent Canvas represent the workflow a customer actually experiences, instead of the workflow a team assumed existed.

A support agent does not become smarter only because the model improves. It becomes smarter when the environment around it is structured more accurately.

Subintent discovery and multilingual support

Multilingual support makes subintent discovery more important, not less.

A broad intent may appear consistent across languages while subintents diverge. Customers in one region may contact support about plan changes. Customers in another may contact support about identity verification. Spanish-speaking callers may use different phrasing for the same operational issue, or mix English product names into Spanish explanations. A taxonomy built only from English examples may miss important structure.

Strong subintent discovery should therefore preserve language as a signal. Language should not be flattened away too early. A translated summary may help the system reason, but the original utterance can reveal patterns that translation obscures.

For a multilingual voice platform, this matters commercially. Language coverage is table stakes. Resolution quality across languages is the real product standard. Subintent discovery helps teams understand whether different language groups are encountering the same problems or different versions of the support system.

From discovered subintent to operational change

A discovered subintent should not end as a label. It should change the support system.

If a new subintent appears frequently and has high escalation, the team may build a dedicated agent scenario. If it has high repeat contact, the team may inspect the policy or tool path. If it appears after a product launch, the product team may need clearer messaging. If it correlates with a specific language, the team may examine translation, voice recognition, or regional policy.

Microsoft’s intent-based suggestions documentation describes a workflow where detected intents help service representatives understand customer needs, ask relevant questions, and access suggested solutions. That is a useful operational pattern. Intent and subintent data should guide the conversation, not simply label it afterward.

For Giga, the stronger story is even more agentic: discovered subintents can inform Smart Insights, update scenarios, shape eval cases, improve policies, and eventually connect to KPI movement.

What to measure

Subintent discovery should be evaluated by whether it improves the support operation.

Useful metrics include subintent coverage, cluster stability, human acceptance rate, escalation by subintent, repeat contact by subintent, resolution by subintent, routing improvement, scenario creation rate, and KPI lift after workflow changes. Another useful metric is taxonomy freshness: how long does it take the system to detect a new meaningful support pattern after it appears in production?

A discovered subintent that never changes routing, policy, evaluation, or workflow design may still be interesting. It is not yet operational.

The goal is not to create a beautiful taxonomy. The goal is to make support behavior easier to improve.

Where subintent discovery can go wrong

Automated discovery has failure modes.

A system can over-segment and create too many tiny categories. It can under-segment and leave distinct workflows trapped in one broad bucket. It can label clusters by surface vocabulary instead of operational meaning. It can confuse product names with customer goals. It can treat language variation as intent variation, or erase language-specific patterns through over-normalization.

Human review protects against these errors. Reviewers should ask whether the proposed subintent is actionable, stable, distinct, and connected to a measurable support outcome. If no workflow, policy, tool, question, or measurement changes because the subintent exists, the category may not be useful yet.

Good taxonomy is not just accurate. It is operationally meaningful.

The new support taxonomy

Support taxonomies used to be static control systems. A team decided the categories, trained agents to use them, and reviewed the dashboard later.

AI support changes the cadence. Production conversations can now help maintain the taxonomy itself. New subintents can emerge from real usage. Existing categories can split when the work inside them becomes too diverse. Old categories can disappear when the product or policy changes. Agent scenarios can be updated based on what customers actually ask.

This is a healthier model.

A support taxonomy should not be a monument. It should be an instrument. It should stay tuned to the customer’s changing language, the company’s changing product, and the agent’s changing responsibilities.

Subintent discovery is one way to keep that instrument in tune.