A contact center needs to route customers into a fixed set of 12 complaint categories based on a short web form field plus a brief free-text reason…

Standard Natural Language Processing (NLP) rules, such as regular expressions or keyword matching, are the best fit. This approach is deterministic, making it highly explainable for audits. It is extremely fast, easily meeting the sub-150ms requirement. Most importantly, rule-based systems do not require a machine learning model for inference, meaning they have zero incremental model spend, adhering to the strict budget constraint. The rules can be centrally managed and deployed as part of a configuration file or simple service, satisfying the governance requirement. Given the fixed, slowly changing taxonomy, a rule-based system is efficient and sufficient.

B. Azure CLU trained intent model project: This involves training and hosting a model, which incurs inference costs, violating the "no increase model spend" requirement.

C. Generative orchestration with tool-based classification: This is far too slow, expensive, and complex for a simple, high-speed routing task. It would grossly violate the performance and budget constraints.

D. Copilot Studio generative answers: This feature is designed for question-answering over a knowledge base, not for high-speed, low-cost classification of inputs into a fixed taxonomy.

1. Jurafsky, D., & Martin, J. H. (2023). Speech and Language Processing (3rd ed. draft). Chapter 2: Regular Expressions, Text Normalization, Edit Distance. This chapter from a foundational university textbook describes the speed and efficiency of rule-based methods for text classification.

2. Microsoft Azure AI Language Documentation, "What is conversational language understanding (CLU)?". The documentation implicitly contrasts CLU's model-based approach with simpler methods by highlighting its ability to understand conversational nuances, which are not required for this simple routing task. The cost is detailed in the "Azure AI services pricing" page.