Manual Conversation Topic Catalogs
A manual conversation topic catalog is usually built as a hierarchical tree with three or four levels.
Purpose: contact center agents need to narrow down their options step by step so they do not have to choose from a list of 400–700 topics at once.
Reason: the tree is a practical compromise based on the limits of human attention and the speed required during manual work.
The Cost of Manual Topic Classification
On average, an agent spends about 10 seconds selecting a call topic manually. A contact center with 100 agents processes approximately 210,000 calls per month. This equals around 583 working hours per month spent only on selecting one call topic. AI can release this time completely.
Moving to AI-Native Classification
Imagine a physical library. A book can be stored on only one shelf. A book titled “The History of Military Aviation” must be placed under either History or Transport. Whichever option is selected, the book becomes difficult to find through the other category. This is hierarchical classification.
AI works with digital labels, or a semantic matrix. It is not limited by physical shelves. The same item can receive the labels #history, #aviation, and #technology at the same time.
AI interprets a conversation in the same way: as a multidimensional object containing several independent layers of meaning.
Problem: Why Hierarchical Trees Cause Data Loss
A tree created for manual input can begin to distort business statistics when it is used for automation. This happens for three main reasons:
Loss of up to 40% of related context. Real customer conversations are usually multi-topic. If a customer reports a failed payment, requests a product return, and complains about a courier, an agent working with a tree must select only one branch. Only one issue enters the statistics, while the others are lost.
Unusual growth of the Other category to 20–30% of calls. A manual catalog is static, but customer topics change. If a new micro-topic does not yet exist in the tree, conversations are assigned to Other, and management loses visibility of them.
High time cost. Agents spend an average of 5–15 seconds selecting categories and clicking through branches for every call. Across a contact center, this becomes hundreds of paid hours that do not create value for the customer.
Solution: Move from a Hierarchy to Vector Classification
AI does not need folders or nested levels. It can analyze the complete conversation and convert it into a flat matrix of semantic labels. Several labels can be assigned to one call at the same time and with different weights.
Comparison of the Approaches
Classification Principle
Current approach — hierarchical tree: one call equals one topic, so related problems are lost.
New approach — AI vector matrix: one call can contain several topics, creating multidimensional analytics and a more complete view of customer requests.
Data Flexibility
Current approach — hierarchical tree: the structure is static, and changes require additional staff training.
New approach — AI vector matrix: the system can provide additional information about previously unknown conversation context alongside the defined topic list.
Operations
Current approach — hierarchical tree: agents spend time selecting branches manually.
New approach — AI vector matrix: AI classifies 100% of conversations automatically and can generate additional detailed descriptions.
Automation Barriers: Why AI Struggles with a Manual Catalog
The main barrier to AI implementation is transferring the old hierarchical catalog without changing it. The manual catalog was designed to reduce agent effort, not to support deep business analytics.
Limitations of the Existing Structure
Insufficient contextual information. Topic names often contain only one or two words, such as Failure or Error. This is not enough for AI to match natural customer language to the correct topic accurately.
Duplicate topics. The same meaning may appear in different branches. For example, Refund may exist under both Products and Finance. Agents follow learned rules, but AI sees two similar values unless the full context explains how to choose between them.
Outdated categories. Manual catalogs may retain topics related to expired campaigns, discontinued products, or old problems, while new topics are introduced too slowly.
A long tail of low-frequency topics. Narrow categories that occur in less than 1% of calls create noise. They complicate AI training without providing meaningful analytical value.
Focus on agent navigation instead of business logic. The old tree is optimized so an agent can complete several clicks quickly. It reflects interface convenience rather than real business processes, products, or customer needs.
Transformation Stage: Preparing the Catalog for AI
Before launching AI classification, complete a required technology preparation stage consisting of three steps:
Audit and clean the catalog: remove the long tail of low-frequency topics, delete outdated categories, and merge duplicates.
Move to matrix-based business logic: organize topics in a flat structure in which every topic is linked to a specific end-to-end process, product, internal team, customer need, or other business context. For example: Product team → Operational process → Specific problem.
Create a semantic matrix for AI: write detailed contextual descriptions for every topic. Explain not only the topic name, but also the context, marker words, and customer meanings associated with it.
Summary
This preparation transforms a limited topic tree created for manual selection into a structured catalog of business contexts. AI can then use the catalog to provide multidimensional business analytics.