6.1.2. When to Use Azure Language Service: PII, Volume, and Specialized Tasks
💡 First Principle: The dedicated Azure Language service wins precisely where LLM-first is weak — guaranteed PII detection/redaction, high-volume low-latency low-cost batch NLP, and specialized fine-tuned capabilities (text analytics for health). It uses fine-tuned models that reduce prompt engineering and run cheaper at scale. Match the constraint (privacy, volume, specialization) to the service.
Azure Language (formerly Azure AI Language, now in Foundry Tools) provides NER, PII detection/redaction, sentiment/opinion mining, language detection, summarization (extractive and abstractive), key-phrase extraction, conversational language understanding (CLU), custom question answering (CQA), and text analytics for health. Two reasons to choose it over a general model: PII redaction is a privacy guarantee often used to preprocess text before sending it to an LLM (remove sensitive data first), and high-volume/cost workloads run cheaper and faster on fine-tuned task models than on a flagship generative model per record. It also integrates with agents — its MCP server exposes NLP tools, and CQA/CLU projects deploy to agents.
| Choose dedicated Language service when | Choose LLM-first when |
|---|---|
| Guaranteed PII detection/redaction | Flexible, novel, or mixed tasks |
| High volume, low latency, low cost | Low/medium volume |
| Specialized (health text analytics) | One model for many tasks |
| Preprocessing before sending to an LLM | Reasoning over the result inline |
⚠️ Exam Trap: For a privacy requirement — "remove PII before sending customer text to a generative model" — the answer is the Language service's PII redaction as a preprocessing step, not prompting the LLM to "ignore the PII." You don't hand sensitive data to the model and ask it to be careful; you redact first. Routing raw PII to the model and trusting instructions is the privacy trap.
Reflection Question: A pipeline must strip Social Security numbers from millions of support transcripts before any LLM sees them. Why is the dedicated PII redaction service the right answer over an LLM-first prompt, on both privacy and cost grounds?
